Hugging Face's logo

busyfarm
/
whisper-coreml-models

whisper
core-ml
apple
Apple-Silicon
Model card Files
xet
 Community
busyfarm-org commited on
Commit
db98e44
·
verified ·
1 Parent(s): 16e0707

Add files using upload-large-folder tool

Browse files
Files changed (44) hide show
  1. base.en/ggml-base.en.bin +3 -0
  2. base/ggml-base.bin +3 -0
  3. large-v1/ggml-large-v1-encoder.mlmodelc/weights/weight.bin +3 -0
  4. large-v1/ggml-large-v1.bin +3 -0
  5. large-v2/ggml-large-v2-encoder.mlmodelc/weights/weight.bin +3 -0
  6. large-v2/ggml-large-v2.bin +3 -0
  7. large-v3/ggml-large-v3-encoder.mlmodelc/weights/weight.bin +3 -0
  8. large-v3/ggml-large-v3.bin +3 -0
  9. medium.en/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin +3 -0
  10. medium.en/ggml-medium.en-encoder.mlmodelc/coremldata.bin +3 -0
  11. medium.en/ggml-medium.en-encoder.mlmodelc/metadata.json +71 -0
  12. medium.en/ggml-medium.en-encoder.mlmodelc/model.mil +0 -0
  13. medium.en/ggml-medium.en-encoder.mlmodelc/weights/weight.bin +3 -0
  14. medium.en/ggml-medium.en.bin +3 -0
  15. medium/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin +3 -0
  16. medium/ggml-medium-encoder.mlmodelc/coremldata.bin +3 -0
  17. medium/ggml-medium-encoder.mlmodelc/metadata.json +71 -0
  18. medium/ggml-medium-encoder.mlmodelc/model.mil +0 -0
  19. medium/ggml-medium-encoder.mlmodelc/weights/weight.bin +3 -0
  20. medium/ggml-medium.bin +3 -0
  21. small.en/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin +3 -0
  22. small.en/ggml-small.en-encoder.mlmodelc/coremldata.bin +3 -0
  23. small.en/ggml-small.en-encoder.mlmodelc/metadata.json +71 -0
  24. small.en/ggml-small.en-encoder.mlmodelc/model.mil +0 -0
  25. small.en/ggml-small.en-encoder.mlmodelc/weights/weight.bin +3 -0
  26. small.en/ggml-small.en.bin +3 -0
  27. small/ggml-small-encoder.mlmodelc/analytics/coremldata.bin +3 -0
  28. small/ggml-small-encoder.mlmodelc/coremldata.bin +3 -0
  29. small/ggml-small-encoder.mlmodelc/metadata.json +71 -0
  30. small/ggml-small-encoder.mlmodelc/model.mil +0 -0
  31. small/ggml-small-encoder.mlmodelc/weights/weight.bin +3 -0
  32. small/ggml-small.bin +3 -0
  33. tiny.en/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin +3 -0
  34. tiny.en/ggml-tiny.en-encoder.mlmodelc/coremldata.bin +3 -0
  35. tiny.en/ggml-tiny.en-encoder.mlmodelc/metadata.json +71 -0
  36. tiny.en/ggml-tiny.en-encoder.mlmodelc/model.mil +463 -0
  37. tiny.en/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin +3 -0
  38. tiny.en/ggml-tiny.en.bin +3 -0
  39. tiny/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin +3 -0
  40. tiny/ggml-tiny-encoder.mlmodelc/coremldata.bin +3 -0
  41. tiny/ggml-tiny-encoder.mlmodelc/metadata.json +71 -0
  42. tiny/ggml-tiny-encoder.mlmodelc/model.mil +463 -0
  43. tiny/ggml-tiny-encoder.mlmodelc/weights/weight.bin +3 -0
  44. tiny/ggml-tiny.bin +3 -0
base.en/ggml-base.en.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a03779c86df3323075f5e796cb2ce5029f00ec8869eee3fdfb897afe36c6d002
3
+ size 147964211
base/ggml-base.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe
3
+ size 147951465
large-v1/ggml-large-v1-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac5cd426142b4d97debcc46d2f536ae257ad80cc168790faf2ad3dcc862462a6
3
+ size 1273600512
large-v1/ggml-large-v1.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d99f41a10525d0206bddadd86760181fa920438b6b33237e3118ff6c83bb53d
3
+ size 3094623691
large-v2/ggml-large-v2-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b0425ba64707bb4c9727963b5e2b0ee1e8ddece596a56c7795aca0d966614b2
3
+ size 1273600512
large-v2/ggml-large-v2.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a423fe4d40c82774b6af34115b8b935f34152246eb19e80e376071d3f999487
3
+ size 3094623691
large-v3/ggml-large-v3-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:806bd7aef5df068fea795e1af8a671cf8817f42f5179e3624632bfbbcbad869f
3
+ size 1273969152
large-v3/ggml-large-v3.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64d182b440b98d5203c4f9bd541544d84c605196c4f7b845dfa11fb23594d1e2
3
+ size 3095033483
medium.en/ggml-medium.en-encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:423209035d398f1e6945478cc49cf8e83022a3dc4e00d735c54289b3f62cdf89
3
+ size 243
medium.en/ggml-medium.en-encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d20a4fe17a031efc213c5c295df6967c6e87eba9cca3f07fa63c2beb835ca420
3
+ size 320
medium.en/ggml-medium.en-encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float16",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32 1 × 1500 × 1024)",
11
+ "shortDescription" : "",
12
+ "shape" : "[1, 1500, 1024]",
13
+ "name" : "output",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 6,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Concat" : 24,
23
+ "Gelu" : 26,
24
+ "LayerNorm" : 49,
25
+ "Transpose" : 25,
26
+ "Softmax" : 384,
27
+ "Squeeze" : 1,
28
+ "Cast" : 2,
29
+ "Add" : 49,
30
+ "Einsum" : 768,
31
+ "ExpandDims" : 1,
32
+ "Split" : 72,
33
+ "Conv" : 146
34
+ },
35
+ "computePrecision" : "Mixed (Float16, Float32, Int32)",
36
+ "isUpdatable" : "0",
37
+ "stateSchema" : [
38
+
39
+ ],
40
+ "availability" : {
41
+ "macOS" : "12.0",
42
+ "tvOS" : "15.0",
43
+ "visionOS" : "1.0",
44
+ "watchOS" : "8.0",
45
+ "iOS" : "15.0",
46
+ "macCatalyst" : "15.0"
47
+ },
48
+ "modelType" : {
49
+ "name" : "MLModelType_mlProgram"
50
+ },
51
+ "userDefinedMetadata" : {
52
+ "com.github.apple.coremltools.source_dialect" : "TorchScript",
53
+ "com.github.apple.coremltools.source" : "torch==2.2.2",
54
+ "com.github.apple.coremltools.version" : "8.3.0"
55
+ },
56
+ "inputSchema" : [
57
+ {
58
+ "hasShapeFlexibility" : "0",
59
+ "isOptional" : "0",
60
+ "dataType" : "Float32",
61
+ "formattedType" : "MultiArray (Float32 1 × 80 × 3000)",
62
+ "shortDescription" : "",
63
+ "shape" : "[1, 80, 3000]",
64
+ "name" : "logmel_data",
65
+ "type" : "MultiArray"
66
+ }
67
+ ],
68
+ "generatedClassName" : "coreml_encoder_medium_en",
69
+ "method" : "predict"
70
+ }
71
+ ]
medium.en/ggml-medium.en-encoder.mlmodelc/model.mil ADDED
The diff for this file is too large to render. See raw diff
 
medium.en/ggml-medium.en-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:74029c2746f01f4f43ebbcdaa98dfb6e0597eb7e3d4008fdef72c2113a4d2483
3
+ size 614456320
medium.en/ggml-medium.en.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cc37e93478338ec7700281a7ac30a10128929eb8f427dda2e865faa8f6da4356
3
+ size 1533774781
medium/ggml-medium-encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:423209035d398f1e6945478cc49cf8e83022a3dc4e00d735c54289b3f62cdf89
3
+ size 243
medium/ggml-medium-encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d20a4fe17a031efc213c5c295df6967c6e87eba9cca3f07fa63c2beb835ca420
3
+ size 320
medium/ggml-medium-encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float16",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32 1 × 1500 × 1024)",
11
+ "shortDescription" : "",
12
+ "shape" : "[1, 1500, 1024]",
13
+ "name" : "output",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 6,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Concat" : 24,
23
+ "Gelu" : 26,
24
+ "LayerNorm" : 49,
25
+ "Transpose" : 25,
26
+ "Softmax" : 384,
27
+ "Squeeze" : 1,
28
+ "Cast" : 2,
29
+ "Add" : 49,
30
+ "Einsum" : 768,
31
+ "ExpandDims" : 1,
32
+ "Split" : 72,
33
+ "Conv" : 146
34
+ },
35
+ "computePrecision" : "Mixed (Float16, Float32, Int32)",
36
+ "isUpdatable" : "0",
37
+ "stateSchema" : [
38
+
39
+ ],
40
+ "availability" : {
41
+ "macOS" : "12.0",
42
+ "tvOS" : "15.0",
43
+ "visionOS" : "1.0",
44
+ "watchOS" : "8.0",
45
+ "iOS" : "15.0",
46
+ "macCatalyst" : "15.0"
47
+ },
48
+ "modelType" : {
49
+ "name" : "MLModelType_mlProgram"
50
+ },
51
+ "userDefinedMetadata" : {
52
+ "com.github.apple.coremltools.source_dialect" : "TorchScript",
53
+ "com.github.apple.coremltools.version" : "8.3.0",
54
+ "com.github.apple.coremltools.source" : "torch==2.2.2"
55
+ },
56
+ "inputSchema" : [
57
+ {
58
+ "hasShapeFlexibility" : "0",
59
+ "isOptional" : "0",
60
+ "dataType" : "Float32",
61
+ "formattedType" : "MultiArray (Float32 1 × 80 × 3000)",
62
+ "shortDescription" : "",
63
+ "shape" : "[1, 80, 3000]",
64
+ "name" : "logmel_data",
65
+ "type" : "MultiArray"
66
+ }
67
+ ],
68
+ "generatedClassName" : "coreml_encoder_medium",
69
+ "method" : "predict"
70
+ }
71
+ ]
medium/ggml-medium-encoder.mlmodelc/model.mil ADDED
The diff for this file is too large to render. See raw diff
 
medium/ggml-medium-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2972798300feedfddbf8f3f898b8492320d538e8ecddebbfdf0d03a82e00b83
3
+ size 614456320
medium/ggml-medium.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c14d5adee5f86394037b4e4e8b59f1673b6cee10e3cf0b11bbdbee79c156208
3
+ size 1533763059
small.en/ggml-small.en-encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ea4d5727e7b1c8762c548f4b80a30447740200e986075b056cf3da16359da56
3
+ size 243
small.en/ggml-small.en-encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a9554f8fae587d40c808e5a6411d79ce20ab6b473d7a3a88df2cd8123978fb4
3
+ size 320
small.en/ggml-small.en-encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float16",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32 1 × 1500 × 768)",
11
+ "shortDescription" : "",
12
+ "shape" : "[1, 1500, 768]",
13
+ "name" : "output",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 6,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Concat" : 12,
23
+ "Gelu" : 14,
24
+ "LayerNorm" : 25,
25
+ "Transpose" : 13,
26
+ "Softmax" : 144,
27
+ "Squeeze" : 1,
28
+ "Cast" : 2,
29
+ "Add" : 25,
30
+ "Einsum" : 288,
31
+ "ExpandDims" : 1,
32
+ "Split" : 36,
33
+ "Conv" : 74
34
+ },
35
+ "computePrecision" : "Mixed (Float16, Float32, Int32)",
36
+ "isUpdatable" : "0",
37
+ "stateSchema" : [
38
+
39
+ ],
40
+ "availability" : {
41
+ "macOS" : "12.0",
42
+ "tvOS" : "15.0",
43
+ "visionOS" : "1.0",
44
+ "watchOS" : "8.0",
45
+ "iOS" : "15.0",
46
+ "macCatalyst" : "15.0"
47
+ },
48
+ "modelType" : {
49
+ "name" : "MLModelType_mlProgram"
50
+ },
51
+ "userDefinedMetadata" : {
52
+ "com.github.apple.coremltools.version" : "8.3.0",
53
+ "com.github.apple.coremltools.source_dialect" : "TorchScript",
54
+ "com.github.apple.coremltools.source" : "torch==2.2.2"
55
+ },
56
+ "inputSchema" : [
57
+ {
58
+ "hasShapeFlexibility" : "0",
59
+ "isOptional" : "0",
60
+ "dataType" : "Float32",
61
+ "formattedType" : "MultiArray (Float32 1 × 80 × 3000)",
62
+ "shortDescription" : "",
63
+ "shape" : "[1, 80, 3000]",
64
+ "name" : "logmel_data",
65
+ "type" : "MultiArray"
66
+ }
67
+ ],
68
+ "generatedClassName" : "coreml_encoder_small_en",
69
+ "method" : "predict"
70
+ }
71
+ ]
small.en/ggml-small.en-encoder.mlmodelc/model.mil ADDED
The diff for this file is too large to render. See raw diff
 
small.en/ggml-small.en-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57d5901ae064655c83da5a7281e643d56974fd875e986e6bd7997038e65e428a
3
+ size 176320256
small.en/ggml-small.en.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c6138d6d58ecc8322097e0f987c32f1be8bb0a18532a3f88f734d1bbf9c41e5d
3
+ size 487614201
small/ggml-small-encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ea4d5727e7b1c8762c548f4b80a30447740200e986075b056cf3da16359da56
3
+ size 243
small/ggml-small-encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a9554f8fae587d40c808e5a6411d79ce20ab6b473d7a3a88df2cd8123978fb4
3
+ size 320
small/ggml-small-encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float16",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32 1 × 1500 × 768)",
11
+ "shortDescription" : "",
12
+ "shape" : "[1, 1500, 768]",
13
+ "name" : "output",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 6,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Concat" : 12,
23
+ "Gelu" : 14,
24
+ "LayerNorm" : 25,
25
+ "Transpose" : 13,
26
+ "Softmax" : 144,
27
+ "Squeeze" : 1,
28
+ "Cast" : 2,
29
+ "Add" : 25,
30
+ "Einsum" : 288,
31
+ "ExpandDims" : 1,
32
+ "Split" : 36,
33
+ "Conv" : 74
34
+ },
35
+ "computePrecision" : "Mixed (Float16, Float32, Int32)",
36
+ "isUpdatable" : "0",
37
+ "stateSchema" : [
38
+
39
+ ],
40
+ "availability" : {
41
+ "macOS" : "12.0",
42
+ "tvOS" : "15.0",
43
+ "visionOS" : "1.0",
44
+ "watchOS" : "8.0",
45
+ "iOS" : "15.0",
46
+ "macCatalyst" : "15.0"
47
+ },
48
+ "modelType" : {
49
+ "name" : "MLModelType_mlProgram"
50
+ },
51
+ "userDefinedMetadata" : {
52
+ "com.github.apple.coremltools.version" : "8.3.0",
53
+ "com.github.apple.coremltools.source" : "torch==2.2.2",
54
+ "com.github.apple.coremltools.source_dialect" : "TorchScript"
55
+ },
56
+ "inputSchema" : [
57
+ {
58
+ "hasShapeFlexibility" : "0",
59
+ "isOptional" : "0",
60
+ "dataType" : "Float32",
61
+ "formattedType" : "MultiArray (Float32 1 × 80 × 3000)",
62
+ "shortDescription" : "",
63
+ "shape" : "[1, 80, 3000]",
64
+ "name" : "logmel_data",
65
+ "type" : "MultiArray"
66
+ }
67
+ ],
68
+ "generatedClassName" : "coreml_encoder_small",
69
+ "method" : "predict"
70
+ }
71
+ ]
small/ggml-small-encoder.mlmodelc/model.mil ADDED
The diff for this file is too large to render. See raw diff
 
small/ggml-small-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fd99222f556ad6a561a7286c4dc94d14fee3d2b60bdeb4e838bfed437f501b1e
3
+ size 176320256
small/ggml-small.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1be3a9b2063867b937e64e2ec7483364a79917e157fa98c5d94b5c1fffea987b
3
+ size 487601967
tiny.en/ggml-tiny.en-encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c811ab131cbe7fef6230b32b61cb04cc99fb5990e5fb70ab5d7ec907a4a124b2
3
+ size 243
tiny.en/ggml-tiny.en-encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93254f2adaadd6535328df3b1200d4922c2f701f84df2c9b37c9a5df9ae4f4b7
3
+ size 320
tiny.en/ggml-tiny.en-encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float16",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32 1 × 1500 × 384)",
11
+ "shortDescription" : "",
12
+ "shape" : "[1, 1500, 384]",
13
+ "name" : "output",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 6,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Concat" : 4,
23
+ "Gelu" : 6,
24
+ "LayerNorm" : 9,
25
+ "Transpose" : 5,
26
+ "Softmax" : 24,
27
+ "Squeeze" : 1,
28
+ "Cast" : 2,
29
+ "Add" : 9,
30
+ "Einsum" : 48,
31
+ "ExpandDims" : 1,
32
+ "Split" : 12,
33
+ "Conv" : 26
34
+ },
35
+ "computePrecision" : "Mixed (Float16, Float32, Int32)",
36
+ "isUpdatable" : "0",
37
+ "stateSchema" : [
38
+
39
+ ],
40
+ "availability" : {
41
+ "macOS" : "12.0",
42
+ "tvOS" : "15.0",
43
+ "visionOS" : "1.0",
44
+ "watchOS" : "8.0",
45
+ "iOS" : "15.0",
46
+ "macCatalyst" : "15.0"
47
+ },
48
+ "modelType" : {
49
+ "name" : "MLModelType_mlProgram"
50
+ },
51
+ "userDefinedMetadata" : {
52
+ "com.github.apple.coremltools.version" : "8.3.0",
53
+ "com.github.apple.coremltools.source" : "torch==2.2.2",
54
+ "com.github.apple.coremltools.source_dialect" : "TorchScript"
55
+ },
56
+ "inputSchema" : [
57
+ {
58
+ "hasShapeFlexibility" : "0",
59
+ "isOptional" : "0",
60
+ "dataType" : "Float32",
61
+ "formattedType" : "MultiArray (Float32 1 × 80 × 3000)",
62
+ "shortDescription" : "",
63
+ "shape" : "[1, 80, 3000]",
64
+ "name" : "logmel_data",
65
+ "type" : "MultiArray"
66
+ }
67
+ ],
68
+ "generatedClassName" : "coreml_encoder_tiny_en",
69
+ "method" : "predict"
70
+ }
71
+ ]
tiny.en/ggml-tiny.en-encoder.mlmodelc/model.mil ADDED
@@ -0,0 +1,463 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ program(1.0)
2
+ [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
3
+ {
4
+ func main<ios15>(tensor<fp32, [1, 80, 3000]> logmel_data) {
5
+ tensor<string, []> var_28_pad_type_0 = const()[name = tensor<string, []>("op_28_pad_type_0"), val = tensor<string, []>("custom")];
6
+ tensor<int32, [2]> var_28_pad_0 = const()[name = tensor<string, []>("op_28_pad_0"), val = tensor<int32, [2]>([1, 1])];
7
+ tensor<int32, [1]> var_28_strides_0 = const()[name = tensor<string, []>("op_28_strides_0"), val = tensor<int32, [1]>([1])];
8
+ tensor<int32, [1]> var_28_dilations_0 = const()[name = tensor<string, []>("op_28_dilations_0"), val = tensor<int32, [1]>([1])];
9
+ tensor<int32, []> var_28_groups_0 = const()[name = tensor<string, []>("op_28_groups_0"), val = tensor<int32, []>(1)];
10
+ tensor<string, []> logmel_data_to_fp16_dtype_0 = const()[name = tensor<string, []>("logmel_data_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
11
+ tensor<fp16, [384, 80, 3]> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, [384, 80, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
12
+ tensor<fp16, [384]> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184448)))];
13
+ tensor<fp16, [1, 80, 3000]> logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor<string, []>("cast_20")];
14
+ tensor<fp16, [1, 384, 3000]> var_28_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_28_dilations_0, groups = var_28_groups_0, pad = var_28_pad_0, pad_type = var_28_pad_type_0, strides = var_28_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
15
+ tensor<string, []> input_1_mode_0 = const()[name = tensor<string, []>("input_1_mode_0"), val = tensor<string, []>("EXACT")];
16
+ tensor<fp16, [1, 384, 3000]> input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_28_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
17
+ tensor<string, []> var_46_pad_type_0 = const()[name = tensor<string, []>("op_46_pad_type_0"), val = tensor<string, []>("custom")];
18
+ tensor<int32, [2]> var_46_pad_0 = const()[name = tensor<string, []>("op_46_pad_0"), val = tensor<int32, [2]>([1, 1])];
19
+ tensor<int32, [1]> var_46_strides_0 = const()[name = tensor<string, []>("op_46_strides_0"), val = tensor<int32, [1]>([2])];
20
+ tensor<int32, [1]> var_46_dilations_0 = const()[name = tensor<string, []>("op_46_dilations_0"), val = tensor<int32, [1]>([1])];
21
+ tensor<int32, []> var_46_groups_0 = const()[name = tensor<string, []>("op_46_groups_0"), val = tensor<int32, []>(1)];
22
+ tensor<fp16, [384, 384, 3]> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, [384, 384, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185280)))];
23
+ tensor<fp16, [384]> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1070080)))];
24
+ tensor<fp16, [1, 384, 1500]> var_46_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("op_46_cast_fp16")];
25
+ tensor<string, []> x_3_mode_0 = const()[name = tensor<string, []>("x_3_mode_0"), val = tensor<string, []>("EXACT")];
26
+ tensor<fp16, [1, 384, 1500]> x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_46_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
27
+ tensor<fp16, [384, 1500]> var_51_to_fp16 = const()[name = tensor<string, []>("op_51_to_fp16"), val = tensor<fp16, [384, 1500]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1070912)))];
28
+ tensor<fp16, [1, 384, 1500]> var_53_cast_fp16 = add(x = x_3_cast_fp16, y = var_51_to_fp16)[name = tensor<string, []>("op_53_cast_fp16")];
29
+ tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([2])];
30
+ tensor<fp16, [1, 384, 1, 1500]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_53_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
31
+ tensor<int32, []> var_68 = const()[name = tensor<string, []>("op_68"), val = tensor<int32, []>(1)];
32
+ tensor<int32, [1]> input_3_axes_0 = const()[name = tensor<string, []>("input_3_axes_0"), val = tensor<int32, [1]>([1])];
33
+ tensor<fp16, [384]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2222976)))];
34
+ tensor<fp16, [384]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2223808)))];
35
+ tensor<fp16, []> var_84_to_fp16 = const()[name = tensor<string, []>("op_84_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
36
+ tensor<fp16, [1, 384, 1, 1500]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_84_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
37
+ tensor<string, []> q_1_pad_type_0 = const()[name = tensor<string, []>("q_1_pad_type_0"), val = tensor<string, []>("valid")];
38
+ tensor<int32, [2]> q_1_strides_0 = const()[name = tensor<string, []>("q_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
39
+ tensor<int32, [4]> q_1_pad_0 = const()[name = tensor<string, []>("q_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
40
+ tensor<int32, [2]> q_1_dilations_0 = const()[name = tensor<string, []>("q_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
41
+ tensor<int32, []> q_1_groups_0 = const()[name = tensor<string, []>("q_1_groups_0"), val = tensor<int32, []>(1)];
42
+ tensor<fp16, [384, 384, 1, 1]> var_119_weight_0_to_fp16 = const()[name = tensor<string, []>("op_119_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2224640)))];
43
+ tensor<fp16, [384]> var_119_bias_0_to_fp16 = const()[name = tensor<string, []>("op_119_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2519616)))];
44
+ tensor<fp16, [1, 384, 1, 1500]> var_119_cast_fp16 = conv(bias = var_119_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_119_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_119_cast_fp16")];
45
+ tensor<string, []> k_1_pad_type_0 = const()[name = tensor<string, []>("k_1_pad_type_0"), val = tensor<string, []>("valid")];
46
+ tensor<int32, [2]> k_1_strides_0 = const()[name = tensor<string, []>("k_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
47
+ tensor<int32, [4]> k_1_pad_0 = const()[name = tensor<string, []>("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
48
+ tensor<int32, [2]> k_1_dilations_0 = const()[name = tensor<string, []>("k_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
49
+ tensor<int32, []> k_1_groups_0 = const()[name = tensor<string, []>("k_1_groups_0"), val = tensor<int32, []>(1)];
50
+ tensor<fp16, [384, 384, 1, 1]> blocks_0_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2520448)))];
51
+ tensor<fp16, [1, 384, 1, 1500]> k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
52
+ tensor<string, []> var_117_pad_type_0 = const()[name = tensor<string, []>("op_117_pad_type_0"), val = tensor<string, []>("valid")];
53
+ tensor<int32, [2]> var_117_strides_0 = const()[name = tensor<string, []>("op_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
54
+ tensor<int32, [4]> var_117_pad_0 = const()[name = tensor<string, []>("op_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
55
+ tensor<int32, [2]> var_117_dilations_0 = const()[name = tensor<string, []>("op_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
56
+ tensor<int32, []> var_117_groups_0 = const()[name = tensor<string, []>("op_117_groups_0"), val = tensor<int32, []>(1)];
57
+ tensor<fp16, [384, 384, 1, 1]> blocks_0_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2815424)))];
58
+ tensor<fp16, [384]> blocks_0_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3110400)))];
59
+ tensor<fp16, [1, 384, 1, 1500]> var_117_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_117_dilations_0, groups = var_117_groups_0, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_117_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_117_cast_fp16")];
60
+ tensor<int32, [6]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
61
+ tensor<int32, []> var_120_axis_0 = const()[name = tensor<string, []>("op_120_axis_0"), val = tensor<int32, []>(1)];
62
+ tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_5 = split(axis = var_120_axis_0, split_sizes = tile_0, x = var_119_cast_fp16)[name = tensor<string, []>("op_120_cast_fp16")];
63
+ tensor<int32, [4]> var_127_perm_0 = const()[name = tensor<string, []>("op_127_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
64
+ tensor<int32, [6]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
65
+ tensor<int32, []> var_128_axis_0 = const()[name = tensor<string, []>("op_128_axis_0"), val = tensor<int32, []>(3)];
66
+ tensor<fp16, [1, 1500, 1, 384]> var_127_cast_fp16 = transpose(perm = var_127_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_4")];
67
+ tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_5 = split(axis = var_128_axis_0, split_sizes = tile_1, x = var_127_cast_fp16)[name = tensor<string, []>("op_128_cast_fp16")];
68
+ tensor<int32, [6]> tile_2 = const()[name = tensor<string, []>("tile_2"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
69
+ tensor<int32, []> var_135_axis_0 = const()[name = tensor<string, []>("op_135_axis_0"), val = tensor<int32, []>(1)];
70
+ tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_5 = split(axis = var_135_axis_0, split_sizes = tile_2, x = var_117_cast_fp16)[name = tensor<string, []>("op_135_cast_fp16")];
71
+ tensor<string, []> aw_1_equation_0 = const()[name = tensor<string, []>("aw_1_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
72
+ tensor<fp16, [1, 1500, 1, 1500]> aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_128_cast_fp16_0, var_120_cast_fp16_0))[name = tensor<string, []>("aw_1_cast_fp16")];
73
+ tensor<string, []> aw_3_equation_0 = const()[name = tensor<string, []>("aw_3_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
74
+ tensor<fp16, [1, 1500, 1, 1500]> aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_128_cast_fp16_1, var_120_cast_fp16_1))[name = tensor<string, []>("aw_3_cast_fp16")];
75
+ tensor<string, []> aw_5_equation_0 = const()[name = tensor<string, []>("aw_5_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
76
+ tensor<fp16, [1, 1500, 1, 1500]> aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_128_cast_fp16_2, var_120_cast_fp16_2))[name = tensor<string, []>("aw_5_cast_fp16")];
77
+ tensor<string, []> aw_7_equation_0 = const()[name = tensor<string, []>("aw_7_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
78
+ tensor<fp16, [1, 1500, 1, 1500]> aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_128_cast_fp16_3, var_120_cast_fp16_3))[name = tensor<string, []>("aw_7_cast_fp16")];
79
+ tensor<string, []> aw_9_equation_0 = const()[name = tensor<string, []>("aw_9_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
80
+ tensor<fp16, [1, 1500, 1, 1500]> aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_128_cast_fp16_4, var_120_cast_fp16_4))[name = tensor<string, []>("aw_9_cast_fp16")];
81
+ tensor<string, []> aw_11_equation_0 = const()[name = tensor<string, []>("aw_11_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
82
+ tensor<fp16, [1, 1500, 1, 1500]> aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_128_cast_fp16_5, var_120_cast_fp16_5))[name = tensor<string, []>("aw_11_cast_fp16")];
83
+ tensor<fp16, [1, 1500, 1, 1500]> var_154_cast_fp16 = softmax(axis = var_68, x = aw_1_cast_fp16)[name = tensor<string, []>("op_154_cast_fp16")];
84
+ tensor<fp16, [1, 1500, 1, 1500]> var_155_cast_fp16 = softmax(axis = var_68, x = aw_3_cast_fp16)[name = tensor<string, []>("op_155_cast_fp16")];
85
+ tensor<fp16, [1, 1500, 1, 1500]> var_156_cast_fp16 = softmax(axis = var_68, x = aw_5_cast_fp16)[name = tensor<string, []>("op_156_cast_fp16")];
86
+ tensor<fp16, [1, 1500, 1, 1500]> var_157_cast_fp16 = softmax(axis = var_68, x = aw_7_cast_fp16)[name = tensor<string, []>("op_157_cast_fp16")];
87
+ tensor<fp16, [1, 1500, 1, 1500]> var_158_cast_fp16 = softmax(axis = var_68, x = aw_9_cast_fp16)[name = tensor<string, []>("op_158_cast_fp16")];
88
+ tensor<fp16, [1, 1500, 1, 1500]> var_159_cast_fp16 = softmax(axis = var_68, x = aw_11_cast_fp16)[name = tensor<string, []>("op_159_cast_fp16")];
89
+ tensor<string, []> var_161_equation_0 = const()[name = tensor<string, []>("op_161_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
90
+ tensor<fp16, [1, 64, 1, 1500]> var_161_cast_fp16 = einsum(equation = var_161_equation_0, values = (var_135_cast_fp16_0, var_154_cast_fp16))[name = tensor<string, []>("op_161_cast_fp16")];
91
+ tensor<string, []> var_163_equation_0 = const()[name = tensor<string, []>("op_163_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
92
+ tensor<fp16, [1, 64, 1, 1500]> var_163_cast_fp16 = einsum(equation = var_163_equation_0, values = (var_135_cast_fp16_1, var_155_cast_fp16))[name = tensor<string, []>("op_163_cast_fp16")];
93
+ tensor<string, []> var_165_equation_0 = const()[name = tensor<string, []>("op_165_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
94
+ tensor<fp16, [1, 64, 1, 1500]> var_165_cast_fp16 = einsum(equation = var_165_equation_0, values = (var_135_cast_fp16_2, var_156_cast_fp16))[name = tensor<string, []>("op_165_cast_fp16")];
95
+ tensor<string, []> var_167_equation_0 = const()[name = tensor<string, []>("op_167_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
96
+ tensor<fp16, [1, 64, 1, 1500]> var_167_cast_fp16 = einsum(equation = var_167_equation_0, values = (var_135_cast_fp16_3, var_157_cast_fp16))[name = tensor<string, []>("op_167_cast_fp16")];
97
+ tensor<string, []> var_169_equation_0 = const()[name = tensor<string, []>("op_169_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
98
+ tensor<fp16, [1, 64, 1, 1500]> var_169_cast_fp16 = einsum(equation = var_169_equation_0, values = (var_135_cast_fp16_4, var_158_cast_fp16))[name = tensor<string, []>("op_169_cast_fp16")];
99
+ tensor<string, []> var_171_equation_0 = const()[name = tensor<string, []>("op_171_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
100
+ tensor<fp16, [1, 64, 1, 1500]> var_171_cast_fp16 = einsum(equation = var_171_equation_0, values = (var_135_cast_fp16_5, var_159_cast_fp16))[name = tensor<string, []>("op_171_cast_fp16")];
101
+ tensor<bool, []> input_5_interleave_0 = const()[name = tensor<string, []>("input_5_interleave_0"), val = tensor<bool, []>(false)];
102
+ tensor<fp16, [1, 384, 1, 1500]> input_5_cast_fp16 = concat(axis = var_68, interleave = input_5_interleave_0, values = (var_161_cast_fp16, var_163_cast_fp16, var_165_cast_fp16, var_167_cast_fp16, var_169_cast_fp16, var_171_cast_fp16))[name = tensor<string, []>("input_5_cast_fp16")];
103
+ tensor<string, []> var_180_pad_type_0 = const()[name = tensor<string, []>("op_180_pad_type_0"), val = tensor<string, []>("valid")];
104
+ tensor<int32, [2]> var_180_strides_0 = const()[name = tensor<string, []>("op_180_strides_0"), val = tensor<int32, [2]>([1, 1])];
105
+ tensor<int32, [4]> var_180_pad_0 = const()[name = tensor<string, []>("op_180_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
106
+ tensor<int32, [2]> var_180_dilations_0 = const()[name = tensor<string, []>("op_180_dilations_0"), val = tensor<int32, [2]>([1, 1])];
107
+ tensor<int32, []> var_180_groups_0 = const()[name = tensor<string, []>("op_180_groups_0"), val = tensor<int32, []>(1)];
108
+ tensor<fp16, [384, 384, 1, 1]> blocks_0_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3111232)))];
109
+ tensor<fp16, [384]> blocks_0_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3406208)))];
110
+ tensor<fp16, [1, 384, 1, 1500]> var_180_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_180_dilations_0, groups = var_180_groups_0, pad = var_180_pad_0, pad_type = var_180_pad_type_0, strides = var_180_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("op_180_cast_fp16")];
111
+ tensor<fp16, [1, 384, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_180_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
112
+ tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([1])];
113
+ tensor<fp16, [384]> input_7_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_7_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407040)))];
114
+ tensor<fp16, [384]> input_7_beta_0_to_fp16 = const()[name = tensor<string, []>("input_7_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407872)))];
115
+ tensor<fp16, []> var_190_to_fp16 = const()[name = tensor<string, []>("op_190_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
116
+ tensor<fp16, [1, 384, 1, 1500]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_190_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
117
+ tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("valid")];
118
+ tensor<int32, [2]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
119
+ tensor<int32, [4]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
120
+ tensor<int32, [2]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
121
+ tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)];
122
+ tensor<fp16, [1536, 384, 1, 1]> blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3408704)))];
123
+ tensor<fp16, [1536]> blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4588416)))];
124
+ tensor<fp16, [1, 1536, 1, 1500]> input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, 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 = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
125
+ tensor<string, []> input_11_mode_0 = const()[name = tensor<string, []>("input_11_mode_0"), val = tensor<string, []>("EXACT")];
126
+ tensor<fp16, [1, 1536, 1, 1500]> input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
127
+ tensor<string, []> var_216_pad_type_0 = const()[name = tensor<string, []>("op_216_pad_type_0"), val = tensor<string, []>("valid")];
128
+ tensor<int32, [2]> var_216_strides_0 = const()[name = tensor<string, []>("op_216_strides_0"), val = tensor<int32, [2]>([1, 1])];
129
+ tensor<int32, [4]> var_216_pad_0 = const()[name = tensor<string, []>("op_216_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
130
+ tensor<int32, [2]> var_216_dilations_0 = const()[name = tensor<string, []>("op_216_dilations_0"), val = tensor<int32, [2]>([1, 1])];
131
+ tensor<int32, []> var_216_groups_0 = const()[name = tensor<string, []>("op_216_groups_0"), val = tensor<int32, []>(1)];
132
+ tensor<fp16, [384, 1536, 1, 1]> blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4591552)))];
133
+ tensor<fp16, [384]> blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5771264)))];
134
+ tensor<fp16, [1, 384, 1, 1500]> var_216_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_216_dilations_0, groups = var_216_groups_0, pad = var_216_pad_0, pad_type = var_216_pad_type_0, strides = var_216_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("op_216_cast_fp16")];
135
+ tensor<fp16, [1, 384, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_216_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
136
+ tensor<int32, []> var_225 = const()[name = tensor<string, []>("op_225"), val = tensor<int32, []>(1)];
137
+ tensor<int32, [1]> input_13_axes_0 = const()[name = tensor<string, []>("input_13_axes_0"), val = tensor<int32, [1]>([1])];
138
+ tensor<fp16, [384]> input_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_13_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5772096)))];
139
+ tensor<fp16, [384]> input_13_beta_0_to_fp16 = const()[name = tensor<string, []>("input_13_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5772928)))];
140
+ tensor<fp16, []> var_241_to_fp16 = const()[name = tensor<string, []>("op_241_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
141
+ tensor<fp16, [1, 384, 1, 1500]> input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_241_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
142
+ tensor<string, []> q_3_pad_type_0 = const()[name = tensor<string, []>("q_3_pad_type_0"), val = tensor<string, []>("valid")];
143
+ tensor<int32, [2]> q_3_strides_0 = const()[name = tensor<string, []>("q_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
144
+ tensor<int32, [4]> q_3_pad_0 = const()[name = tensor<string, []>("q_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
145
+ tensor<int32, [2]> q_3_dilations_0 = const()[name = tensor<string, []>("q_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
146
+ tensor<int32, []> q_3_groups_0 = const()[name = tensor<string, []>("q_3_groups_0"), val = tensor<int32, []>(1)];
147
+ tensor<fp16, [384, 384, 1, 1]> var_276_weight_0_to_fp16 = const()[name = tensor<string, []>("op_276_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5773760)))];
148
+ tensor<fp16, [384]> var_276_bias_0_to_fp16 = const()[name = tensor<string, []>("op_276_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6068736)))];
149
+ tensor<fp16, [1, 384, 1, 1500]> var_276_cast_fp16 = conv(bias = var_276_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_276_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
150
+ tensor<string, []> k_3_pad_type_0 = const()[name = tensor<string, []>("k_3_pad_type_0"), val = tensor<string, []>("valid")];
151
+ tensor<int32, [2]> k_3_strides_0 = const()[name = tensor<string, []>("k_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
152
+ tensor<int32, [4]> k_3_pad_0 = const()[name = tensor<string, []>("k_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
153
+ tensor<int32, [2]> k_3_dilations_0 = const()[name = tensor<string, []>("k_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
154
+ tensor<int32, []> k_3_groups_0 = const()[name = tensor<string, []>("k_3_groups_0"), val = tensor<int32, []>(1)];
155
+ tensor<fp16, [384, 384, 1, 1]> blocks_1_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6069568)))];
156
+ tensor<fp16, [1, 384, 1, 1500]> k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("k_3_cast_fp16")];
157
+ tensor<string, []> var_274_pad_type_0 = const()[name = tensor<string, []>("op_274_pad_type_0"), val = tensor<string, []>("valid")];
158
+ tensor<int32, [2]> var_274_strides_0 = const()[name = tensor<string, []>("op_274_strides_0"), val = tensor<int32, [2]>([1, 1])];
159
+ tensor<int32, [4]> var_274_pad_0 = const()[name = tensor<string, []>("op_274_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
160
+ tensor<int32, [2]> var_274_dilations_0 = const()[name = tensor<string, []>("op_274_dilations_0"), val = tensor<int32, [2]>([1, 1])];
161
+ tensor<int32, []> var_274_groups_0 = const()[name = tensor<string, []>("op_274_groups_0"), val = tensor<int32, []>(1)];
162
+ tensor<fp16, [384, 384, 1, 1]> blocks_1_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6364544)))];
163
+ tensor<fp16, [384]> blocks_1_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6659520)))];
164
+ tensor<fp16, [1, 384, 1, 1500]> var_274_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_274_dilations_0, groups = var_274_groups_0, pad = var_274_pad_0, pad_type = var_274_pad_type_0, strides = var_274_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("op_274_cast_fp16")];
165
+ tensor<int32, [6]> tile_3 = const()[name = tensor<string, []>("tile_3"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
166
+ tensor<int32, []> var_277_axis_0 = const()[name = tensor<string, []>("op_277_axis_0"), val = tensor<int32, []>(1)];
167
+ tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_5 = split(axis = var_277_axis_0, split_sizes = tile_3, x = var_276_cast_fp16)[name = tensor<string, []>("op_277_cast_fp16")];
168
+ tensor<int32, [4]> var_284_perm_0 = const()[name = tensor<string, []>("op_284_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
169
+ tensor<int32, [6]> tile_4 = const()[name = tensor<string, []>("tile_4"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
170
+ tensor<int32, []> var_285_axis_0 = const()[name = tensor<string, []>("op_285_axis_0"), val = tensor<int32, []>(3)];
171
+ tensor<fp16, [1, 1500, 1, 384]> var_284_cast_fp16 = transpose(perm = var_284_perm_0, x = k_3_cast_fp16)[name = tensor<string, []>("transpose_3")];
172
+ tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_5 = split(axis = var_285_axis_0, split_sizes = tile_4, x = var_284_cast_fp16)[name = tensor<string, []>("op_285_cast_fp16")];
173
+ tensor<int32, [6]> tile_5 = const()[name = tensor<string, []>("tile_5"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
174
+ tensor<int32, []> var_292_axis_0 = const()[name = tensor<string, []>("op_292_axis_0"), val = tensor<int32, []>(1)];
175
+ tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_5 = split(axis = var_292_axis_0, split_sizes = tile_5, x = var_274_cast_fp16)[name = tensor<string, []>("op_292_cast_fp16")];
176
+ tensor<string, []> aw_13_equation_0 = const()[name = tensor<string, []>("aw_13_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
177
+ tensor<fp16, [1, 1500, 1, 1500]> aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_285_cast_fp16_0, var_277_cast_fp16_0))[name = tensor<string, []>("aw_13_cast_fp16")];
178
+ tensor<string, []> aw_15_equation_0 = const()[name = tensor<string, []>("aw_15_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
179
+ tensor<fp16, [1, 1500, 1, 1500]> aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_285_cast_fp16_1, var_277_cast_fp16_1))[name = tensor<string, []>("aw_15_cast_fp16")];
180
+ tensor<string, []> aw_17_equation_0 = const()[name = tensor<string, []>("aw_17_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
181
+ tensor<fp16, [1, 1500, 1, 1500]> aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_285_cast_fp16_2, var_277_cast_fp16_2))[name = tensor<string, []>("aw_17_cast_fp16")];
182
+ tensor<string, []> aw_19_equation_0 = const()[name = tensor<string, []>("aw_19_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
183
+ tensor<fp16, [1, 1500, 1, 1500]> aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_285_cast_fp16_3, var_277_cast_fp16_3))[name = tensor<string, []>("aw_19_cast_fp16")];
184
+ tensor<string, []> aw_21_equation_0 = const()[name = tensor<string, []>("aw_21_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
185
+ tensor<fp16, [1, 1500, 1, 1500]> aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_285_cast_fp16_4, var_277_cast_fp16_4))[name = tensor<string, []>("aw_21_cast_fp16")];
186
+ tensor<string, []> aw_23_equation_0 = const()[name = tensor<string, []>("aw_23_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
187
+ tensor<fp16, [1, 1500, 1, 1500]> aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_285_cast_fp16_5, var_277_cast_fp16_5))[name = tensor<string, []>("aw_23_cast_fp16")];
188
+ tensor<fp16, [1, 1500, 1, 1500]> var_311_cast_fp16 = softmax(axis = var_225, x = aw_13_cast_fp16)[name = tensor<string, []>("op_311_cast_fp16")];
189
+ tensor<fp16, [1, 1500, 1, 1500]> var_312_cast_fp16 = softmax(axis = var_225, x = aw_15_cast_fp16)[name = tensor<string, []>("op_312_cast_fp16")];
190
+ tensor<fp16, [1, 1500, 1, 1500]> var_313_cast_fp16 = softmax(axis = var_225, x = aw_17_cast_fp16)[name = tensor<string, []>("op_313_cast_fp16")];
191
+ tensor<fp16, [1, 1500, 1, 1500]> var_314_cast_fp16 = softmax(axis = var_225, x = aw_19_cast_fp16)[name = tensor<string, []>("op_314_cast_fp16")];
192
+ tensor<fp16, [1, 1500, 1, 1500]> var_315_cast_fp16 = softmax(axis = var_225, x = aw_21_cast_fp16)[name = tensor<string, []>("op_315_cast_fp16")];
193
+ tensor<fp16, [1, 1500, 1, 1500]> var_316_cast_fp16 = softmax(axis = var_225, x = aw_23_cast_fp16)[name = tensor<string, []>("op_316_cast_fp16")];
194
+ tensor<string, []> var_318_equation_0 = const()[name = tensor<string, []>("op_318_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
195
+ tensor<fp16, [1, 64, 1, 1500]> var_318_cast_fp16 = einsum(equation = var_318_equation_0, values = (var_292_cast_fp16_0, var_311_cast_fp16))[name = tensor<string, []>("op_318_cast_fp16")];
196
+ tensor<string, []> var_320_equation_0 = const()[name = tensor<string, []>("op_320_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
197
+ tensor<fp16, [1, 64, 1, 1500]> var_320_cast_fp16 = einsum(equation = var_320_equation_0, values = (var_292_cast_fp16_1, var_312_cast_fp16))[name = tensor<string, []>("op_320_cast_fp16")];
198
+ tensor<string, []> var_322_equation_0 = const()[name = tensor<string, []>("op_322_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
199
+ tensor<fp16, [1, 64, 1, 1500]> var_322_cast_fp16 = einsum(equation = var_322_equation_0, values = (var_292_cast_fp16_2, var_313_cast_fp16))[name = tensor<string, []>("op_322_cast_fp16")];
200
+ tensor<string, []> var_324_equation_0 = const()[name = tensor<string, []>("op_324_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
201
+ tensor<fp16, [1, 64, 1, 1500]> var_324_cast_fp16 = einsum(equation = var_324_equation_0, values = (var_292_cast_fp16_3, var_314_cast_fp16))[name = tensor<string, []>("op_324_cast_fp16")];
202
+ tensor<string, []> var_326_equation_0 = const()[name = tensor<string, []>("op_326_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
203
+ tensor<fp16, [1, 64, 1, 1500]> var_326_cast_fp16 = einsum(equation = var_326_equation_0, values = (var_292_cast_fp16_4, var_315_cast_fp16))[name = tensor<string, []>("op_326_cast_fp16")];
204
+ tensor<string, []> var_328_equation_0 = const()[name = tensor<string, []>("op_328_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
205
+ tensor<fp16, [1, 64, 1, 1500]> var_328_cast_fp16 = einsum(equation = var_328_equation_0, values = (var_292_cast_fp16_5, var_316_cast_fp16))[name = tensor<string, []>("op_328_cast_fp16")];
206
+ tensor<bool, []> input_15_interleave_0 = const()[name = tensor<string, []>("input_15_interleave_0"), val = tensor<bool, []>(false)];
207
+ tensor<fp16, [1, 384, 1, 1500]> input_15_cast_fp16 = concat(axis = var_225, interleave = input_15_interleave_0, values = (var_318_cast_fp16, var_320_cast_fp16, var_322_cast_fp16, var_324_cast_fp16, var_326_cast_fp16, var_328_cast_fp16))[name = tensor<string, []>("input_15_cast_fp16")];
208
+ tensor<string, []> var_337_pad_type_0 = const()[name = tensor<string, []>("op_337_pad_type_0"), val = tensor<string, []>("valid")];
209
+ tensor<int32, [2]> var_337_strides_0 = const()[name = tensor<string, []>("op_337_strides_0"), val = tensor<int32, [2]>([1, 1])];
210
+ tensor<int32, [4]> var_337_pad_0 = const()[name = tensor<string, []>("op_337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
211
+ tensor<int32, [2]> var_337_dilations_0 = const()[name = tensor<string, []>("op_337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
212
+ tensor<int32, []> var_337_groups_0 = const()[name = tensor<string, []>("op_337_groups_0"), val = tensor<int32, []>(1)];
213
+ tensor<fp16, [384, 384, 1, 1]> blocks_1_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6660352)))];
214
+ tensor<fp16, [384]> blocks_1_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6955328)))];
215
+ tensor<fp16, [1, 384, 1, 1500]> var_337_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_337_dilations_0, groups = var_337_groups_0, pad = var_337_pad_0, pad_type = var_337_pad_type_0, strides = var_337_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("op_337_cast_fp16")];
216
+ tensor<fp16, [1, 384, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_337_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
217
+ tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([1])];
218
+ tensor<fp16, [384]> input_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_17_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6956160)))];
219
+ tensor<fp16, [384]> input_17_beta_0_to_fp16 = const()[name = tensor<string, []>("input_17_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6956992)))];
220
+ tensor<fp16, []> var_347_to_fp16 = const()[name = tensor<string, []>("op_347_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
221
+ tensor<fp16, [1, 384, 1, 1500]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_347_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
222
+ tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
223
+ tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
224
+ tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
225
+ tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
226
+ tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
227
+ tensor<fp16, [1536, 384, 1, 1]> blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6957824)))];
228
+ tensor<fp16, [1536]> blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8137536)))];
229
+ tensor<fp16, [1, 1536, 1, 1500]> input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, 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 = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
230
+ tensor<string, []> input_21_mode_0 = const()[name = tensor<string, []>("input_21_mode_0"), val = tensor<string, []>("EXACT")];
231
+ tensor<fp16, [1, 1536, 1, 1500]> input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
232
+ tensor<string, []> var_373_pad_type_0 = const()[name = tensor<string, []>("op_373_pad_type_0"), val = tensor<string, []>("valid")];
233
+ tensor<int32, [2]> var_373_strides_0 = const()[name = tensor<string, []>("op_373_strides_0"), val = tensor<int32, [2]>([1, 1])];
234
+ tensor<int32, [4]> var_373_pad_0 = const()[name = tensor<string, []>("op_373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
235
+ tensor<int32, [2]> var_373_dilations_0 = const()[name = tensor<string, []>("op_373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
236
+ tensor<int32, []> var_373_groups_0 = const()[name = tensor<string, []>("op_373_groups_0"), val = tensor<int32, []>(1)];
237
+ tensor<fp16, [384, 1536, 1, 1]> blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8140672)))];
238
+ tensor<fp16, [384]> blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9320384)))];
239
+ tensor<fp16, [1, 384, 1, 1500]> var_373_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_373_dilations_0, groups = var_373_groups_0, pad = var_373_pad_0, pad_type = var_373_pad_type_0, strides = var_373_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_373_cast_fp16")];
240
+ tensor<fp16, [1, 384, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_373_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
241
+ tensor<int32, []> var_382 = const()[name = tensor<string, []>("op_382"), val = tensor<int32, []>(1)];
242
+ tensor<int32, [1]> input_23_axes_0 = const()[name = tensor<string, []>("input_23_axes_0"), val = tensor<int32, [1]>([1])];
243
+ tensor<fp16, [384]> input_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_23_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9321216)))];
244
+ tensor<fp16, [384]> input_23_beta_0_to_fp16 = const()[name = tensor<string, []>("input_23_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322048)))];
245
+ tensor<fp16, []> var_398_to_fp16 = const()[name = tensor<string, []>("op_398_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
246
+ tensor<fp16, [1, 384, 1, 1500]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_398_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
247
+ tensor<string, []> q_5_pad_type_0 = const()[name = tensor<string, []>("q_5_pad_type_0"), val = tensor<string, []>("valid")];
248
+ tensor<int32, [2]> q_5_strides_0 = const()[name = tensor<string, []>("q_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
249
+ tensor<int32, [4]> q_5_pad_0 = const()[name = tensor<string, []>("q_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
250
+ tensor<int32, [2]> q_5_dilations_0 = const()[name = tensor<string, []>("q_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
251
+ tensor<int32, []> q_5_groups_0 = const()[name = tensor<string, []>("q_5_groups_0"), val = tensor<int32, []>(1)];
252
+ tensor<fp16, [384, 384, 1, 1]> var_433_weight_0_to_fp16 = const()[name = tensor<string, []>("op_433_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322880)))];
253
+ tensor<fp16, [384]> var_433_bias_0_to_fp16 = const()[name = tensor<string, []>("op_433_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9617856)))];
254
+ tensor<fp16, [1, 384, 1, 1500]> var_433_cast_fp16 = conv(bias = var_433_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_433_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_433_cast_fp16")];
255
+ tensor<string, []> k_5_pad_type_0 = const()[name = tensor<string, []>("k_5_pad_type_0"), val = tensor<string, []>("valid")];
256
+ tensor<int32, [2]> k_5_strides_0 = const()[name = tensor<string, []>("k_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
257
+ tensor<int32, [4]> k_5_pad_0 = const()[name = tensor<string, []>("k_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
258
+ tensor<int32, [2]> k_5_dilations_0 = const()[name = tensor<string, []>("k_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
259
+ tensor<int32, []> k_5_groups_0 = const()[name = tensor<string, []>("k_5_groups_0"), val = tensor<int32, []>(1)];
260
+ tensor<fp16, [384, 384, 1, 1]> blocks_2_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9618688)))];
261
+ tensor<fp16, [1, 384, 1, 1500]> k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
262
+ tensor<string, []> var_431_pad_type_0 = const()[name = tensor<string, []>("op_431_pad_type_0"), val = tensor<string, []>("valid")];
263
+ tensor<int32, [2]> var_431_strides_0 = const()[name = tensor<string, []>("op_431_strides_0"), val = tensor<int32, [2]>([1, 1])];
264
+ tensor<int32, [4]> var_431_pad_0 = const()[name = tensor<string, []>("op_431_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
265
+ tensor<int32, [2]> var_431_dilations_0 = const()[name = tensor<string, []>("op_431_dilations_0"), val = tensor<int32, [2]>([1, 1])];
266
+ tensor<int32, []> var_431_groups_0 = const()[name = tensor<string, []>("op_431_groups_0"), val = tensor<int32, []>(1)];
267
+ tensor<fp16, [384, 384, 1, 1]> blocks_2_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9913664)))];
268
+ tensor<fp16, [384]> blocks_2_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10208640)))];
269
+ tensor<fp16, [1, 384, 1, 1500]> var_431_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_431_dilations_0, groups = var_431_groups_0, pad = var_431_pad_0, pad_type = var_431_pad_type_0, strides = var_431_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")];
270
+ tensor<int32, [6]> tile_6 = const()[name = tensor<string, []>("tile_6"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
271
+ tensor<int32, []> var_434_axis_0 = const()[name = tensor<string, []>("op_434_axis_0"), val = tensor<int32, []>(1)];
272
+ tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_5 = split(axis = var_434_axis_0, split_sizes = tile_6, x = var_433_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
273
+ tensor<int32, [4]> var_441_perm_0 = const()[name = tensor<string, []>("op_441_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
274
+ tensor<int32, [6]> tile_7 = const()[name = tensor<string, []>("tile_7"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
275
+ tensor<int32, []> var_442_axis_0 = const()[name = tensor<string, []>("op_442_axis_0"), val = tensor<int32, []>(3)];
276
+ tensor<fp16, [1, 1500, 1, 384]> var_441_cast_fp16 = transpose(perm = var_441_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_2")];
277
+ tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_5 = split(axis = var_442_axis_0, split_sizes = tile_7, x = var_441_cast_fp16)[name = tensor<string, []>("op_442_cast_fp16")];
278
+ tensor<int32, [6]> tile_8 = const()[name = tensor<string, []>("tile_8"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
279
+ tensor<int32, []> var_449_axis_0 = const()[name = tensor<string, []>("op_449_axis_0"), val = tensor<int32, []>(1)];
280
+ tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_5 = split(axis = var_449_axis_0, split_sizes = tile_8, x = var_431_cast_fp16)[name = tensor<string, []>("op_449_cast_fp16")];
281
+ tensor<string, []> aw_25_equation_0 = const()[name = tensor<string, []>("aw_25_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
282
+ tensor<fp16, [1, 1500, 1, 1500]> aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_442_cast_fp16_0, var_434_cast_fp16_0))[name = tensor<string, []>("aw_25_cast_fp16")];
283
+ tensor<string, []> aw_27_equation_0 = const()[name = tensor<string, []>("aw_27_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
284
+ tensor<fp16, [1, 1500, 1, 1500]> aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_442_cast_fp16_1, var_434_cast_fp16_1))[name = tensor<string, []>("aw_27_cast_fp16")];
285
+ tensor<string, []> aw_29_equation_0 = const()[name = tensor<string, []>("aw_29_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
286
+ tensor<fp16, [1, 1500, 1, 1500]> aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_442_cast_fp16_2, var_434_cast_fp16_2))[name = tensor<string, []>("aw_29_cast_fp16")];
287
+ tensor<string, []> aw_31_equation_0 = const()[name = tensor<string, []>("aw_31_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
288
+ tensor<fp16, [1, 1500, 1, 1500]> aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_442_cast_fp16_3, var_434_cast_fp16_3))[name = tensor<string, []>("aw_31_cast_fp16")];
289
+ tensor<string, []> aw_33_equation_0 = const()[name = tensor<string, []>("aw_33_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
290
+ tensor<fp16, [1, 1500, 1, 1500]> aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_442_cast_fp16_4, var_434_cast_fp16_4))[name = tensor<string, []>("aw_33_cast_fp16")];
291
+ tensor<string, []> aw_35_equation_0 = const()[name = tensor<string, []>("aw_35_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
292
+ tensor<fp16, [1, 1500, 1, 1500]> aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_442_cast_fp16_5, var_434_cast_fp16_5))[name = tensor<string, []>("aw_35_cast_fp16")];
293
+ tensor<fp16, [1, 1500, 1, 1500]> var_468_cast_fp16 = softmax(axis = var_382, x = aw_25_cast_fp16)[name = tensor<string, []>("op_468_cast_fp16")];
294
+ tensor<fp16, [1, 1500, 1, 1500]> var_469_cast_fp16 = softmax(axis = var_382, x = aw_27_cast_fp16)[name = tensor<string, []>("op_469_cast_fp16")];
295
+ tensor<fp16, [1, 1500, 1, 1500]> var_470_cast_fp16 = softmax(axis = var_382, x = aw_29_cast_fp16)[name = tensor<string, []>("op_470_cast_fp16")];
296
+ tensor<fp16, [1, 1500, 1, 1500]> var_471_cast_fp16 = softmax(axis = var_382, x = aw_31_cast_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
297
+ tensor<fp16, [1, 1500, 1, 1500]> var_472_cast_fp16 = softmax(axis = var_382, x = aw_33_cast_fp16)[name = tensor<string, []>("op_472_cast_fp16")];
298
+ tensor<fp16, [1, 1500, 1, 1500]> var_473_cast_fp16 = softmax(axis = var_382, x = aw_35_cast_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
299
+ tensor<string, []> var_475_equation_0 = const()[name = tensor<string, []>("op_475_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
300
+ tensor<fp16, [1, 64, 1, 1500]> var_475_cast_fp16 = einsum(equation = var_475_equation_0, values = (var_449_cast_fp16_0, var_468_cast_fp16))[name = tensor<string, []>("op_475_cast_fp16")];
301
+ tensor<string, []> var_477_equation_0 = const()[name = tensor<string, []>("op_477_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
302
+ tensor<fp16, [1, 64, 1, 1500]> var_477_cast_fp16 = einsum(equation = var_477_equation_0, values = (var_449_cast_fp16_1, var_469_cast_fp16))[name = tensor<string, []>("op_477_cast_fp16")];
303
+ tensor<string, []> var_479_equation_0 = const()[name = tensor<string, []>("op_479_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
304
+ tensor<fp16, [1, 64, 1, 1500]> var_479_cast_fp16 = einsum(equation = var_479_equation_0, values = (var_449_cast_fp16_2, var_470_cast_fp16))[name = tensor<string, []>("op_479_cast_fp16")];
305
+ tensor<string, []> var_481_equation_0 = const()[name = tensor<string, []>("op_481_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
306
+ tensor<fp16, [1, 64, 1, 1500]> var_481_cast_fp16 = einsum(equation = var_481_equation_0, values = (var_449_cast_fp16_3, var_471_cast_fp16))[name = tensor<string, []>("op_481_cast_fp16")];
307
+ tensor<string, []> var_483_equation_0 = const()[name = tensor<string, []>("op_483_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
308
+ tensor<fp16, [1, 64, 1, 1500]> var_483_cast_fp16 = einsum(equation = var_483_equation_0, values = (var_449_cast_fp16_4, var_472_cast_fp16))[name = tensor<string, []>("op_483_cast_fp16")];
309
+ tensor<string, []> var_485_equation_0 = const()[name = tensor<string, []>("op_485_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
310
+ tensor<fp16, [1, 64, 1, 1500]> var_485_cast_fp16 = einsum(equation = var_485_equation_0, values = (var_449_cast_fp16_5, var_473_cast_fp16))[name = tensor<string, []>("op_485_cast_fp16")];
311
+ tensor<bool, []> input_25_interleave_0 = const()[name = tensor<string, []>("input_25_interleave_0"), val = tensor<bool, []>(false)];
312
+ tensor<fp16, [1, 384, 1, 1500]> input_25_cast_fp16 = concat(axis = var_382, interleave = input_25_interleave_0, values = (var_475_cast_fp16, var_477_cast_fp16, var_479_cast_fp16, var_481_cast_fp16, var_483_cast_fp16, var_485_cast_fp16))[name = tensor<string, []>("input_25_cast_fp16")];
313
+ tensor<string, []> var_494_pad_type_0 = const()[name = tensor<string, []>("op_494_pad_type_0"), val = tensor<string, []>("valid")];
314
+ tensor<int32, [2]> var_494_strides_0 = const()[name = tensor<string, []>("op_494_strides_0"), val = tensor<int32, [2]>([1, 1])];
315
+ tensor<int32, [4]> var_494_pad_0 = const()[name = tensor<string, []>("op_494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
316
+ tensor<int32, [2]> var_494_dilations_0 = const()[name = tensor<string, []>("op_494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
317
+ tensor<int32, []> var_494_groups_0 = const()[name = tensor<string, []>("op_494_groups_0"), val = tensor<int32, []>(1)];
318
+ tensor<fp16, [384, 384, 1, 1]> blocks_2_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10209472)))];
319
+ tensor<fp16, [384]> blocks_2_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10504448)))];
320
+ tensor<fp16, [1, 384, 1, 1500]> var_494_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_494_dilations_0, groups = var_494_groups_0, pad = var_494_pad_0, pad_type = var_494_pad_type_0, strides = var_494_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("op_494_cast_fp16")];
321
+ tensor<fp16, [1, 384, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_494_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
322
+ tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([1])];
323
+ tensor<fp16, [384]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10505280)))];
324
+ tensor<fp16, [384]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506112)))];
325
+ tensor<fp16, []> var_504_to_fp16 = const()[name = tensor<string, []>("op_504_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
326
+ tensor<fp16, [1, 384, 1, 1500]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_504_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
327
+ tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")];
328
+ tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
329
+ tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
330
+ tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
331
+ tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
332
+ tensor<fp16, [1536, 384, 1, 1]> blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506944)))];
333
+ tensor<fp16, [1536]> blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11686656)))];
334
+ tensor<fp16, [1, 1536, 1, 1500]> input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
335
+ tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")];
336
+ tensor<fp16, [1, 1536, 1, 1500]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
337
+ tensor<string, []> var_530_pad_type_0 = const()[name = tensor<string, []>("op_530_pad_type_0"), val = tensor<string, []>("valid")];
338
+ tensor<int32, [2]> var_530_strides_0 = const()[name = tensor<string, []>("op_530_strides_0"), val = tensor<int32, [2]>([1, 1])];
339
+ tensor<int32, [4]> var_530_pad_0 = const()[name = tensor<string, []>("op_530_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
340
+ tensor<int32, [2]> var_530_dilations_0 = const()[name = tensor<string, []>("op_530_dilations_0"), val = tensor<int32, [2]>([1, 1])];
341
+ tensor<int32, []> var_530_groups_0 = const()[name = tensor<string, []>("op_530_groups_0"), val = tensor<int32, []>(1)];
342
+ tensor<fp16, [384, 1536, 1, 1]> blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11689792)))];
343
+ tensor<fp16, [384]> blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12869504)))];
344
+ tensor<fp16, [1, 384, 1, 1500]> var_530_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_530_dilations_0, groups = var_530_groups_0, pad = var_530_pad_0, pad_type = var_530_pad_type_0, strides = var_530_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("op_530_cast_fp16")];
345
+ tensor<fp16, [1, 384, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_530_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
346
+ tensor<int32, []> var_539 = const()[name = tensor<string, []>("op_539"), val = tensor<int32, []>(1)];
347
+ tensor<int32, [1]> input_33_axes_0 = const()[name = tensor<string, []>("input_33_axes_0"), val = tensor<int32, [1]>([1])];
348
+ tensor<fp16, [384]> input_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_33_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12870336)))];
349
+ tensor<fp16, [384]> input_33_beta_0_to_fp16 = const()[name = tensor<string, []>("input_33_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12871168)))];
350
+ tensor<fp16, []> var_555_to_fp16 = const()[name = tensor<string, []>("op_555_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
351
+ tensor<fp16, [1, 384, 1, 1500]> input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_555_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
352
+ tensor<string, []> q_pad_type_0 = const()[name = tensor<string, []>("q_pad_type_0"), val = tensor<string, []>("valid")];
353
+ tensor<int32, [2]> q_strides_0 = const()[name = tensor<string, []>("q_strides_0"), val = tensor<int32, [2]>([1, 1])];
354
+ tensor<int32, [4]> q_pad_0 = const()[name = tensor<string, []>("q_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
355
+ tensor<int32, [2]> q_dilations_0 = const()[name = tensor<string, []>("q_dilations_0"), val = tensor<int32, [2]>([1, 1])];
356
+ tensor<int32, []> q_groups_0 = const()[name = tensor<string, []>("q_groups_0"), val = tensor<int32, []>(1)];
357
+ tensor<fp16, [384, 384, 1, 1]> var_590_weight_0_to_fp16 = const()[name = tensor<string, []>("op_590_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12872000)))];
358
+ tensor<fp16, [384]> var_590_bias_0_to_fp16 = const()[name = tensor<string, []>("op_590_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13166976)))];
359
+ tensor<fp16, [1, 384, 1, 1500]> var_590_cast_fp16 = conv(bias = var_590_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_590_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_590_cast_fp16")];
360
+ tensor<string, []> k_pad_type_0 = const()[name = tensor<string, []>("k_pad_type_0"), val = tensor<string, []>("valid")];
361
+ tensor<int32, [2]> k_strides_0 = const()[name = tensor<string, []>("k_strides_0"), val = tensor<int32, [2]>([1, 1])];
362
+ tensor<int32, [4]> k_pad_0 = const()[name = tensor<string, []>("k_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
363
+ tensor<int32, [2]> k_dilations_0 = const()[name = tensor<string, []>("k_dilations_0"), val = tensor<int32, [2]>([1, 1])];
364
+ tensor<int32, []> k_groups_0 = const()[name = tensor<string, []>("k_groups_0"), val = tensor<int32, []>(1)];
365
+ tensor<fp16, [384, 384, 1, 1]> blocks_3_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13167808)))];
366
+ tensor<fp16, [1, 384, 1, 1500]> k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
367
+ tensor<string, []> var_588_pad_type_0 = const()[name = tensor<string, []>("op_588_pad_type_0"), val = tensor<string, []>("valid")];
368
+ tensor<int32, [2]> var_588_strides_0 = const()[name = tensor<string, []>("op_588_strides_0"), val = tensor<int32, [2]>([1, 1])];
369
+ tensor<int32, [4]> var_588_pad_0 = const()[name = tensor<string, []>("op_588_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
370
+ tensor<int32, [2]> var_588_dilations_0 = const()[name = tensor<string, []>("op_588_dilations_0"), val = tensor<int32, [2]>([1, 1])];
371
+ tensor<int32, []> var_588_groups_0 = const()[name = tensor<string, []>("op_588_groups_0"), val = tensor<int32, []>(1)];
372
+ tensor<fp16, [384, 384, 1, 1]> blocks_3_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13462784)))];
373
+ tensor<fp16, [384]> blocks_3_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13757760)))];
374
+ tensor<fp16, [1, 384, 1, 1500]> var_588_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_588_dilations_0, groups = var_588_groups_0, pad = var_588_pad_0, pad_type = var_588_pad_type_0, strides = var_588_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_588_cast_fp16")];
375
+ tensor<int32, [6]> tile_9 = const()[name = tensor<string, []>("tile_9"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
376
+ tensor<int32, []> var_591_axis_0 = const()[name = tensor<string, []>("op_591_axis_0"), val = tensor<int32, []>(1)];
377
+ tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_5 = split(axis = var_591_axis_0, split_sizes = tile_9, x = var_590_cast_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
378
+ tensor<int32, [4]> var_598_perm_0 = const()[name = tensor<string, []>("op_598_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
379
+ tensor<int32, [6]> tile_10 = const()[name = tensor<string, []>("tile_10"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
380
+ tensor<int32, []> var_599_axis_0 = const()[name = tensor<string, []>("op_599_axis_0"), val = tensor<int32, []>(3)];
381
+ tensor<fp16, [1, 1500, 1, 384]> var_598_cast_fp16 = transpose(perm = var_598_perm_0, x = k_cast_fp16)[name = tensor<string, []>("transpose_1")];
382
+ tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_5 = split(axis = var_599_axis_0, split_sizes = tile_10, x = var_598_cast_fp16)[name = tensor<string, []>("op_599_cast_fp16")];
383
+ tensor<int32, [6]> tile_11 = const()[name = tensor<string, []>("tile_11"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
384
+ tensor<int32, []> var_606_axis_0 = const()[name = tensor<string, []>("op_606_axis_0"), val = tensor<int32, []>(1)];
385
+ tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_5 = split(axis = var_606_axis_0, split_sizes = tile_11, x = var_588_cast_fp16)[name = tensor<string, []>("op_606_cast_fp16")];
386
+ tensor<string, []> aw_37_equation_0 = const()[name = tensor<string, []>("aw_37_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
387
+ tensor<fp16, [1, 1500, 1, 1500]> aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_599_cast_fp16_0, var_591_cast_fp16_0))[name = tensor<string, []>("aw_37_cast_fp16")];
388
+ tensor<string, []> aw_39_equation_0 = const()[name = tensor<string, []>("aw_39_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
389
+ tensor<fp16, [1, 1500, 1, 1500]> aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_599_cast_fp16_1, var_591_cast_fp16_1))[name = tensor<string, []>("aw_39_cast_fp16")];
390
+ tensor<string, []> aw_41_equation_0 = const()[name = tensor<string, []>("aw_41_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
391
+ tensor<fp16, [1, 1500, 1, 1500]> aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_599_cast_fp16_2, var_591_cast_fp16_2))[name = tensor<string, []>("aw_41_cast_fp16")];
392
+ tensor<string, []> aw_43_equation_0 = const()[name = tensor<string, []>("aw_43_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
393
+ tensor<fp16, [1, 1500, 1, 1500]> aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_599_cast_fp16_3, var_591_cast_fp16_3))[name = tensor<string, []>("aw_43_cast_fp16")];
394
+ tensor<string, []> aw_45_equation_0 = const()[name = tensor<string, []>("aw_45_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
395
+ tensor<fp16, [1, 1500, 1, 1500]> aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_599_cast_fp16_4, var_591_cast_fp16_4))[name = tensor<string, []>("aw_45_cast_fp16")];
396
+ tensor<string, []> aw_equation_0 = const()[name = tensor<string, []>("aw_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
397
+ tensor<fp16, [1, 1500, 1, 1500]> aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_599_cast_fp16_5, var_591_cast_fp16_5))[name = tensor<string, []>("aw_cast_fp16")];
398
+ tensor<fp16, [1, 1500, 1, 1500]> var_625_cast_fp16 = softmax(axis = var_539, x = aw_37_cast_fp16)[name = tensor<string, []>("op_625_cast_fp16")];
399
+ tensor<fp16, [1, 1500, 1, 1500]> var_626_cast_fp16 = softmax(axis = var_539, x = aw_39_cast_fp16)[name = tensor<string, []>("op_626_cast_fp16")];
400
+ tensor<fp16, [1, 1500, 1, 1500]> var_627_cast_fp16 = softmax(axis = var_539, x = aw_41_cast_fp16)[name = tensor<string, []>("op_627_cast_fp16")];
401
+ tensor<fp16, [1, 1500, 1, 1500]> var_628_cast_fp16 = softmax(axis = var_539, x = aw_43_cast_fp16)[name = tensor<string, []>("op_628_cast_fp16")];
402
+ tensor<fp16, [1, 1500, 1, 1500]> var_629_cast_fp16 = softmax(axis = var_539, x = aw_45_cast_fp16)[name = tensor<string, []>("op_629_cast_fp16")];
403
+ tensor<fp16, [1, 1500, 1, 1500]> var_630_cast_fp16 = softmax(axis = var_539, x = aw_cast_fp16)[name = tensor<string, []>("op_630_cast_fp16")];
404
+ tensor<string, []> var_632_equation_0 = const()[name = tensor<string, []>("op_632_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
405
+ tensor<fp16, [1, 64, 1, 1500]> var_632_cast_fp16 = einsum(equation = var_632_equation_0, values = (var_606_cast_fp16_0, var_625_cast_fp16))[name = tensor<string, []>("op_632_cast_fp16")];
406
+ tensor<string, []> var_634_equation_0 = const()[name = tensor<string, []>("op_634_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
407
+ tensor<fp16, [1, 64, 1, 1500]> var_634_cast_fp16 = einsum(equation = var_634_equation_0, values = (var_606_cast_fp16_1, var_626_cast_fp16))[name = tensor<string, []>("op_634_cast_fp16")];
408
+ tensor<string, []> var_636_equation_0 = const()[name = tensor<string, []>("op_636_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
409
+ tensor<fp16, [1, 64, 1, 1500]> var_636_cast_fp16 = einsum(equation = var_636_equation_0, values = (var_606_cast_fp16_2, var_627_cast_fp16))[name = tensor<string, []>("op_636_cast_fp16")];
410
+ tensor<string, []> var_638_equation_0 = const()[name = tensor<string, []>("op_638_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
411
+ tensor<fp16, [1, 64, 1, 1500]> var_638_cast_fp16 = einsum(equation = var_638_equation_0, values = (var_606_cast_fp16_3, var_628_cast_fp16))[name = tensor<string, []>("op_638_cast_fp16")];
412
+ tensor<string, []> var_640_equation_0 = const()[name = tensor<string, []>("op_640_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
413
+ tensor<fp16, [1, 64, 1, 1500]> var_640_cast_fp16 = einsum(equation = var_640_equation_0, values = (var_606_cast_fp16_4, var_629_cast_fp16))[name = tensor<string, []>("op_640_cast_fp16")];
414
+ tensor<string, []> var_642_equation_0 = const()[name = tensor<string, []>("op_642_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
415
+ tensor<fp16, [1, 64, 1, 1500]> var_642_cast_fp16 = einsum(equation = var_642_equation_0, values = (var_606_cast_fp16_5, var_630_cast_fp16))[name = tensor<string, []>("op_642_cast_fp16")];
416
+ tensor<bool, []> input_35_interleave_0 = const()[name = tensor<string, []>("input_35_interleave_0"), val = tensor<bool, []>(false)];
417
+ tensor<fp16, [1, 384, 1, 1500]> input_35_cast_fp16 = concat(axis = var_539, interleave = input_35_interleave_0, values = (var_632_cast_fp16, var_634_cast_fp16, var_636_cast_fp16, var_638_cast_fp16, var_640_cast_fp16, var_642_cast_fp16))[name = tensor<string, []>("input_35_cast_fp16")];
418
+ tensor<string, []> var_651_pad_type_0 = const()[name = tensor<string, []>("op_651_pad_type_0"), val = tensor<string, []>("valid")];
419
+ tensor<int32, [2]> var_651_strides_0 = const()[name = tensor<string, []>("op_651_strides_0"), val = tensor<int32, [2]>([1, 1])];
420
+ tensor<int32, [4]> var_651_pad_0 = const()[name = tensor<string, []>("op_651_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
421
+ tensor<int32, [2]> var_651_dilations_0 = const()[name = tensor<string, []>("op_651_dilations_0"), val = tensor<int32, [2]>([1, 1])];
422
+ tensor<int32, []> var_651_groups_0 = const()[name = tensor<string, []>("op_651_groups_0"), val = tensor<int32, []>(1)];
423
+ tensor<fp16, [384, 384, 1, 1]> blocks_3_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13758592)))];
424
+ tensor<fp16, [384]> blocks_3_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14053568)))];
425
+ tensor<fp16, [1, 384, 1, 1500]> var_651_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_651_dilations_0, groups = var_651_groups_0, pad = var_651_pad_0, pad_type = var_651_pad_type_0, strides = var_651_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("op_651_cast_fp16")];
426
+ tensor<fp16, [1, 384, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_651_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
427
+ tensor<int32, [1]> input_37_axes_0 = const()[name = tensor<string, []>("input_37_axes_0"), val = tensor<int32, [1]>([1])];
428
+ tensor<fp16, [384]> input_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14054400)))];
429
+ tensor<fp16, [384]> input_37_beta_0_to_fp16 = const()[name = tensor<string, []>("input_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14055232)))];
430
+ tensor<fp16, []> var_661_to_fp16 = const()[name = tensor<string, []>("op_661_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
431
+ tensor<fp16, [1, 384, 1, 1500]> input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_661_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
432
+ tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("valid")];
433
+ tensor<int32, [2]> input_39_strides_0 = const()[name = tensor<string, []>("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
434
+ tensor<int32, [4]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
435
+ tensor<int32, [2]> input_39_dilations_0 = const()[name = tensor<string, []>("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
436
+ tensor<int32, []> input_39_groups_0 = const()[name = tensor<string, []>("input_39_groups_0"), val = tensor<int32, []>(1)];
437
+ tensor<fp16, [1536, 384, 1, 1]> blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14056064)))];
438
+ tensor<fp16, [1536]> blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15235776)))];
439
+ tensor<fp16, [1, 1536, 1, 1500]> input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
440
+ tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
441
+ tensor<fp16, [1, 1536, 1, 1500]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_39_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
442
+ tensor<string, []> var_687_pad_type_0 = const()[name = tensor<string, []>("op_687_pad_type_0"), val = tensor<string, []>("valid")];
443
+ tensor<int32, [2]> var_687_strides_0 = const()[name = tensor<string, []>("op_687_strides_0"), val = tensor<int32, [2]>([1, 1])];
444
+ tensor<int32, [4]> var_687_pad_0 = const()[name = tensor<string, []>("op_687_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
445
+ tensor<int32, [2]> var_687_dilations_0 = const()[name = tensor<string, []>("op_687_dilations_0"), val = tensor<int32, [2]>([1, 1])];
446
+ tensor<int32, []> var_687_groups_0 = const()[name = tensor<string, []>("op_687_groups_0"), val = tensor<int32, []>(1)];
447
+ tensor<fp16, [384, 1536, 1, 1]> blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15238912)))];
448
+ tensor<fp16, [384]> blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16418624)))];
449
+ tensor<fp16, [1, 384, 1, 1500]> var_687_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_687_dilations_0, groups = var_687_groups_0, pad = var_687_pad_0, pad_type = var_687_pad_type_0, strides = var_687_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("op_687_cast_fp16")];
450
+ tensor<fp16, [1, 384, 1, 1500]> inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_687_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
451
+ tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([1])];
452
+ tensor<fp16, [384]> x_gamma_0_to_fp16 = const()[name = tensor<string, []>("x_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16419456)))];
453
+ tensor<fp16, [384]> x_beta_0_to_fp16 = const()[name = tensor<string, []>("x_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16420288)))];
454
+ tensor<fp16, []> var_701_to_fp16 = const()[name = tensor<string, []>("op_701_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
455
+ tensor<fp16, [1, 384, 1, 1500]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_701_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
456
+ tensor<int32, [1]> var_712_axes_0 = const()[name = tensor<string, []>("op_712_axes_0"), val = tensor<int32, [1]>([2])];
457
+ tensor<fp16, [1, 384, 1500]> var_712_cast_fp16 = squeeze(axes = var_712_axes_0, x = x_cast_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
458
+ tensor<int32, [3]> var_715_perm_0 = const()[name = tensor<string, []>("op_715_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
459
+ tensor<string, []> var_715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_715_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
460
+ tensor<fp16, [1, 1500, 384]> var_715_cast_fp16 = transpose(perm = var_715_perm_0, x = var_712_cast_fp16)[name = tensor<string, []>("transpose_0")];
461
+ tensor<fp32, [1, 1500, 384]> output = cast(dtype = var_715_cast_fp16_to_fp32_dtype_0, x = var_715_cast_fp16)[name = tensor<string, []>("cast_19")];
462
+ } -> (output);
463
+ }
tiny.en/ggml-tiny.en-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:040cc1dc03624b30f9f01e567d71b651729da26d98de36c72ab3266c85f68fab
3
+ size 16421120
tiny.en/ggml-tiny.en.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:921e4cf8686fdd993dcd081a5da5b6c365bfde1162e72b08d75ac75289920b1f
3
+ size 77704715
tiny/ggml-tiny-encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c811ab131cbe7fef6230b32b61cb04cc99fb5990e5fb70ab5d7ec907a4a124b2
3
+ size 243
tiny/ggml-tiny-encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9df7e387da981d548738ae3570db8a375e2e55d16ae8e1374da84809fdd0c5a
3
+ size 320
tiny/ggml-tiny-encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float16",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32 1 × 1500 × 384)",
11
+ "shortDescription" : "",
12
+ "shape" : "[1, 1500, 384]",
13
+ "name" : "output",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 6,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Concat" : 4,
23
+ "Gelu" : 6,
24
+ "LayerNorm" : 9,
25
+ "Transpose" : 5,
26
+ "Softmax" : 24,
27
+ "Squeeze" : 1,
28
+ "Cast" : 2,
29
+ "Add" : 9,
30
+ "Einsum" : 48,
31
+ "ExpandDims" : 1,
32
+ "Split" : 12,
33
+ "Conv" : 26
34
+ },
35
+ "computePrecision" : "Mixed (Float16, Float32, Int32)",
36
+ "isUpdatable" : "0",
37
+ "stateSchema" : [
38
+
39
+ ],
40
+ "availability" : {
41
+ "macOS" : "12.0",
42
+ "tvOS" : "15.0",
43
+ "visionOS" : "1.0",
44
+ "watchOS" : "8.0",
45
+ "iOS" : "15.0",
46
+ "macCatalyst" : "15.0"
47
+ },
48
+ "modelType" : {
49
+ "name" : "MLModelType_mlProgram"
50
+ },
51
+ "userDefinedMetadata" : {
52
+ "com.github.apple.coremltools.source_dialect" : "TorchScript",
53
+ "com.github.apple.coremltools.version" : "8.3.0",
54
+ "com.github.apple.coremltools.source" : "torch==2.2.2"
55
+ },
56
+ "inputSchema" : [
57
+ {
58
+ "hasShapeFlexibility" : "0",
59
+ "isOptional" : "0",
60
+ "dataType" : "Float32",
61
+ "formattedType" : "MultiArray (Float32 1 × 80 × 3000)",
62
+ "shortDescription" : "",
63
+ "shape" : "[1, 80, 3000]",
64
+ "name" : "logmel_data",
65
+ "type" : "MultiArray"
66
+ }
67
+ ],
68
+ "generatedClassName" : "coreml_encoder_tiny",
69
+ "method" : "predict"
70
+ }
71
+ ]
tiny/ggml-tiny-encoder.mlmodelc/model.mil ADDED
@@ -0,0 +1,463 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ program(1.0)
2
+ [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.2.2"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
3
+ {
4
+ func main<ios15>(tensor<fp32, [1, 80, 3000]> logmel_data) {
5
+ tensor<string, []> var_28_pad_type_0 = const()[name = tensor<string, []>("op_28_pad_type_0"), val = tensor<string, []>("custom")];
6
+ tensor<int32, [2]> var_28_pad_0 = const()[name = tensor<string, []>("op_28_pad_0"), val = tensor<int32, [2]>([1, 1])];
7
+ tensor<int32, [1]> var_28_strides_0 = const()[name = tensor<string, []>("op_28_strides_0"), val = tensor<int32, [1]>([1])];
8
+ tensor<int32, [1]> var_28_dilations_0 = const()[name = tensor<string, []>("op_28_dilations_0"), val = tensor<int32, [1]>([1])];
9
+ tensor<int32, []> var_28_groups_0 = const()[name = tensor<string, []>("op_28_groups_0"), val = tensor<int32, []>(1)];
10
+ tensor<string, []> logmel_data_to_fp16_dtype_0 = const()[name = tensor<string, []>("logmel_data_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
11
+ tensor<fp16, [384, 80, 3]> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, [384, 80, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
12
+ tensor<fp16, [384]> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184448)))];
13
+ tensor<fp16, [1, 80, 3000]> logmel_data_to_fp16 = cast(dtype = logmel_data_to_fp16_dtype_0, x = logmel_data)[name = tensor<string, []>("cast_20")];
14
+ tensor<fp16, [1, 384, 3000]> var_28_cast_fp16 = conv(bias = const_1_to_fp16, dilations = var_28_dilations_0, groups = var_28_groups_0, pad = var_28_pad_0, pad_type = var_28_pad_type_0, strides = var_28_strides_0, weight = const_0_to_fp16, x = logmel_data_to_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
15
+ tensor<string, []> input_1_mode_0 = const()[name = tensor<string, []>("input_1_mode_0"), val = tensor<string, []>("EXACT")];
16
+ tensor<fp16, [1, 384, 3000]> input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_28_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
17
+ tensor<string, []> var_46_pad_type_0 = const()[name = tensor<string, []>("op_46_pad_type_0"), val = tensor<string, []>("custom")];
18
+ tensor<int32, [2]> var_46_pad_0 = const()[name = tensor<string, []>("op_46_pad_0"), val = tensor<int32, [2]>([1, 1])];
19
+ tensor<int32, [1]> var_46_strides_0 = const()[name = tensor<string, []>("op_46_strides_0"), val = tensor<int32, [1]>([2])];
20
+ tensor<int32, [1]> var_46_dilations_0 = const()[name = tensor<string, []>("op_46_dilations_0"), val = tensor<int32, [1]>([1])];
21
+ tensor<int32, []> var_46_groups_0 = const()[name = tensor<string, []>("op_46_groups_0"), val = tensor<int32, []>(1)];
22
+ tensor<fp16, [384, 384, 3]> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, [384, 384, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185280)))];
23
+ tensor<fp16, [384]> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1070080)))];
24
+ tensor<fp16, [1, 384, 1500]> var_46_cast_fp16 = conv(bias = const_3_to_fp16, dilations = var_46_dilations_0, groups = var_46_groups_0, pad = var_46_pad_0, pad_type = var_46_pad_type_0, strides = var_46_strides_0, weight = const_2_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("op_46_cast_fp16")];
25
+ tensor<string, []> x_3_mode_0 = const()[name = tensor<string, []>("x_3_mode_0"), val = tensor<string, []>("EXACT")];
26
+ tensor<fp16, [1, 384, 1500]> x_3_cast_fp16 = gelu(mode = x_3_mode_0, x = var_46_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
27
+ tensor<fp16, [384, 1500]> var_51_to_fp16 = const()[name = tensor<string, []>("op_51_to_fp16"), val = tensor<fp16, [384, 1500]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1070912)))];
28
+ tensor<fp16, [1, 384, 1500]> var_53_cast_fp16 = add(x = x_3_cast_fp16, y = var_51_to_fp16)[name = tensor<string, []>("op_53_cast_fp16")];
29
+ tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([2])];
30
+ tensor<fp16, [1, 384, 1, 1500]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_53_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")];
31
+ tensor<int32, []> var_68 = const()[name = tensor<string, []>("op_68"), val = tensor<int32, []>(1)];
32
+ tensor<int32, [1]> input_3_axes_0 = const()[name = tensor<string, []>("input_3_axes_0"), val = tensor<int32, [1]>([1])];
33
+ tensor<fp16, [384]> input_3_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_3_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2222976)))];
34
+ tensor<fp16, [384]> input_3_beta_0_to_fp16 = const()[name = tensor<string, []>("input_3_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2223808)))];
35
+ tensor<fp16, []> var_84_to_fp16 = const()[name = tensor<string, []>("op_84_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
36
+ tensor<fp16, [1, 384, 1, 1500]> input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = input_3_beta_0_to_fp16, epsilon = var_84_to_fp16, gamma = input_3_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
37
+ tensor<string, []> q_1_pad_type_0 = const()[name = tensor<string, []>("q_1_pad_type_0"), val = tensor<string, []>("valid")];
38
+ tensor<int32, [2]> q_1_strides_0 = const()[name = tensor<string, []>("q_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
39
+ tensor<int32, [4]> q_1_pad_0 = const()[name = tensor<string, []>("q_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
40
+ tensor<int32, [2]> q_1_dilations_0 = const()[name = tensor<string, []>("q_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
41
+ tensor<int32, []> q_1_groups_0 = const()[name = tensor<string, []>("q_1_groups_0"), val = tensor<int32, []>(1)];
42
+ tensor<fp16, [384, 384, 1, 1]> var_119_weight_0_to_fp16 = const()[name = tensor<string, []>("op_119_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2224640)))];
43
+ tensor<fp16, [384]> var_119_bias_0_to_fp16 = const()[name = tensor<string, []>("op_119_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2519616)))];
44
+ tensor<fp16, [1, 384, 1, 1500]> var_119_cast_fp16 = conv(bias = var_119_bias_0_to_fp16, dilations = q_1_dilations_0, groups = q_1_groups_0, pad = q_1_pad_0, pad_type = q_1_pad_type_0, strides = q_1_strides_0, weight = var_119_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_119_cast_fp16")];
45
+ tensor<string, []> k_1_pad_type_0 = const()[name = tensor<string, []>("k_1_pad_type_0"), val = tensor<string, []>("valid")];
46
+ tensor<int32, [2]> k_1_strides_0 = const()[name = tensor<string, []>("k_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
47
+ tensor<int32, [4]> k_1_pad_0 = const()[name = tensor<string, []>("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
48
+ tensor<int32, [2]> k_1_dilations_0 = const()[name = tensor<string, []>("k_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
49
+ tensor<int32, []> k_1_groups_0 = const()[name = tensor<string, []>("k_1_groups_0"), val = tensor<int32, []>(1)];
50
+ tensor<fp16, [384, 384, 1, 1]> blocks_0_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2520448)))];
51
+ tensor<fp16, [1, 384, 1, 1500]> k_1_cast_fp16 = conv(dilations = k_1_dilations_0, groups = k_1_groups_0, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = k_1_strides_0, weight = blocks_0_attn_key_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
52
+ tensor<string, []> var_117_pad_type_0 = const()[name = tensor<string, []>("op_117_pad_type_0"), val = tensor<string, []>("valid")];
53
+ tensor<int32, [2]> var_117_strides_0 = const()[name = tensor<string, []>("op_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
54
+ tensor<int32, [4]> var_117_pad_0 = const()[name = tensor<string, []>("op_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
55
+ tensor<int32, [2]> var_117_dilations_0 = const()[name = tensor<string, []>("op_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
56
+ tensor<int32, []> var_117_groups_0 = const()[name = tensor<string, []>("op_117_groups_0"), val = tensor<int32, []>(1)];
57
+ tensor<fp16, [384, 384, 1, 1]> blocks_0_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2815424)))];
58
+ tensor<fp16, [384]> blocks_0_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3110400)))];
59
+ tensor<fp16, [1, 384, 1, 1500]> var_117_cast_fp16 = conv(bias = blocks_0_attn_value_bias_to_fp16, dilations = var_117_dilations_0, groups = var_117_groups_0, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_117_strides_0, weight = blocks_0_attn_value_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("op_117_cast_fp16")];
60
+ tensor<int32, [6]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
61
+ tensor<int32, []> var_120_axis_0 = const()[name = tensor<string, []>("op_120_axis_0"), val = tensor<int32, []>(1)];
62
+ tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_120_cast_fp16_5 = split(axis = var_120_axis_0, split_sizes = tile_0, x = var_119_cast_fp16)[name = tensor<string, []>("op_120_cast_fp16")];
63
+ tensor<int32, [4]> var_127_perm_0 = const()[name = tensor<string, []>("op_127_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
64
+ tensor<int32, [6]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
65
+ tensor<int32, []> var_128_axis_0 = const()[name = tensor<string, []>("op_128_axis_0"), val = tensor<int32, []>(3)];
66
+ tensor<fp16, [1, 1500, 1, 384]> var_127_cast_fp16 = transpose(perm = var_127_perm_0, x = k_1_cast_fp16)[name = tensor<string, []>("transpose_4")];
67
+ tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_128_cast_fp16_5 = split(axis = var_128_axis_0, split_sizes = tile_1, x = var_127_cast_fp16)[name = tensor<string, []>("op_128_cast_fp16")];
68
+ tensor<int32, [6]> tile_2 = const()[name = tensor<string, []>("tile_2"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
69
+ tensor<int32, []> var_135_axis_0 = const()[name = tensor<string, []>("op_135_axis_0"), val = tensor<int32, []>(1)];
70
+ tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_135_cast_fp16_5 = split(axis = var_135_axis_0, split_sizes = tile_2, x = var_117_cast_fp16)[name = tensor<string, []>("op_135_cast_fp16")];
71
+ tensor<string, []> aw_1_equation_0 = const()[name = tensor<string, []>("aw_1_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
72
+ tensor<fp16, [1, 1500, 1, 1500]> aw_1_cast_fp16 = einsum(equation = aw_1_equation_0, values = (var_128_cast_fp16_0, var_120_cast_fp16_0))[name = tensor<string, []>("aw_1_cast_fp16")];
73
+ tensor<string, []> aw_3_equation_0 = const()[name = tensor<string, []>("aw_3_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
74
+ tensor<fp16, [1, 1500, 1, 1500]> aw_3_cast_fp16 = einsum(equation = aw_3_equation_0, values = (var_128_cast_fp16_1, var_120_cast_fp16_1))[name = tensor<string, []>("aw_3_cast_fp16")];
75
+ tensor<string, []> aw_5_equation_0 = const()[name = tensor<string, []>("aw_5_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
76
+ tensor<fp16, [1, 1500, 1, 1500]> aw_5_cast_fp16 = einsum(equation = aw_5_equation_0, values = (var_128_cast_fp16_2, var_120_cast_fp16_2))[name = tensor<string, []>("aw_5_cast_fp16")];
77
+ tensor<string, []> aw_7_equation_0 = const()[name = tensor<string, []>("aw_7_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
78
+ tensor<fp16, [1, 1500, 1, 1500]> aw_7_cast_fp16 = einsum(equation = aw_7_equation_0, values = (var_128_cast_fp16_3, var_120_cast_fp16_3))[name = tensor<string, []>("aw_7_cast_fp16")];
79
+ tensor<string, []> aw_9_equation_0 = const()[name = tensor<string, []>("aw_9_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
80
+ tensor<fp16, [1, 1500, 1, 1500]> aw_9_cast_fp16 = einsum(equation = aw_9_equation_0, values = (var_128_cast_fp16_4, var_120_cast_fp16_4))[name = tensor<string, []>("aw_9_cast_fp16")];
81
+ tensor<string, []> aw_11_equation_0 = const()[name = tensor<string, []>("aw_11_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
82
+ tensor<fp16, [1, 1500, 1, 1500]> aw_11_cast_fp16 = einsum(equation = aw_11_equation_0, values = (var_128_cast_fp16_5, var_120_cast_fp16_5))[name = tensor<string, []>("aw_11_cast_fp16")];
83
+ tensor<fp16, [1, 1500, 1, 1500]> var_154_cast_fp16 = softmax(axis = var_68, x = aw_1_cast_fp16)[name = tensor<string, []>("op_154_cast_fp16")];
84
+ tensor<fp16, [1, 1500, 1, 1500]> var_155_cast_fp16 = softmax(axis = var_68, x = aw_3_cast_fp16)[name = tensor<string, []>("op_155_cast_fp16")];
85
+ tensor<fp16, [1, 1500, 1, 1500]> var_156_cast_fp16 = softmax(axis = var_68, x = aw_5_cast_fp16)[name = tensor<string, []>("op_156_cast_fp16")];
86
+ tensor<fp16, [1, 1500, 1, 1500]> var_157_cast_fp16 = softmax(axis = var_68, x = aw_7_cast_fp16)[name = tensor<string, []>("op_157_cast_fp16")];
87
+ tensor<fp16, [1, 1500, 1, 1500]> var_158_cast_fp16 = softmax(axis = var_68, x = aw_9_cast_fp16)[name = tensor<string, []>("op_158_cast_fp16")];
88
+ tensor<fp16, [1, 1500, 1, 1500]> var_159_cast_fp16 = softmax(axis = var_68, x = aw_11_cast_fp16)[name = tensor<string, []>("op_159_cast_fp16")];
89
+ tensor<string, []> var_161_equation_0 = const()[name = tensor<string, []>("op_161_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
90
+ tensor<fp16, [1, 64, 1, 1500]> var_161_cast_fp16 = einsum(equation = var_161_equation_0, values = (var_135_cast_fp16_0, var_154_cast_fp16))[name = tensor<string, []>("op_161_cast_fp16")];
91
+ tensor<string, []> var_163_equation_0 = const()[name = tensor<string, []>("op_163_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
92
+ tensor<fp16, [1, 64, 1, 1500]> var_163_cast_fp16 = einsum(equation = var_163_equation_0, values = (var_135_cast_fp16_1, var_155_cast_fp16))[name = tensor<string, []>("op_163_cast_fp16")];
93
+ tensor<string, []> var_165_equation_0 = const()[name = tensor<string, []>("op_165_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
94
+ tensor<fp16, [1, 64, 1, 1500]> var_165_cast_fp16 = einsum(equation = var_165_equation_0, values = (var_135_cast_fp16_2, var_156_cast_fp16))[name = tensor<string, []>("op_165_cast_fp16")];
95
+ tensor<string, []> var_167_equation_0 = const()[name = tensor<string, []>("op_167_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
96
+ tensor<fp16, [1, 64, 1, 1500]> var_167_cast_fp16 = einsum(equation = var_167_equation_0, values = (var_135_cast_fp16_3, var_157_cast_fp16))[name = tensor<string, []>("op_167_cast_fp16")];
97
+ tensor<string, []> var_169_equation_0 = const()[name = tensor<string, []>("op_169_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
98
+ tensor<fp16, [1, 64, 1, 1500]> var_169_cast_fp16 = einsum(equation = var_169_equation_0, values = (var_135_cast_fp16_4, var_158_cast_fp16))[name = tensor<string, []>("op_169_cast_fp16")];
99
+ tensor<string, []> var_171_equation_0 = const()[name = tensor<string, []>("op_171_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
100
+ tensor<fp16, [1, 64, 1, 1500]> var_171_cast_fp16 = einsum(equation = var_171_equation_0, values = (var_135_cast_fp16_5, var_159_cast_fp16))[name = tensor<string, []>("op_171_cast_fp16")];
101
+ tensor<bool, []> input_5_interleave_0 = const()[name = tensor<string, []>("input_5_interleave_0"), val = tensor<bool, []>(false)];
102
+ tensor<fp16, [1, 384, 1, 1500]> input_5_cast_fp16 = concat(axis = var_68, interleave = input_5_interleave_0, values = (var_161_cast_fp16, var_163_cast_fp16, var_165_cast_fp16, var_167_cast_fp16, var_169_cast_fp16, var_171_cast_fp16))[name = tensor<string, []>("input_5_cast_fp16")];
103
+ tensor<string, []> var_180_pad_type_0 = const()[name = tensor<string, []>("op_180_pad_type_0"), val = tensor<string, []>("valid")];
104
+ tensor<int32, [2]> var_180_strides_0 = const()[name = tensor<string, []>("op_180_strides_0"), val = tensor<int32, [2]>([1, 1])];
105
+ tensor<int32, [4]> var_180_pad_0 = const()[name = tensor<string, []>("op_180_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
106
+ tensor<int32, [2]> var_180_dilations_0 = const()[name = tensor<string, []>("op_180_dilations_0"), val = tensor<int32, [2]>([1, 1])];
107
+ tensor<int32, []> var_180_groups_0 = const()[name = tensor<string, []>("op_180_groups_0"), val = tensor<int32, []>(1)];
108
+ tensor<fp16, [384, 384, 1, 1]> blocks_0_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3111232)))];
109
+ tensor<fp16, [384]> blocks_0_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3406208)))];
110
+ tensor<fp16, [1, 384, 1, 1500]> var_180_cast_fp16 = conv(bias = blocks_0_attn_out_bias_to_fp16, dilations = var_180_dilations_0, groups = var_180_groups_0, pad = var_180_pad_0, pad_type = var_180_pad_type_0, strides = var_180_strides_0, weight = blocks_0_attn_out_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("op_180_cast_fp16")];
111
+ tensor<fp16, [1, 384, 1, 1500]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_180_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")];
112
+ tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([1])];
113
+ tensor<fp16, [384]> input_7_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_7_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407040)))];
114
+ tensor<fp16, [384]> input_7_beta_0_to_fp16 = const()[name = tensor<string, []>("input_7_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3407872)))];
115
+ tensor<fp16, []> var_190_to_fp16 = const()[name = tensor<string, []>("op_190_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
116
+ tensor<fp16, [1, 384, 1, 1500]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_190_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
117
+ tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("valid")];
118
+ tensor<int32, [2]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
119
+ tensor<int32, [4]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
120
+ tensor<int32, [2]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
121
+ tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)];
122
+ tensor<fp16, [1536, 384, 1, 1]> blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3408704)))];
123
+ tensor<fp16, [1536]> blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4588416)))];
124
+ tensor<fp16, [1, 1536, 1, 1500]> input_9_cast_fp16 = conv(bias = blocks_0_mlp_0_bias_to_fp16, 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 = blocks_0_mlp_0_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
125
+ tensor<string, []> input_11_mode_0 = const()[name = tensor<string, []>("input_11_mode_0"), val = tensor<string, []>("EXACT")];
126
+ tensor<fp16, [1, 1536, 1, 1500]> input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
127
+ tensor<string, []> var_216_pad_type_0 = const()[name = tensor<string, []>("op_216_pad_type_0"), val = tensor<string, []>("valid")];
128
+ tensor<int32, [2]> var_216_strides_0 = const()[name = tensor<string, []>("op_216_strides_0"), val = tensor<int32, [2]>([1, 1])];
129
+ tensor<int32, [4]> var_216_pad_0 = const()[name = tensor<string, []>("op_216_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
130
+ tensor<int32, [2]> var_216_dilations_0 = const()[name = tensor<string, []>("op_216_dilations_0"), val = tensor<int32, [2]>([1, 1])];
131
+ tensor<int32, []> var_216_groups_0 = const()[name = tensor<string, []>("op_216_groups_0"), val = tensor<int32, []>(1)];
132
+ tensor<fp16, [384, 1536, 1, 1]> blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4591552)))];
133
+ tensor<fp16, [384]> blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_0_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5771264)))];
134
+ tensor<fp16, [1, 384, 1, 1500]> var_216_cast_fp16 = conv(bias = blocks_0_mlp_2_bias_to_fp16, dilations = var_216_dilations_0, groups = var_216_groups_0, pad = var_216_pad_0, pad_type = var_216_pad_type_0, strides = var_216_strides_0, weight = blocks_0_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("op_216_cast_fp16")];
135
+ tensor<fp16, [1, 384, 1, 1500]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = var_216_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")];
136
+ tensor<int32, []> var_225 = const()[name = tensor<string, []>("op_225"), val = tensor<int32, []>(1)];
137
+ tensor<int32, [1]> input_13_axes_0 = const()[name = tensor<string, []>("input_13_axes_0"), val = tensor<int32, [1]>([1])];
138
+ tensor<fp16, [384]> input_13_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_13_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5772096)))];
139
+ tensor<fp16, [384]> input_13_beta_0_to_fp16 = const()[name = tensor<string, []>("input_13_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5772928)))];
140
+ tensor<fp16, []> var_241_to_fp16 = const()[name = tensor<string, []>("op_241_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
141
+ tensor<fp16, [1, 384, 1, 1500]> input_13_cast_fp16 = layer_norm(axes = input_13_axes_0, beta = input_13_beta_0_to_fp16, epsilon = var_241_to_fp16, gamma = input_13_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
142
+ tensor<string, []> q_3_pad_type_0 = const()[name = tensor<string, []>("q_3_pad_type_0"), val = tensor<string, []>("valid")];
143
+ tensor<int32, [2]> q_3_strides_0 = const()[name = tensor<string, []>("q_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
144
+ tensor<int32, [4]> q_3_pad_0 = const()[name = tensor<string, []>("q_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
145
+ tensor<int32, [2]> q_3_dilations_0 = const()[name = tensor<string, []>("q_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
146
+ tensor<int32, []> q_3_groups_0 = const()[name = tensor<string, []>("q_3_groups_0"), val = tensor<int32, []>(1)];
147
+ tensor<fp16, [384, 384, 1, 1]> var_276_weight_0_to_fp16 = const()[name = tensor<string, []>("op_276_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5773760)))];
148
+ tensor<fp16, [384]> var_276_bias_0_to_fp16 = const()[name = tensor<string, []>("op_276_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6068736)))];
149
+ tensor<fp16, [1, 384, 1, 1500]> var_276_cast_fp16 = conv(bias = var_276_bias_0_to_fp16, dilations = q_3_dilations_0, groups = q_3_groups_0, pad = q_3_pad_0, pad_type = q_3_pad_type_0, strides = q_3_strides_0, weight = var_276_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("op_276_cast_fp16")];
150
+ tensor<string, []> k_3_pad_type_0 = const()[name = tensor<string, []>("k_3_pad_type_0"), val = tensor<string, []>("valid")];
151
+ tensor<int32, [2]> k_3_strides_0 = const()[name = tensor<string, []>("k_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
152
+ tensor<int32, [4]> k_3_pad_0 = const()[name = tensor<string, []>("k_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
153
+ tensor<int32, [2]> k_3_dilations_0 = const()[name = tensor<string, []>("k_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
154
+ tensor<int32, []> k_3_groups_0 = const()[name = tensor<string, []>("k_3_groups_0"), val = tensor<int32, []>(1)];
155
+ tensor<fp16, [384, 384, 1, 1]> blocks_1_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6069568)))];
156
+ tensor<fp16, [1, 384, 1, 1500]> k_3_cast_fp16 = conv(dilations = k_3_dilations_0, groups = k_3_groups_0, pad = k_3_pad_0, pad_type = k_3_pad_type_0, strides = k_3_strides_0, weight = blocks_1_attn_key_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("k_3_cast_fp16")];
157
+ tensor<string, []> var_274_pad_type_0 = const()[name = tensor<string, []>("op_274_pad_type_0"), val = tensor<string, []>("valid")];
158
+ tensor<int32, [2]> var_274_strides_0 = const()[name = tensor<string, []>("op_274_strides_0"), val = tensor<int32, [2]>([1, 1])];
159
+ tensor<int32, [4]> var_274_pad_0 = const()[name = tensor<string, []>("op_274_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
160
+ tensor<int32, [2]> var_274_dilations_0 = const()[name = tensor<string, []>("op_274_dilations_0"), val = tensor<int32, [2]>([1, 1])];
161
+ tensor<int32, []> var_274_groups_0 = const()[name = tensor<string, []>("op_274_groups_0"), val = tensor<int32, []>(1)];
162
+ tensor<fp16, [384, 384, 1, 1]> blocks_1_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6364544)))];
163
+ tensor<fp16, [384]> blocks_1_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6659520)))];
164
+ tensor<fp16, [1, 384, 1, 1500]> var_274_cast_fp16 = conv(bias = blocks_1_attn_value_bias_to_fp16, dilations = var_274_dilations_0, groups = var_274_groups_0, pad = var_274_pad_0, pad_type = var_274_pad_type_0, strides = var_274_strides_0, weight = blocks_1_attn_value_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("op_274_cast_fp16")];
165
+ tensor<int32, [6]> tile_3 = const()[name = tensor<string, []>("tile_3"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
166
+ tensor<int32, []> var_277_axis_0 = const()[name = tensor<string, []>("op_277_axis_0"), val = tensor<int32, []>(1)];
167
+ tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_277_cast_fp16_5 = split(axis = var_277_axis_0, split_sizes = tile_3, x = var_276_cast_fp16)[name = tensor<string, []>("op_277_cast_fp16")];
168
+ tensor<int32, [4]> var_284_perm_0 = const()[name = tensor<string, []>("op_284_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
169
+ tensor<int32, [6]> tile_4 = const()[name = tensor<string, []>("tile_4"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
170
+ tensor<int32, []> var_285_axis_0 = const()[name = tensor<string, []>("op_285_axis_0"), val = tensor<int32, []>(3)];
171
+ tensor<fp16, [1, 1500, 1, 384]> var_284_cast_fp16 = transpose(perm = var_284_perm_0, x = k_3_cast_fp16)[name = tensor<string, []>("transpose_3")];
172
+ tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_285_cast_fp16_5 = split(axis = var_285_axis_0, split_sizes = tile_4, x = var_284_cast_fp16)[name = tensor<string, []>("op_285_cast_fp16")];
173
+ tensor<int32, [6]> tile_5 = const()[name = tensor<string, []>("tile_5"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
174
+ tensor<int32, []> var_292_axis_0 = const()[name = tensor<string, []>("op_292_axis_0"), val = tensor<int32, []>(1)];
175
+ tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_292_cast_fp16_5 = split(axis = var_292_axis_0, split_sizes = tile_5, x = var_274_cast_fp16)[name = tensor<string, []>("op_292_cast_fp16")];
176
+ tensor<string, []> aw_13_equation_0 = const()[name = tensor<string, []>("aw_13_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
177
+ tensor<fp16, [1, 1500, 1, 1500]> aw_13_cast_fp16 = einsum(equation = aw_13_equation_0, values = (var_285_cast_fp16_0, var_277_cast_fp16_0))[name = tensor<string, []>("aw_13_cast_fp16")];
178
+ tensor<string, []> aw_15_equation_0 = const()[name = tensor<string, []>("aw_15_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
179
+ tensor<fp16, [1, 1500, 1, 1500]> aw_15_cast_fp16 = einsum(equation = aw_15_equation_0, values = (var_285_cast_fp16_1, var_277_cast_fp16_1))[name = tensor<string, []>("aw_15_cast_fp16")];
180
+ tensor<string, []> aw_17_equation_0 = const()[name = tensor<string, []>("aw_17_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
181
+ tensor<fp16, [1, 1500, 1, 1500]> aw_17_cast_fp16 = einsum(equation = aw_17_equation_0, values = (var_285_cast_fp16_2, var_277_cast_fp16_2))[name = tensor<string, []>("aw_17_cast_fp16")];
182
+ tensor<string, []> aw_19_equation_0 = const()[name = tensor<string, []>("aw_19_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
183
+ tensor<fp16, [1, 1500, 1, 1500]> aw_19_cast_fp16 = einsum(equation = aw_19_equation_0, values = (var_285_cast_fp16_3, var_277_cast_fp16_3))[name = tensor<string, []>("aw_19_cast_fp16")];
184
+ tensor<string, []> aw_21_equation_0 = const()[name = tensor<string, []>("aw_21_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
185
+ tensor<fp16, [1, 1500, 1, 1500]> aw_21_cast_fp16 = einsum(equation = aw_21_equation_0, values = (var_285_cast_fp16_4, var_277_cast_fp16_4))[name = tensor<string, []>("aw_21_cast_fp16")];
186
+ tensor<string, []> aw_23_equation_0 = const()[name = tensor<string, []>("aw_23_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
187
+ tensor<fp16, [1, 1500, 1, 1500]> aw_23_cast_fp16 = einsum(equation = aw_23_equation_0, values = (var_285_cast_fp16_5, var_277_cast_fp16_5))[name = tensor<string, []>("aw_23_cast_fp16")];
188
+ tensor<fp16, [1, 1500, 1, 1500]> var_311_cast_fp16 = softmax(axis = var_225, x = aw_13_cast_fp16)[name = tensor<string, []>("op_311_cast_fp16")];
189
+ tensor<fp16, [1, 1500, 1, 1500]> var_312_cast_fp16 = softmax(axis = var_225, x = aw_15_cast_fp16)[name = tensor<string, []>("op_312_cast_fp16")];
190
+ tensor<fp16, [1, 1500, 1, 1500]> var_313_cast_fp16 = softmax(axis = var_225, x = aw_17_cast_fp16)[name = tensor<string, []>("op_313_cast_fp16")];
191
+ tensor<fp16, [1, 1500, 1, 1500]> var_314_cast_fp16 = softmax(axis = var_225, x = aw_19_cast_fp16)[name = tensor<string, []>("op_314_cast_fp16")];
192
+ tensor<fp16, [1, 1500, 1, 1500]> var_315_cast_fp16 = softmax(axis = var_225, x = aw_21_cast_fp16)[name = tensor<string, []>("op_315_cast_fp16")];
193
+ tensor<fp16, [1, 1500, 1, 1500]> var_316_cast_fp16 = softmax(axis = var_225, x = aw_23_cast_fp16)[name = tensor<string, []>("op_316_cast_fp16")];
194
+ tensor<string, []> var_318_equation_0 = const()[name = tensor<string, []>("op_318_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
195
+ tensor<fp16, [1, 64, 1, 1500]> var_318_cast_fp16 = einsum(equation = var_318_equation_0, values = (var_292_cast_fp16_0, var_311_cast_fp16))[name = tensor<string, []>("op_318_cast_fp16")];
196
+ tensor<string, []> var_320_equation_0 = const()[name = tensor<string, []>("op_320_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
197
+ tensor<fp16, [1, 64, 1, 1500]> var_320_cast_fp16 = einsum(equation = var_320_equation_0, values = (var_292_cast_fp16_1, var_312_cast_fp16))[name = tensor<string, []>("op_320_cast_fp16")];
198
+ tensor<string, []> var_322_equation_0 = const()[name = tensor<string, []>("op_322_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
199
+ tensor<fp16, [1, 64, 1, 1500]> var_322_cast_fp16 = einsum(equation = var_322_equation_0, values = (var_292_cast_fp16_2, var_313_cast_fp16))[name = tensor<string, []>("op_322_cast_fp16")];
200
+ tensor<string, []> var_324_equation_0 = const()[name = tensor<string, []>("op_324_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
201
+ tensor<fp16, [1, 64, 1, 1500]> var_324_cast_fp16 = einsum(equation = var_324_equation_0, values = (var_292_cast_fp16_3, var_314_cast_fp16))[name = tensor<string, []>("op_324_cast_fp16")];
202
+ tensor<string, []> var_326_equation_0 = const()[name = tensor<string, []>("op_326_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
203
+ tensor<fp16, [1, 64, 1, 1500]> var_326_cast_fp16 = einsum(equation = var_326_equation_0, values = (var_292_cast_fp16_4, var_315_cast_fp16))[name = tensor<string, []>("op_326_cast_fp16")];
204
+ tensor<string, []> var_328_equation_0 = const()[name = tensor<string, []>("op_328_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
205
+ tensor<fp16, [1, 64, 1, 1500]> var_328_cast_fp16 = einsum(equation = var_328_equation_0, values = (var_292_cast_fp16_5, var_316_cast_fp16))[name = tensor<string, []>("op_328_cast_fp16")];
206
+ tensor<bool, []> input_15_interleave_0 = const()[name = tensor<string, []>("input_15_interleave_0"), val = tensor<bool, []>(false)];
207
+ tensor<fp16, [1, 384, 1, 1500]> input_15_cast_fp16 = concat(axis = var_225, interleave = input_15_interleave_0, values = (var_318_cast_fp16, var_320_cast_fp16, var_322_cast_fp16, var_324_cast_fp16, var_326_cast_fp16, var_328_cast_fp16))[name = tensor<string, []>("input_15_cast_fp16")];
208
+ tensor<string, []> var_337_pad_type_0 = const()[name = tensor<string, []>("op_337_pad_type_0"), val = tensor<string, []>("valid")];
209
+ tensor<int32, [2]> var_337_strides_0 = const()[name = tensor<string, []>("op_337_strides_0"), val = tensor<int32, [2]>([1, 1])];
210
+ tensor<int32, [4]> var_337_pad_0 = const()[name = tensor<string, []>("op_337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
211
+ tensor<int32, [2]> var_337_dilations_0 = const()[name = tensor<string, []>("op_337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
212
+ tensor<int32, []> var_337_groups_0 = const()[name = tensor<string, []>("op_337_groups_0"), val = tensor<int32, []>(1)];
213
+ tensor<fp16, [384, 384, 1, 1]> blocks_1_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6660352)))];
214
+ tensor<fp16, [384]> blocks_1_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6955328)))];
215
+ tensor<fp16, [1, 384, 1, 1500]> var_337_cast_fp16 = conv(bias = blocks_1_attn_out_bias_to_fp16, dilations = var_337_dilations_0, groups = var_337_groups_0, pad = var_337_pad_0, pad_type = var_337_pad_type_0, strides = var_337_strides_0, weight = blocks_1_attn_out_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("op_337_cast_fp16")];
216
+ tensor<fp16, [1, 384, 1, 1500]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = var_337_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")];
217
+ tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([1])];
218
+ tensor<fp16, [384]> input_17_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_17_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6956160)))];
219
+ tensor<fp16, [384]> input_17_beta_0_to_fp16 = const()[name = tensor<string, []>("input_17_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6956992)))];
220
+ tensor<fp16, []> var_347_to_fp16 = const()[name = tensor<string, []>("op_347_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
221
+ tensor<fp16, [1, 384, 1, 1500]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_347_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
222
+ tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
223
+ tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
224
+ tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
225
+ tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
226
+ tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
227
+ tensor<fp16, [1536, 384, 1, 1]> blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6957824)))];
228
+ tensor<fp16, [1536]> blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8137536)))];
229
+ tensor<fp16, [1, 1536, 1, 1500]> input_19_cast_fp16 = conv(bias = blocks_1_mlp_0_bias_to_fp16, 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 = blocks_1_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
230
+ tensor<string, []> input_21_mode_0 = const()[name = tensor<string, []>("input_21_mode_0"), val = tensor<string, []>("EXACT")];
231
+ tensor<fp16, [1, 1536, 1, 1500]> input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
232
+ tensor<string, []> var_373_pad_type_0 = const()[name = tensor<string, []>("op_373_pad_type_0"), val = tensor<string, []>("valid")];
233
+ tensor<int32, [2]> var_373_strides_0 = const()[name = tensor<string, []>("op_373_strides_0"), val = tensor<int32, [2]>([1, 1])];
234
+ tensor<int32, [4]> var_373_pad_0 = const()[name = tensor<string, []>("op_373_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
235
+ tensor<int32, [2]> var_373_dilations_0 = const()[name = tensor<string, []>("op_373_dilations_0"), val = tensor<int32, [2]>([1, 1])];
236
+ tensor<int32, []> var_373_groups_0 = const()[name = tensor<string, []>("op_373_groups_0"), val = tensor<int32, []>(1)];
237
+ tensor<fp16, [384, 1536, 1, 1]> blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8140672)))];
238
+ tensor<fp16, [384]> blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_1_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9320384)))];
239
+ tensor<fp16, [1, 384, 1, 1500]> var_373_cast_fp16 = conv(bias = blocks_1_mlp_2_bias_to_fp16, dilations = var_373_dilations_0, groups = var_373_groups_0, pad = var_373_pad_0, pad_type = var_373_pad_type_0, strides = var_373_strides_0, weight = blocks_1_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("op_373_cast_fp16")];
240
+ tensor<fp16, [1, 384, 1, 1500]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_373_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")];
241
+ tensor<int32, []> var_382 = const()[name = tensor<string, []>("op_382"), val = tensor<int32, []>(1)];
242
+ tensor<int32, [1]> input_23_axes_0 = const()[name = tensor<string, []>("input_23_axes_0"), val = tensor<int32, [1]>([1])];
243
+ tensor<fp16, [384]> input_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_23_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9321216)))];
244
+ tensor<fp16, [384]> input_23_beta_0_to_fp16 = const()[name = tensor<string, []>("input_23_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322048)))];
245
+ tensor<fp16, []> var_398_to_fp16 = const()[name = tensor<string, []>("op_398_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
246
+ tensor<fp16, [1, 384, 1, 1500]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_398_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
247
+ tensor<string, []> q_5_pad_type_0 = const()[name = tensor<string, []>("q_5_pad_type_0"), val = tensor<string, []>("valid")];
248
+ tensor<int32, [2]> q_5_strides_0 = const()[name = tensor<string, []>("q_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
249
+ tensor<int32, [4]> q_5_pad_0 = const()[name = tensor<string, []>("q_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
250
+ tensor<int32, [2]> q_5_dilations_0 = const()[name = tensor<string, []>("q_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
251
+ tensor<int32, []> q_5_groups_0 = const()[name = tensor<string, []>("q_5_groups_0"), val = tensor<int32, []>(1)];
252
+ tensor<fp16, [384, 384, 1, 1]> var_433_weight_0_to_fp16 = const()[name = tensor<string, []>("op_433_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9322880)))];
253
+ tensor<fp16, [384]> var_433_bias_0_to_fp16 = const()[name = tensor<string, []>("op_433_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9617856)))];
254
+ tensor<fp16, [1, 384, 1, 1500]> var_433_cast_fp16 = conv(bias = var_433_bias_0_to_fp16, dilations = q_5_dilations_0, groups = q_5_groups_0, pad = q_5_pad_0, pad_type = q_5_pad_type_0, strides = q_5_strides_0, weight = var_433_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_433_cast_fp16")];
255
+ tensor<string, []> k_5_pad_type_0 = const()[name = tensor<string, []>("k_5_pad_type_0"), val = tensor<string, []>("valid")];
256
+ tensor<int32, [2]> k_5_strides_0 = const()[name = tensor<string, []>("k_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
257
+ tensor<int32, [4]> k_5_pad_0 = const()[name = tensor<string, []>("k_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
258
+ tensor<int32, [2]> k_5_dilations_0 = const()[name = tensor<string, []>("k_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
259
+ tensor<int32, []> k_5_groups_0 = const()[name = tensor<string, []>("k_5_groups_0"), val = tensor<int32, []>(1)];
260
+ tensor<fp16, [384, 384, 1, 1]> blocks_2_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9618688)))];
261
+ tensor<fp16, [1, 384, 1, 1500]> k_5_cast_fp16 = conv(dilations = k_5_dilations_0, groups = k_5_groups_0, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = k_5_strides_0, weight = blocks_2_attn_key_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
262
+ tensor<string, []> var_431_pad_type_0 = const()[name = tensor<string, []>("op_431_pad_type_0"), val = tensor<string, []>("valid")];
263
+ tensor<int32, [2]> var_431_strides_0 = const()[name = tensor<string, []>("op_431_strides_0"), val = tensor<int32, [2]>([1, 1])];
264
+ tensor<int32, [4]> var_431_pad_0 = const()[name = tensor<string, []>("op_431_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
265
+ tensor<int32, [2]> var_431_dilations_0 = const()[name = tensor<string, []>("op_431_dilations_0"), val = tensor<int32, [2]>([1, 1])];
266
+ tensor<int32, []> var_431_groups_0 = const()[name = tensor<string, []>("op_431_groups_0"), val = tensor<int32, []>(1)];
267
+ tensor<fp16, [384, 384, 1, 1]> blocks_2_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9913664)))];
268
+ tensor<fp16, [384]> blocks_2_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10208640)))];
269
+ tensor<fp16, [1, 384, 1, 1500]> var_431_cast_fp16 = conv(bias = blocks_2_attn_value_bias_to_fp16, dilations = var_431_dilations_0, groups = var_431_groups_0, pad = var_431_pad_0, pad_type = var_431_pad_type_0, strides = var_431_strides_0, weight = blocks_2_attn_value_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("op_431_cast_fp16")];
270
+ tensor<int32, [6]> tile_6 = const()[name = tensor<string, []>("tile_6"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
271
+ tensor<int32, []> var_434_axis_0 = const()[name = tensor<string, []>("op_434_axis_0"), val = tensor<int32, []>(1)];
272
+ tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_434_cast_fp16_5 = split(axis = var_434_axis_0, split_sizes = tile_6, x = var_433_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
273
+ tensor<int32, [4]> var_441_perm_0 = const()[name = tensor<string, []>("op_441_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
274
+ tensor<int32, [6]> tile_7 = const()[name = tensor<string, []>("tile_7"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
275
+ tensor<int32, []> var_442_axis_0 = const()[name = tensor<string, []>("op_442_axis_0"), val = tensor<int32, []>(3)];
276
+ tensor<fp16, [1, 1500, 1, 384]> var_441_cast_fp16 = transpose(perm = var_441_perm_0, x = k_5_cast_fp16)[name = tensor<string, []>("transpose_2")];
277
+ tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_442_cast_fp16_5 = split(axis = var_442_axis_0, split_sizes = tile_7, x = var_441_cast_fp16)[name = tensor<string, []>("op_442_cast_fp16")];
278
+ tensor<int32, [6]> tile_8 = const()[name = tensor<string, []>("tile_8"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
279
+ tensor<int32, []> var_449_axis_0 = const()[name = tensor<string, []>("op_449_axis_0"), val = tensor<int32, []>(1)];
280
+ tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_449_cast_fp16_5 = split(axis = var_449_axis_0, split_sizes = tile_8, x = var_431_cast_fp16)[name = tensor<string, []>("op_449_cast_fp16")];
281
+ tensor<string, []> aw_25_equation_0 = const()[name = tensor<string, []>("aw_25_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
282
+ tensor<fp16, [1, 1500, 1, 1500]> aw_25_cast_fp16 = einsum(equation = aw_25_equation_0, values = (var_442_cast_fp16_0, var_434_cast_fp16_0))[name = tensor<string, []>("aw_25_cast_fp16")];
283
+ tensor<string, []> aw_27_equation_0 = const()[name = tensor<string, []>("aw_27_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
284
+ tensor<fp16, [1, 1500, 1, 1500]> aw_27_cast_fp16 = einsum(equation = aw_27_equation_0, values = (var_442_cast_fp16_1, var_434_cast_fp16_1))[name = tensor<string, []>("aw_27_cast_fp16")];
285
+ tensor<string, []> aw_29_equation_0 = const()[name = tensor<string, []>("aw_29_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
286
+ tensor<fp16, [1, 1500, 1, 1500]> aw_29_cast_fp16 = einsum(equation = aw_29_equation_0, values = (var_442_cast_fp16_2, var_434_cast_fp16_2))[name = tensor<string, []>("aw_29_cast_fp16")];
287
+ tensor<string, []> aw_31_equation_0 = const()[name = tensor<string, []>("aw_31_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
288
+ tensor<fp16, [1, 1500, 1, 1500]> aw_31_cast_fp16 = einsum(equation = aw_31_equation_0, values = (var_442_cast_fp16_3, var_434_cast_fp16_3))[name = tensor<string, []>("aw_31_cast_fp16")];
289
+ tensor<string, []> aw_33_equation_0 = const()[name = tensor<string, []>("aw_33_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
290
+ tensor<fp16, [1, 1500, 1, 1500]> aw_33_cast_fp16 = einsum(equation = aw_33_equation_0, values = (var_442_cast_fp16_4, var_434_cast_fp16_4))[name = tensor<string, []>("aw_33_cast_fp16")];
291
+ tensor<string, []> aw_35_equation_0 = const()[name = tensor<string, []>("aw_35_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
292
+ tensor<fp16, [1, 1500, 1, 1500]> aw_35_cast_fp16 = einsum(equation = aw_35_equation_0, values = (var_442_cast_fp16_5, var_434_cast_fp16_5))[name = tensor<string, []>("aw_35_cast_fp16")];
293
+ tensor<fp16, [1, 1500, 1, 1500]> var_468_cast_fp16 = softmax(axis = var_382, x = aw_25_cast_fp16)[name = tensor<string, []>("op_468_cast_fp16")];
294
+ tensor<fp16, [1, 1500, 1, 1500]> var_469_cast_fp16 = softmax(axis = var_382, x = aw_27_cast_fp16)[name = tensor<string, []>("op_469_cast_fp16")];
295
+ tensor<fp16, [1, 1500, 1, 1500]> var_470_cast_fp16 = softmax(axis = var_382, x = aw_29_cast_fp16)[name = tensor<string, []>("op_470_cast_fp16")];
296
+ tensor<fp16, [1, 1500, 1, 1500]> var_471_cast_fp16 = softmax(axis = var_382, x = aw_31_cast_fp16)[name = tensor<string, []>("op_471_cast_fp16")];
297
+ tensor<fp16, [1, 1500, 1, 1500]> var_472_cast_fp16 = softmax(axis = var_382, x = aw_33_cast_fp16)[name = tensor<string, []>("op_472_cast_fp16")];
298
+ tensor<fp16, [1, 1500, 1, 1500]> var_473_cast_fp16 = softmax(axis = var_382, x = aw_35_cast_fp16)[name = tensor<string, []>("op_473_cast_fp16")];
299
+ tensor<string, []> var_475_equation_0 = const()[name = tensor<string, []>("op_475_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
300
+ tensor<fp16, [1, 64, 1, 1500]> var_475_cast_fp16 = einsum(equation = var_475_equation_0, values = (var_449_cast_fp16_0, var_468_cast_fp16))[name = tensor<string, []>("op_475_cast_fp16")];
301
+ tensor<string, []> var_477_equation_0 = const()[name = tensor<string, []>("op_477_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
302
+ tensor<fp16, [1, 64, 1, 1500]> var_477_cast_fp16 = einsum(equation = var_477_equation_0, values = (var_449_cast_fp16_1, var_469_cast_fp16))[name = tensor<string, []>("op_477_cast_fp16")];
303
+ tensor<string, []> var_479_equation_0 = const()[name = tensor<string, []>("op_479_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
304
+ tensor<fp16, [1, 64, 1, 1500]> var_479_cast_fp16 = einsum(equation = var_479_equation_0, values = (var_449_cast_fp16_2, var_470_cast_fp16))[name = tensor<string, []>("op_479_cast_fp16")];
305
+ tensor<string, []> var_481_equation_0 = const()[name = tensor<string, []>("op_481_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
306
+ tensor<fp16, [1, 64, 1, 1500]> var_481_cast_fp16 = einsum(equation = var_481_equation_0, values = (var_449_cast_fp16_3, var_471_cast_fp16))[name = tensor<string, []>("op_481_cast_fp16")];
307
+ tensor<string, []> var_483_equation_0 = const()[name = tensor<string, []>("op_483_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
308
+ tensor<fp16, [1, 64, 1, 1500]> var_483_cast_fp16 = einsum(equation = var_483_equation_0, values = (var_449_cast_fp16_4, var_472_cast_fp16))[name = tensor<string, []>("op_483_cast_fp16")];
309
+ tensor<string, []> var_485_equation_0 = const()[name = tensor<string, []>("op_485_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
310
+ tensor<fp16, [1, 64, 1, 1500]> var_485_cast_fp16 = einsum(equation = var_485_equation_0, values = (var_449_cast_fp16_5, var_473_cast_fp16))[name = tensor<string, []>("op_485_cast_fp16")];
311
+ tensor<bool, []> input_25_interleave_0 = const()[name = tensor<string, []>("input_25_interleave_0"), val = tensor<bool, []>(false)];
312
+ tensor<fp16, [1, 384, 1, 1500]> input_25_cast_fp16 = concat(axis = var_382, interleave = input_25_interleave_0, values = (var_475_cast_fp16, var_477_cast_fp16, var_479_cast_fp16, var_481_cast_fp16, var_483_cast_fp16, var_485_cast_fp16))[name = tensor<string, []>("input_25_cast_fp16")];
313
+ tensor<string, []> var_494_pad_type_0 = const()[name = tensor<string, []>("op_494_pad_type_0"), val = tensor<string, []>("valid")];
314
+ tensor<int32, [2]> var_494_strides_0 = const()[name = tensor<string, []>("op_494_strides_0"), val = tensor<int32, [2]>([1, 1])];
315
+ tensor<int32, [4]> var_494_pad_0 = const()[name = tensor<string, []>("op_494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
316
+ tensor<int32, [2]> var_494_dilations_0 = const()[name = tensor<string, []>("op_494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
317
+ tensor<int32, []> var_494_groups_0 = const()[name = tensor<string, []>("op_494_groups_0"), val = tensor<int32, []>(1)];
318
+ tensor<fp16, [384, 384, 1, 1]> blocks_2_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10209472)))];
319
+ tensor<fp16, [384]> blocks_2_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10504448)))];
320
+ tensor<fp16, [1, 384, 1, 1500]> var_494_cast_fp16 = conv(bias = blocks_2_attn_out_bias_to_fp16, dilations = var_494_dilations_0, groups = var_494_groups_0, pad = var_494_pad_0, pad_type = var_494_pad_type_0, strides = var_494_strides_0, weight = blocks_2_attn_out_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("op_494_cast_fp16")];
321
+ tensor<fp16, [1, 384, 1, 1500]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = var_494_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")];
322
+ tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([1])];
323
+ tensor<fp16, [384]> input_27_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_27_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10505280)))];
324
+ tensor<fp16, [384]> input_27_beta_0_to_fp16 = const()[name = tensor<string, []>("input_27_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506112)))];
325
+ tensor<fp16, []> var_504_to_fp16 = const()[name = tensor<string, []>("op_504_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
326
+ tensor<fp16, [1, 384, 1, 1500]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = input_27_beta_0_to_fp16, epsilon = var_504_to_fp16, gamma = input_27_gamma_0_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
327
+ tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")];
328
+ tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
329
+ tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
330
+ tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
331
+ tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
332
+ tensor<fp16, [1536, 384, 1, 1]> blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10506944)))];
333
+ tensor<fp16, [1536]> blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11686656)))];
334
+ tensor<fp16, [1, 1536, 1, 1500]> input_29_cast_fp16 = conv(bias = blocks_2_mlp_0_bias_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = blocks_2_mlp_0_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
335
+ tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")];
336
+ tensor<fp16, [1, 1536, 1, 1500]> input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
337
+ tensor<string, []> var_530_pad_type_0 = const()[name = tensor<string, []>("op_530_pad_type_0"), val = tensor<string, []>("valid")];
338
+ tensor<int32, [2]> var_530_strides_0 = const()[name = tensor<string, []>("op_530_strides_0"), val = tensor<int32, [2]>([1, 1])];
339
+ tensor<int32, [4]> var_530_pad_0 = const()[name = tensor<string, []>("op_530_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
340
+ tensor<int32, [2]> var_530_dilations_0 = const()[name = tensor<string, []>("op_530_dilations_0"), val = tensor<int32, [2]>([1, 1])];
341
+ tensor<int32, []> var_530_groups_0 = const()[name = tensor<string, []>("op_530_groups_0"), val = tensor<int32, []>(1)];
342
+ tensor<fp16, [384, 1536, 1, 1]> blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11689792)))];
343
+ tensor<fp16, [384]> blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_2_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12869504)))];
344
+ tensor<fp16, [1, 384, 1, 1500]> var_530_cast_fp16 = conv(bias = blocks_2_mlp_2_bias_to_fp16, dilations = var_530_dilations_0, groups = var_530_groups_0, pad = var_530_pad_0, pad_type = var_530_pad_type_0, strides = var_530_strides_0, weight = blocks_2_mlp_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("op_530_cast_fp16")];
345
+ tensor<fp16, [1, 384, 1, 1500]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_530_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")];
346
+ tensor<int32, []> var_539 = const()[name = tensor<string, []>("op_539"), val = tensor<int32, []>(1)];
347
+ tensor<int32, [1]> input_33_axes_0 = const()[name = tensor<string, []>("input_33_axes_0"), val = tensor<int32, [1]>([1])];
348
+ tensor<fp16, [384]> input_33_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_33_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12870336)))];
349
+ tensor<fp16, [384]> input_33_beta_0_to_fp16 = const()[name = tensor<string, []>("input_33_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12871168)))];
350
+ tensor<fp16, []> var_555_to_fp16 = const()[name = tensor<string, []>("op_555_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
351
+ tensor<fp16, [1, 384, 1, 1500]> input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_555_to_fp16, gamma = input_33_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
352
+ tensor<string, []> q_pad_type_0 = const()[name = tensor<string, []>("q_pad_type_0"), val = tensor<string, []>("valid")];
353
+ tensor<int32, [2]> q_strides_0 = const()[name = tensor<string, []>("q_strides_0"), val = tensor<int32, [2]>([1, 1])];
354
+ tensor<int32, [4]> q_pad_0 = const()[name = tensor<string, []>("q_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
355
+ tensor<int32, [2]> q_dilations_0 = const()[name = tensor<string, []>("q_dilations_0"), val = tensor<int32, [2]>([1, 1])];
356
+ tensor<int32, []> q_groups_0 = const()[name = tensor<string, []>("q_groups_0"), val = tensor<int32, []>(1)];
357
+ tensor<fp16, [384, 384, 1, 1]> var_590_weight_0_to_fp16 = const()[name = tensor<string, []>("op_590_weight_0_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12872000)))];
358
+ tensor<fp16, [384]> var_590_bias_0_to_fp16 = const()[name = tensor<string, []>("op_590_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13166976)))];
359
+ tensor<fp16, [1, 384, 1, 1500]> var_590_cast_fp16 = conv(bias = var_590_bias_0_to_fp16, dilations = q_dilations_0, groups = q_groups_0, pad = q_pad_0, pad_type = q_pad_type_0, strides = q_strides_0, weight = var_590_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_590_cast_fp16")];
360
+ tensor<string, []> k_pad_type_0 = const()[name = tensor<string, []>("k_pad_type_0"), val = tensor<string, []>("valid")];
361
+ tensor<int32, [2]> k_strides_0 = const()[name = tensor<string, []>("k_strides_0"), val = tensor<int32, [2]>([1, 1])];
362
+ tensor<int32, [4]> k_pad_0 = const()[name = tensor<string, []>("k_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
363
+ tensor<int32, [2]> k_dilations_0 = const()[name = tensor<string, []>("k_dilations_0"), val = tensor<int32, [2]>([1, 1])];
364
+ tensor<int32, []> k_groups_0 = const()[name = tensor<string, []>("k_groups_0"), val = tensor<int32, []>(1)];
365
+ tensor<fp16, [384, 384, 1, 1]> blocks_3_attn_key_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_key_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13167808)))];
366
+ tensor<fp16, [1, 384, 1, 1500]> k_cast_fp16 = conv(dilations = k_dilations_0, groups = k_groups_0, pad = k_pad_0, pad_type = k_pad_type_0, strides = k_strides_0, weight = blocks_3_attn_key_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
367
+ tensor<string, []> var_588_pad_type_0 = const()[name = tensor<string, []>("op_588_pad_type_0"), val = tensor<string, []>("valid")];
368
+ tensor<int32, [2]> var_588_strides_0 = const()[name = tensor<string, []>("op_588_strides_0"), val = tensor<int32, [2]>([1, 1])];
369
+ tensor<int32, [4]> var_588_pad_0 = const()[name = tensor<string, []>("op_588_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
370
+ tensor<int32, [2]> var_588_dilations_0 = const()[name = tensor<string, []>("op_588_dilations_0"), val = tensor<int32, [2]>([1, 1])];
371
+ tensor<int32, []> var_588_groups_0 = const()[name = tensor<string, []>("op_588_groups_0"), val = tensor<int32, []>(1)];
372
+ tensor<fp16, [384, 384, 1, 1]> blocks_3_attn_value_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_value_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13462784)))];
373
+ tensor<fp16, [384]> blocks_3_attn_value_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13757760)))];
374
+ tensor<fp16, [1, 384, 1, 1500]> var_588_cast_fp16 = conv(bias = blocks_3_attn_value_bias_to_fp16, dilations = var_588_dilations_0, groups = var_588_groups_0, pad = var_588_pad_0, pad_type = var_588_pad_type_0, strides = var_588_strides_0, weight = blocks_3_attn_value_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("op_588_cast_fp16")];
375
+ tensor<int32, [6]> tile_9 = const()[name = tensor<string, []>("tile_9"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
376
+ tensor<int32, []> var_591_axis_0 = const()[name = tensor<string, []>("op_591_axis_0"), val = tensor<int32, []>(1)];
377
+ tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_591_cast_fp16_5 = split(axis = var_591_axis_0, split_sizes = tile_9, x = var_590_cast_fp16)[name = tensor<string, []>("op_591_cast_fp16")];
378
+ tensor<int32, [4]> var_598_perm_0 = const()[name = tensor<string, []>("op_598_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
379
+ tensor<int32, [6]> tile_10 = const()[name = tensor<string, []>("tile_10"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
380
+ tensor<int32, []> var_599_axis_0 = const()[name = tensor<string, []>("op_599_axis_0"), val = tensor<int32, []>(3)];
381
+ tensor<fp16, [1, 1500, 1, 384]> var_598_cast_fp16 = transpose(perm = var_598_perm_0, x = k_cast_fp16)[name = tensor<string, []>("transpose_1")];
382
+ tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_0, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_1, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_2, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_3, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_4, tensor<fp16, [1, 1500, 1, 64]> var_599_cast_fp16_5 = split(axis = var_599_axis_0, split_sizes = tile_10, x = var_598_cast_fp16)[name = tensor<string, []>("op_599_cast_fp16")];
383
+ tensor<int32, [6]> tile_11 = const()[name = tensor<string, []>("tile_11"), val = tensor<int32, [6]>([64, 64, 64, 64, 64, 64])];
384
+ tensor<int32, []> var_606_axis_0 = const()[name = tensor<string, []>("op_606_axis_0"), val = tensor<int32, []>(1)];
385
+ tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_0, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_1, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_2, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_3, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_4, tensor<fp16, [1, 64, 1, 1500]> var_606_cast_fp16_5 = split(axis = var_606_axis_0, split_sizes = tile_11, x = var_588_cast_fp16)[name = tensor<string, []>("op_606_cast_fp16")];
386
+ tensor<string, []> aw_37_equation_0 = const()[name = tensor<string, []>("aw_37_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
387
+ tensor<fp16, [1, 1500, 1, 1500]> aw_37_cast_fp16 = einsum(equation = aw_37_equation_0, values = (var_599_cast_fp16_0, var_591_cast_fp16_0))[name = tensor<string, []>("aw_37_cast_fp16")];
388
+ tensor<string, []> aw_39_equation_0 = const()[name = tensor<string, []>("aw_39_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
389
+ tensor<fp16, [1, 1500, 1, 1500]> aw_39_cast_fp16 = einsum(equation = aw_39_equation_0, values = (var_599_cast_fp16_1, var_591_cast_fp16_1))[name = tensor<string, []>("aw_39_cast_fp16")];
390
+ tensor<string, []> aw_41_equation_0 = const()[name = tensor<string, []>("aw_41_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
391
+ tensor<fp16, [1, 1500, 1, 1500]> aw_41_cast_fp16 = einsum(equation = aw_41_equation_0, values = (var_599_cast_fp16_2, var_591_cast_fp16_2))[name = tensor<string, []>("aw_41_cast_fp16")];
392
+ tensor<string, []> aw_43_equation_0 = const()[name = tensor<string, []>("aw_43_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
393
+ tensor<fp16, [1, 1500, 1, 1500]> aw_43_cast_fp16 = einsum(equation = aw_43_equation_0, values = (var_599_cast_fp16_3, var_591_cast_fp16_3))[name = tensor<string, []>("aw_43_cast_fp16")];
394
+ tensor<string, []> aw_45_equation_0 = const()[name = tensor<string, []>("aw_45_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
395
+ tensor<fp16, [1, 1500, 1, 1500]> aw_45_cast_fp16 = einsum(equation = aw_45_equation_0, values = (var_599_cast_fp16_4, var_591_cast_fp16_4))[name = tensor<string, []>("aw_45_cast_fp16")];
396
+ tensor<string, []> aw_equation_0 = const()[name = tensor<string, []>("aw_equation_0"), val = tensor<string, []>("bkhc,bchq->bkhq")];
397
+ tensor<fp16, [1, 1500, 1, 1500]> aw_cast_fp16 = einsum(equation = aw_equation_0, values = (var_599_cast_fp16_5, var_591_cast_fp16_5))[name = tensor<string, []>("aw_cast_fp16")];
398
+ tensor<fp16, [1, 1500, 1, 1500]> var_625_cast_fp16 = softmax(axis = var_539, x = aw_37_cast_fp16)[name = tensor<string, []>("op_625_cast_fp16")];
399
+ tensor<fp16, [1, 1500, 1, 1500]> var_626_cast_fp16 = softmax(axis = var_539, x = aw_39_cast_fp16)[name = tensor<string, []>("op_626_cast_fp16")];
400
+ tensor<fp16, [1, 1500, 1, 1500]> var_627_cast_fp16 = softmax(axis = var_539, x = aw_41_cast_fp16)[name = tensor<string, []>("op_627_cast_fp16")];
401
+ tensor<fp16, [1, 1500, 1, 1500]> var_628_cast_fp16 = softmax(axis = var_539, x = aw_43_cast_fp16)[name = tensor<string, []>("op_628_cast_fp16")];
402
+ tensor<fp16, [1, 1500, 1, 1500]> var_629_cast_fp16 = softmax(axis = var_539, x = aw_45_cast_fp16)[name = tensor<string, []>("op_629_cast_fp16")];
403
+ tensor<fp16, [1, 1500, 1, 1500]> var_630_cast_fp16 = softmax(axis = var_539, x = aw_cast_fp16)[name = tensor<string, []>("op_630_cast_fp16")];
404
+ tensor<string, []> var_632_equation_0 = const()[name = tensor<string, []>("op_632_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
405
+ tensor<fp16, [1, 64, 1, 1500]> var_632_cast_fp16 = einsum(equation = var_632_equation_0, values = (var_606_cast_fp16_0, var_625_cast_fp16))[name = tensor<string, []>("op_632_cast_fp16")];
406
+ tensor<string, []> var_634_equation_0 = const()[name = tensor<string, []>("op_634_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
407
+ tensor<fp16, [1, 64, 1, 1500]> var_634_cast_fp16 = einsum(equation = var_634_equation_0, values = (var_606_cast_fp16_1, var_626_cast_fp16))[name = tensor<string, []>("op_634_cast_fp16")];
408
+ tensor<string, []> var_636_equation_0 = const()[name = tensor<string, []>("op_636_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
409
+ tensor<fp16, [1, 64, 1, 1500]> var_636_cast_fp16 = einsum(equation = var_636_equation_0, values = (var_606_cast_fp16_2, var_627_cast_fp16))[name = tensor<string, []>("op_636_cast_fp16")];
410
+ tensor<string, []> var_638_equation_0 = const()[name = tensor<string, []>("op_638_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
411
+ tensor<fp16, [1, 64, 1, 1500]> var_638_cast_fp16 = einsum(equation = var_638_equation_0, values = (var_606_cast_fp16_3, var_628_cast_fp16))[name = tensor<string, []>("op_638_cast_fp16")];
412
+ tensor<string, []> var_640_equation_0 = const()[name = tensor<string, []>("op_640_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
413
+ tensor<fp16, [1, 64, 1, 1500]> var_640_cast_fp16 = einsum(equation = var_640_equation_0, values = (var_606_cast_fp16_4, var_629_cast_fp16))[name = tensor<string, []>("op_640_cast_fp16")];
414
+ tensor<string, []> var_642_equation_0 = const()[name = tensor<string, []>("op_642_equation_0"), val = tensor<string, []>("bchk,bkhq->bchq")];
415
+ tensor<fp16, [1, 64, 1, 1500]> var_642_cast_fp16 = einsum(equation = var_642_equation_0, values = (var_606_cast_fp16_5, var_630_cast_fp16))[name = tensor<string, []>("op_642_cast_fp16")];
416
+ tensor<bool, []> input_35_interleave_0 = const()[name = tensor<string, []>("input_35_interleave_0"), val = tensor<bool, []>(false)];
417
+ tensor<fp16, [1, 384, 1, 1500]> input_35_cast_fp16 = concat(axis = var_539, interleave = input_35_interleave_0, values = (var_632_cast_fp16, var_634_cast_fp16, var_636_cast_fp16, var_638_cast_fp16, var_640_cast_fp16, var_642_cast_fp16))[name = tensor<string, []>("input_35_cast_fp16")];
418
+ tensor<string, []> var_651_pad_type_0 = const()[name = tensor<string, []>("op_651_pad_type_0"), val = tensor<string, []>("valid")];
419
+ tensor<int32, [2]> var_651_strides_0 = const()[name = tensor<string, []>("op_651_strides_0"), val = tensor<int32, [2]>([1, 1])];
420
+ tensor<int32, [4]> var_651_pad_0 = const()[name = tensor<string, []>("op_651_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
421
+ tensor<int32, [2]> var_651_dilations_0 = const()[name = tensor<string, []>("op_651_dilations_0"), val = tensor<int32, [2]>([1, 1])];
422
+ tensor<int32, []> var_651_groups_0 = const()[name = tensor<string, []>("op_651_groups_0"), val = tensor<int32, []>(1)];
423
+ tensor<fp16, [384, 384, 1, 1]> blocks_3_attn_out_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_out_weight_to_fp16"), val = tensor<fp16, [384, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13758592)))];
424
+ tensor<fp16, [384]> blocks_3_attn_out_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_attn_out_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14053568)))];
425
+ tensor<fp16, [1, 384, 1, 1500]> var_651_cast_fp16 = conv(bias = blocks_3_attn_out_bias_to_fp16, dilations = var_651_dilations_0, groups = var_651_groups_0, pad = var_651_pad_0, pad_type = var_651_pad_type_0, strides = var_651_strides_0, weight = blocks_3_attn_out_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("op_651_cast_fp16")];
426
+ tensor<fp16, [1, 384, 1, 1500]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = var_651_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")];
427
+ tensor<int32, [1]> input_37_axes_0 = const()[name = tensor<string, []>("input_37_axes_0"), val = tensor<int32, [1]>([1])];
428
+ tensor<fp16, [384]> input_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_37_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14054400)))];
429
+ tensor<fp16, [384]> input_37_beta_0_to_fp16 = const()[name = tensor<string, []>("input_37_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14055232)))];
430
+ tensor<fp16, []> var_661_to_fp16 = const()[name = tensor<string, []>("op_661_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
431
+ tensor<fp16, [1, 384, 1, 1500]> input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = input_37_beta_0_to_fp16, epsilon = var_661_to_fp16, gamma = input_37_gamma_0_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
432
+ tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("valid")];
433
+ tensor<int32, [2]> input_39_strides_0 = const()[name = tensor<string, []>("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
434
+ tensor<int32, [4]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
435
+ tensor<int32, [2]> input_39_dilations_0 = const()[name = tensor<string, []>("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
436
+ tensor<int32, []> input_39_groups_0 = const()[name = tensor<string, []>("input_39_groups_0"), val = tensor<int32, []>(1)];
437
+ tensor<fp16, [1536, 384, 1, 1]> blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_0_weight_to_fp16"), val = tensor<fp16, [1536, 384, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14056064)))];
438
+ tensor<fp16, [1536]> blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_0_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15235776)))];
439
+ tensor<fp16, [1, 1536, 1, 1500]> input_39_cast_fp16 = conv(bias = blocks_3_mlp_0_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = blocks_3_mlp_0_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
440
+ tensor<string, []> input_mode_0 = const()[name = tensor<string, []>("input_mode_0"), val = tensor<string, []>("EXACT")];
441
+ tensor<fp16, [1, 1536, 1, 1500]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_39_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
442
+ tensor<string, []> var_687_pad_type_0 = const()[name = tensor<string, []>("op_687_pad_type_0"), val = tensor<string, []>("valid")];
443
+ tensor<int32, [2]> var_687_strides_0 = const()[name = tensor<string, []>("op_687_strides_0"), val = tensor<int32, [2]>([1, 1])];
444
+ tensor<int32, [4]> var_687_pad_0 = const()[name = tensor<string, []>("op_687_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
445
+ tensor<int32, [2]> var_687_dilations_0 = const()[name = tensor<string, []>("op_687_dilations_0"), val = tensor<int32, [2]>([1, 1])];
446
+ tensor<int32, []> var_687_groups_0 = const()[name = tensor<string, []>("op_687_groups_0"), val = tensor<int32, []>(1)];
447
+ tensor<fp16, [384, 1536, 1, 1]> blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_2_weight_to_fp16"), val = tensor<fp16, [384, 1536, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15238912)))];
448
+ tensor<fp16, [384]> blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor<string, []>("blocks_3_mlp_2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16418624)))];
449
+ tensor<fp16, [1, 384, 1, 1500]> var_687_cast_fp16 = conv(bias = blocks_3_mlp_2_bias_to_fp16, dilations = var_687_dilations_0, groups = var_687_groups_0, pad = var_687_pad_0, pad_type = var_687_pad_type_0, strides = var_687_strides_0, weight = blocks_3_mlp_2_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("op_687_cast_fp16")];
450
+ tensor<fp16, [1, 384, 1, 1500]> inputs_cast_fp16 = add(x = inputs_15_cast_fp16, y = var_687_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")];
451
+ tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([1])];
452
+ tensor<fp16, [384]> x_gamma_0_to_fp16 = const()[name = tensor<string, []>("x_gamma_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16419456)))];
453
+ tensor<fp16, [384]> x_beta_0_to_fp16 = const()[name = tensor<string, []>("x_beta_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16420288)))];
454
+ tensor<fp16, []> var_701_to_fp16 = const()[name = tensor<string, []>("op_701_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
455
+ tensor<fp16, [1, 384, 1, 1500]> x_cast_fp16 = layer_norm(axes = x_axes_0, beta = x_beta_0_to_fp16, epsilon = var_701_to_fp16, gamma = x_gamma_0_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
456
+ tensor<int32, [1]> var_712_axes_0 = const()[name = tensor<string, []>("op_712_axes_0"), val = tensor<int32, [1]>([2])];
457
+ tensor<fp16, [1, 384, 1500]> var_712_cast_fp16 = squeeze(axes = var_712_axes_0, x = x_cast_fp16)[name = tensor<string, []>("op_712_cast_fp16")];
458
+ tensor<int32, [3]> var_715_perm_0 = const()[name = tensor<string, []>("op_715_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
459
+ tensor<string, []> var_715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_715_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
460
+ tensor<fp16, [1, 1500, 384]> var_715_cast_fp16 = transpose(perm = var_715_perm_0, x = var_712_cast_fp16)[name = tensor<string, []>("transpose_0")];
461
+ tensor<fp32, [1, 1500, 384]> output = cast(dtype = var_715_cast_fp16_to_fp32_dtype_0, x = var_715_cast_fp16)[name = tensor<string, []>("cast_19")];
462
+ } -> (output);
463
+ }
tiny/ggml-tiny-encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:17aa929eb61d433fa68217d73a8aec7125af1b4afe39b5c8f27d61d80e2c1a80
3
+ size 16421120
tiny/ggml-tiny.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be07e048e1e599ad46341c8d2a135645097a538221678b7acdd1b1919c6e1b21
3
+ size 77691713