Hugging Face's logo

OpenVINO
/
distil-whisper-large-v3-int4-ov

OpenVINO
whisper
Model card Files Community
openvino-ci commited on
Commit
43f28c8
•
1 Parent(s): 566b185

Upload folder using huggingface_hub

Browse files
Files changed (7) hide show
  1. config.json +2 -1
  2. generation_config.json +1 -1
  3. openvino_decoder_model.xml +197 -196
  4. openvino_detokenizer.bin +1 -1
  5. openvino_detokenizer.xml +32 -12
  6. openvino_encoder_model.xml +0 -0
  7. openvino_tokenizer.xml +74 -54
config.json CHANGED
@@ -1,4 +1,5 @@
1
  {
 
2
  "_name_or_path": "distil-whisper/distil-large-v3",
3
  "activation_dropout": 0.0,
4
  "activation_function": "gelu",
@@ -39,7 +40,7 @@
39
  "num_mel_bins": 128,
40
  "pad_token_id": 50256,
41
  "scale_embedding": false,
42
- "transformers_version": "4.45.2",
43
  "use_cache": true,
44
  "use_weighted_layer_sum": false,
45
  "vocab_size": 51866
 
1
  {
2
+ "_attn_implementation_autoset": true,
3
  "_name_or_path": "distil-whisper/distil-large-v3",
4
  "activation_dropout": 0.0,
5
  "activation_function": "gelu",
 
40
  "num_mel_bins": 128,
41
  "pad_token_id": 50256,
42
  "scale_embedding": false,
43
+ "transformers_version": "4.46.3",
44
  "use_cache": true,
45
  "use_weighted_layer_sum": false,
46
  "vocab_size": 51866
generation_config.json CHANGED
@@ -303,5 +303,5 @@
303
  "transcribe": 50360,
304
  "translate": 50359
305
  },
306
- "transformers_version": "4.45.2"
307
  }
 
303
  "transcribe": 50360,
304
  "translate": 50359
305
  },
306
+ "transformers_version": "4.46.3"
307
  }
openvino_decoder_model.xml CHANGED
@@ -10,7 +10,7 @@
10
  </port>
11
  </output>
12
  </layer>
13
- <layer id="0" name="Parameter_25934" type="Parameter" version="opset1">
14
  <data shape="?,?,?" element_type="f32" />
15
  <output>
16
  <port id="0" precision="FP32" names="encoder_hidden_states">
@@ -29,7 +29,7 @@
29
  </port>
30
  </output>
31
  </layer>
32
- <layer id="3" name="Convert_42647" type="Convert" version="opset1">
33
  <data destination_type="f16" />
34
  <input>
35
  <port id="0" precision="U8">
@@ -53,7 +53,7 @@
53
  </port>
54
  </output>
55
  </layer>
56
- <layer id="5" name="Convert_42650" type="Convert" version="opset1">
57
  <data destination_type="f16" />
58
  <input>
59
  <port id="0" precision="U8">
@@ -130,7 +130,7 @@
130
  </port>
131
  </output>
132
  </layer>
133
- <layer id="10" name="Constant_31368" type="Const" version="opset1">
134
  <data element_type="i64" shape="2" offset="66544078" size="16" />
135
  <output>
136
  <port id="0" precision="I64">
@@ -207,7 +207,7 @@
207
  </port>
208
  </output>
209
  </layer>
210
- <layer id="16" name="Convert_42878" type="Convert" version="opset1">
211
  <data destination_type="f16" />
212
  <input>
213
  <port id="0" precision="U8">
@@ -231,7 +231,7 @@
231
  </port>
232
  </output>
233
  </layer>
234
- <layer id="18" name="Convert_42881" type="Convert" version="opset1">
235
  <data destination_type="f16" />
236
  <input>
237
  <port id="0" precision="U8">
@@ -314,7 +314,7 @@
314
  <port id="0" precision="I64" names="23" />
315
  </output>
316
  </layer>
317
- <layer id="24" name="ShapeOf_31341" type="ShapeOf" version="opset3">
318
  <data output_type="i64" />
319
  <input>
320
  <port id="0" precision="I64">
@@ -328,19 +328,19 @@
328
  </port>
329
  </output>
330
  </layer>
331
- <layer id="25" name="Constant_31342" type="Const" version="opset1">
332
  <data element_type="i64" shape="" offset="67118890" size="8" />
333
  <output>
334
  <port id="0" precision="I64" />
335
  </output>
336
  </layer>
337
- <layer id="26" name="Constant_31343" type="Const" version="opset1">
338
  <data element_type="i64" shape="" offset="67118882" size="8" />
339
  <output>
340
  <port id="0" precision="I64" />
341
  </output>
342
  </layer>
343
- <layer id="27" name="Gather_31344" type="Gather" version="opset8">
344
  <data batch_dims="0" />
345
  <input>
346
  <port id="0" precision="I64">
@@ -401,7 +401,7 @@
401
  </port>
402
  </output>
403
  </layer>
404
- <layer id="32" name="Constant_25931" type="Const" version="opset1">
405
  <data element_type="i32" shape="" offset="66544094" size="4" />
406
  <output>
407
  <port id="0" precision="I32" />
@@ -478,7 +478,7 @@
478
  </port>
479
  </output>
480
  </layer>
481
- <layer id="37" name="Constant_31109" type="Const" version="opset1">
482
  <data element_type="f32" shape="1, 1, 1280" offset="67118906" size="5120" />
483
  <output>
484
  <port id="0" precision="FP32">
@@ -510,7 +510,7 @@
510
  </port>
511
  </output>
512
  </layer>
513
- <layer id="39" name="Constant_31110" type="Const" version="opset1">
514
  <data element_type="f32" shape="1, 1, 1280" offset="67124026" size="5120" />
515
  <output>
516
  <port id="0" precision="FP32">
@@ -542,7 +542,7 @@
542
  </port>
543
  </output>
544
  </layer>
545
- <layer id="41" name="Constant_31012" type="Const" version="opset1">
546
  <data element_type="u4" shape="1280, 10, 128" offset="67129146" size="819200" />
547
  <output>
548
  <port id="0" precision="U4">
@@ -552,7 +552,7 @@
552
  </port>
553
  </output>
554
  </layer>
555
- <layer id="42" name="Convert_42767" type="Convert" version="opset1">
556
  <data destination_type="f16" />
557
  <input>
558
  <port id="0" precision="U4">
@@ -569,7 +569,7 @@
569
  </port>
570
  </output>
571
  </layer>
572
- <layer id="43" name="Constant_31012/zero_point" type="Const" version="opset1">
573
  <data element_type="u4" shape="1280, 10, 1" offset="67948346" size="6400" />
574
  <output>
575
  <port id="0" precision="U4">
@@ -579,7 +579,7 @@
579
  </port>
580
  </output>
581
  </layer>
582
- <layer id="44" name="Convert_42769" type="Convert" version="opset1">
583
  <data destination_type="f16" />
584
  <input>
585
  <port id="0" precision="U4">
@@ -596,7 +596,7 @@
596
  </port>
597
  </output>
598
  </layer>
599
- <layer id="45" name="Constant_31012/zero_point/subtract" type="Subtract" version="opset1">
600
  <data auto_broadcast="numpy" />
601
  <input>
602
  <port id="0" precision="FP16">
@@ -618,7 +618,7 @@
618
  </port>
619
  </output>
620
  </layer>
621
- <layer id="46" name="Constant_31012/scale" type="Const" version="opset1">
622
  <data element_type="f16" shape="1280, 10, 1" offset="67954746" size="25600" />
623
  <output>
624
  <port id="0" precision="FP16">
@@ -628,7 +628,7 @@
628
  </port>
629
  </output>
630
  </layer>
631
- <layer id="47" name="Constant_31012/fq_weights_1" type="Multiply" version="opset1">
632
  <data auto_broadcast="numpy" />
633
  <input>
634
  <port id="0" precision="FP16">
@@ -650,7 +650,7 @@
650
  </port>
651
  </output>
652
  </layer>
653
- <layer id="48" name="Constant_42774" type="Const" version="opset1">
654
  <data element_type="i64" shape="2" offset="67980346" size="16" />
655
  <output>
656
  <port id="0" precision="I64">
@@ -658,7 +658,7 @@
658
  </port>
659
  </output>
660
  </layer>
661
- <layer id="49" name="Reshape_42775" type="Reshape" version="opset1">
662
  <data special_zero="false" />
663
  <input>
664
  <port id="0" precision="FP16">
@@ -677,7 +677,7 @@
677
  </port>
678
  </output>
679
  </layer>
680
- <layer id="50" name="Constant_31012/fq_weights_1/convert" type="Convert" version="opset1">
681
  <data destination_type="f32" />
682
  <input>
683
  <port id="0" precision="FP16">
@@ -692,7 +692,7 @@
692
  </port>
693
  </output>
694
  </layer>
695
- <layer id="51" name="Multiply_30995" type="MatMul" version="opset1">
696
  <data transpose_a="false" transpose_b="true" />
697
  <input>
698
  <port id="0" precision="FP32">
@@ -713,7 +713,7 @@
713
  </port>
714
  </output>
715
  </layer>
716
- <layer id="52" name="Constant_31111" type="Const" version="opset1">
717
  <data element_type="f32" shape="1, 1, 1280" offset="67980362" size="5120" />
718
  <output>
719
  <port id="0" precision="FP32">
@@ -745,7 +745,7 @@
745
  </port>
746
  </output>
747
  </layer>
748
- <layer id="54" name="Constant_31369" type="Const" version="opset1">
749
  <data element_type="i64" shape="4" offset="67985482" size="32" />
750
  <output>
751
  <port id="0" precision="I64">
@@ -813,7 +813,7 @@
813
  </port>
814
  </output>
815
  </layer>
816
- <layer id="59" name="Convert_42756" type="Convert" version="opset1">
817
  <data destination_type="f16" />
818
  <input>
819
  <port id="0" precision="U4">
@@ -840,7 +840,7 @@
840
  </port>
841
  </output>
842
  </layer>
843
- <layer id="61" name="Convert_42758" type="Convert" version="opset1">
844
  <data destination_type="f16" />
845
  <input>
846
  <port id="0" precision="U4">
@@ -911,7 +911,7 @@
911
  </port>
912
  </output>
913
  </layer>
914
- <layer id="65" name="Constant_42763" type="Const" version="opset1">
915
  <data element_type="i64" shape="2" offset="67980346" size="16" />
916
  <output>
917
  <port id="0" precision="I64">
@@ -919,7 +919,7 @@
919
  </port>
920
  </output>
921
  </layer>
922
- <layer id="66" name="Reshape_42764" type="Reshape" version="opset1">
923
  <data special_zero="false" />
924
  <input>
925
  <port id="0" precision="FP16">
@@ -974,7 +974,7 @@
974
  </port>
975
  </output>
976
  </layer>
977
- <layer id="69" name="Constant_31370" type="Const" version="opset1">
978
  <data element_type="i64" shape="4" offset="67985482" size="32" />
979
  <output>
980
  <port id="0" precision="I64">
@@ -1063,13 +1063,13 @@
1063
  <port id="0" precision="I32" />
1064
  </output>
1065
  </layer>
1066
- <layer id="75" name="__module.model.model.decoder/aten::full/Convert" type="Const" version="opset1">
1067
  <data element_type="f32" shape="" offset="68836730" size="4" />
1068
  <output>
1069
- <port id="0" precision="FP32" />
1070
  </output>
1071
  </layer>
1072
- <layer id="76" name="Constant_31351" type="Const" version="opset1">
1073
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1074
  <output>
1075
  <port id="0" precision="I64">
@@ -1077,7 +1077,7 @@
1077
  </port>
1078
  </output>
1079
  </layer>
1080
- <layer id="77" name="Reshape_31352" type="Reshape" version="opset1">
1081
  <data special_zero="false" />
1082
  <input>
1083
  <port id="0" precision="I64" />
@@ -1107,13 +1107,13 @@
1107
  <port id="2" precision="I64" names="62,64" />
1108
  </output>
1109
  </layer>
1110
- <layer id="80" name="Constant_23811" type="Const" version="opset1">
1111
  <data element_type="i32" shape="" offset="66544094" size="4" />
1112
  <output>
1113
  <port id="0" precision="I32" />
1114
  </output>
1115
  </layer>
1116
- <layer id="81" name="Unsqueeze_23812" type="Unsqueeze" version="opset1">
1117
  <input>
1118
  <port id="0" precision="I64" />
1119
  <port id="1" precision="I32" />
@@ -1155,7 +1155,7 @@
1155
  </port>
1156
  </output>
1157
  </layer>
1158
- <layer id="84" name="ShapeOf_31353" type="ShapeOf" version="opset3">
1159
  <data output_type="i32" />
1160
  <input>
1161
  <port id="0" precision="FP32">
@@ -1169,19 +1169,19 @@
1169
  </port>
1170
  </output>
1171
  </layer>
1172
- <layer id="85" name="Constant_31354" type="Const" version="opset1">
1173
  <data element_type="i64" shape="" offset="67118890" size="8" />
1174
  <output>
1175
  <port id="0" precision="I64" />
1176
  </output>
1177
  </layer>
1178
- <layer id="86" name="Constant_31355" type="Const" version="opset1">
1179
  <data element_type="i64" shape="" offset="67118882" size="8" />
1180
  <output>
1181
  <port id="0" precision="I64" />
1182
  </output>
1183
  </layer>
1184
- <layer id="87" name="Gather_31356" type="Gather" version="opset8">
1185
  <data batch_dims="0" />
1186
  <input>
1187
  <port id="0" precision="I32">
@@ -1233,7 +1233,7 @@
1233
  <port id="0" precision="I32" />
1234
  </output>
1235
  </layer>
1236
- <layer id="92" name="Convert_31359" type="Convert" version="opset1">
1237
  <data destination_type="i32" />
1238
  <input>
1239
  <port id="0" precision="I64" />
@@ -1375,7 +1375,7 @@
1375
  </port>
1376
  </output>
1377
  </layer>
1378
- <layer id="104" name="Constant_23914" type="Const" version="opset1">
1379
  <data element_type="i64" shape="2" offset="68836734" size="16" />
1380
  <output>
1381
  <port id="0" precision="I64" names="70">
@@ -1492,7 +1492,7 @@
1492
  </port>
1493
  </output>
1494
  </layer>
1495
- <layer id="112" name="Constant_31348" type="Const" version="opset1">
1496
  <data element_type="i64" shape="1" offset="67118882" size="8" />
1497
  <output>
1498
  <port id="0" precision="I64">
@@ -1500,13 +1500,13 @@
1500
  </port>
1501
  </output>
1502
  </layer>
1503
- <layer id="113" name="Constant_31349" type="Const" version="opset1">
1504
  <data element_type="i64" shape="" offset="67118882" size="8" />
1505
  <output>
1506
  <port id="0" precision="I64" />
1507
  </output>
1508
  </layer>
1509
- <layer id="114" name="Gather_31350" type="Gather" version="opset8">
1510
  <data batch_dims="0" />
1511
  <input>
1512
  <port id="0" precision="I64">
@@ -1523,7 +1523,7 @@
1523
  </port>
1524
  </output>
1525
  </layer>
1526
- <layer id="115" name="Constant_30340" type="Const" version="opset1">
1527
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1528
  <output>
1529
  <port id="0" precision="I64">
@@ -1531,7 +1531,7 @@
1531
  </port>
1532
  </output>
1533
  </layer>
1534
- <layer id="116" name="Constant_30342" type="Const" version="opset1">
1535
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1536
  <output>
1537
  <port id="0" precision="I64">
@@ -1539,7 +1539,7 @@
1539
  </port>
1540
  </output>
1541
  </layer>
1542
- <layer id="117" name="Constant_30344" type="Const" version="opset1">
1543
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1544
  <output>
1545
  <port id="0" precision="I64">
@@ -1591,7 +1591,7 @@
1591
  </port>
1592
  </output>
1593
  </layer>
1594
- <layer id="120" name="Constant_24348" type="Const" version="opset1">
1595
  <data element_type="i64" shape="1" offset="67118882" size="8" />
1596
  <output>
1597
  <port id="0" precision="I64">
@@ -1599,7 +1599,7 @@
1599
  </port>
1600
  </output>
1601
  </layer>
1602
- <layer id="121" name="ShapeOf_31364" type="ShapeOf" version="opset3">
1603
  <data output_type="i64" />
1604
  <input>
1605
  <port id="0" precision="FP32">
@@ -1614,7 +1614,7 @@
1614
  </port>
1615
  </output>
1616
  </layer>
1617
- <layer id="122" name="Constant_31365" type="Const" version="opset1">
1618
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1619
  <output>
1620
  <port id="0" precision="I64">
@@ -1622,13 +1622,13 @@
1622
  </port>
1623
  </output>
1624
  </layer>
1625
- <layer id="123" name="Constant_31366" type="Const" version="opset1">
1626
  <data element_type="i64" shape="" offset="67118882" size="8" />
1627
  <output>
1628
  <port id="0" precision="I64" />
1629
  </output>
1630
  </layer>
1631
- <layer id="124" name="Gather_31367" type="Gather" version="opset8">
1632
  <data batch_dims="0" />
1633
  <input>
1634
  <port id="0" precision="I64">
@@ -1645,7 +1645,7 @@
1645
  </port>
1646
  </output>
1647
  </layer>
1648
- <layer id="125" name="Constant_24351" type="Const" version="opset1">
1649
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1650
  <output>
1651
  <port id="0" precision="I64">
@@ -1745,7 +1745,7 @@
1745
  </port>
1746
  </output>
1747
  </layer>
1748
- <layer id="131" name="Convert_42745" type="Convert" version="opset1">
1749
  <data destination_type="f16" />
1750
  <input>
1751
  <port id="0" precision="U4">
@@ -1772,7 +1772,7 @@
1772
  </port>
1773
  </output>
1774
  </layer>
1775
- <layer id="133" name="Convert_42747" type="Convert" version="opset1">
1776
  <data destination_type="f16" />
1777
  <input>
1778
  <port id="0" precision="U4">
@@ -1843,7 +1843,7 @@
1843
  </port>
1844
  </output>
1845
  </layer>
1846
- <layer id="137" name="Constant_42752" type="Const" version="opset1">
1847
  <data element_type="i64" shape="2" offset="67980346" size="16" />
1848
  <output>
1849
  <port id="0" precision="I64">
@@ -1851,7 +1851,7 @@
1851
  </port>
1852
  </output>
1853
  </layer>
1854
- <layer id="138" name="Reshape_42753" type="Reshape" version="opset1">
1855
  <data special_zero="false" />
1856
  <input>
1857
  <port id="0" precision="FP16">
@@ -1906,7 +1906,7 @@
1906
  </port>
1907
  </output>
1908
  </layer>
1909
- <layer id="141" name="Constant_31112" type="Const" version="opset1">
1910
  <data element_type="f32" shape="1, 1, 1280" offset="69687958" size="5120" />
1911
  <output>
1912
  <port id="0" precision="FP32">
@@ -1938,7 +1938,7 @@
1938
  </port>
1939
  </output>
1940
  </layer>
1941
- <layer id="143" name="Constant_31371" type="Const" version="opset1">
1942
  <data element_type="i64" shape="4" offset="67985482" size="32" />
1943
  <output>
1944
  <port id="0" precision="I64">
@@ -2050,7 +2050,7 @@
2050
  </port>
2051
  </output>
2052
  </layer>
2053
- <layer id="150" name="Constant_31372" type="Const" version="opset1">
2054
  <data element_type="i64" shape="3" offset="69693078" size="24" />
2055
  <output>
2056
  <port id="0" precision="I64">
@@ -2089,7 +2089,7 @@
2089
  </port>
2090
  </output>
2091
  </layer>
2092
- <layer id="153" name="Convert_42778" type="Convert" version="opset1">
2093
  <data destination_type="f16" />
2094
  <input>
2095
  <port id="0" precision="U4">
@@ -2116,7 +2116,7 @@
2116
  </port>
2117
  </output>
2118
  </layer>
2119
- <layer id="155" name="Convert_42780" type="Convert" version="opset1">
2120
  <data destination_type="f16" />
2121
  <input>
2122
  <port id="0" precision="U4">
@@ -2187,7 +2187,7 @@
2187
  </port>
2188
  </output>
2189
  </layer>
2190
- <layer id="159" name="Constant_42785" type="Const" version="opset1">
2191
  <data element_type="i64" shape="2" offset="67980346" size="16" />
2192
  <output>
2193
  <port id="0" precision="I64">
@@ -2195,7 +2195,7 @@
2195
  </port>
2196
  </output>
2197
  </layer>
2198
- <layer id="160" name="Reshape_42786" type="Reshape" version="opset1">
2199
  <data special_zero="false" />
2200
  <input>
2201
  <port id="0" precision="FP16">
@@ -2250,7 +2250,7 @@
2250
  </port>
2251
  </output>
2252
  </layer>
2253
- <layer id="163" name="Constant_31113" type="Const" version="opset1">
2254
  <data element_type="f32" shape="1, 1, 1280" offset="70544302" size="5120" />
2255
  <output>
2256
  <port id="0" precision="FP32">
@@ -2332,7 +2332,7 @@
2332
  </port>
2333
  </output>
2334
  </layer>
2335
- <layer id="168" name="Constant_31114" type="Const" version="opset1">
2336
  <data element_type="f32" shape="1, 1, 1280" offset="70549422" size="5120" />
2337
  <output>
2338
  <port id="0" precision="FP32">
@@ -2364,7 +2364,7 @@
2364
  </port>
2365
  </output>
2366
  </layer>
2367
- <layer id="170" name="Constant_31115" type="Const" version="opset1">
2368
  <data element_type="f32" shape="1, 1, 1280" offset="70554542" size="5120" />
2369
  <output>
2370
  <port id="0" precision="FP32">
@@ -2396,7 +2396,7 @@
2396
  </port>
2397
  </output>
2398
  </layer>
2399
- <layer id="172" name="Constant_31015" type="Const" version="opset1">
2400
  <data element_type="u4" shape="1280, 10, 128" offset="70559662" size="819200" />
2401
  <output>
2402
  <port id="0" precision="U4">
@@ -2406,7 +2406,7 @@
2406
  </port>
2407
  </output>
2408
  </layer>
2409
- <layer id="173" name="Convert_42789" type="Convert" version="opset1">
2410
  <data destination_type="f16" />
2411
  <input>
2412
  <port id="0" precision="U4">
@@ -2423,7 +2423,7 @@
2423
  </port>
2424
  </output>
2425
  </layer>
2426
- <layer id="174" name="Constant_31015/zero_point" type="Const" version="opset1">
2427
  <data element_type="u4" shape="1280, 10, 1" offset="71378862" size="6400" />
2428
  <output>
2429
  <port id="0" precision="U4">
@@ -2433,7 +2433,7 @@
2433
  </port>
2434
  </output>
2435
  </layer>
2436
- <layer id="175" name="Convert_42791" type="Convert" version="opset1">
2437
  <data destination_type="f16" />
2438
  <input>
2439
  <port id="0" precision="U4">
@@ -2450,7 +2450,7 @@
2450
  </port>
2451
  </output>
2452
  </layer>
2453
- <layer id="176" name="Constant_31015/zero_point/subtract" type="Subtract" version="opset1">
2454
  <data auto_broadcast="numpy" />
2455
  <input>
2456
  <port id="0" precision="FP16">
@@ -2472,7 +2472,7 @@
2472
  </port>
2473
  </output>
2474
  </layer>
2475
- <layer id="177" name="Constant_31015/scale" type="Const" version="opset1">
2476
  <data element_type="f16" shape="1280, 10, 1" offset="71385262" size="25600" />
2477
  <output>
2478
  <port id="0" precision="FP16">
@@ -2482,7 +2482,7 @@
2482
  </port>
2483
  </output>
2484
  </layer>
2485
- <layer id="178" name="Constant_31015/fq_weights_1" type="Multiply" version="opset1">
2486
  <data auto_broadcast="numpy" />
2487
  <input>
2488
  <port id="0" precision="FP16">
@@ -2504,7 +2504,7 @@
2504
  </port>
2505
  </output>
2506
  </layer>
2507
- <layer id="179" name="Constant_42796" type="Const" version="opset1">
2508
  <data element_type="i64" shape="2" offset="67980346" size="16" />
2509
  <output>
2510
  <port id="0" precision="I64">
@@ -2512,7 +2512,7 @@
2512
  </port>
2513
  </output>
2514
  </layer>
2515
- <layer id="180" name="Reshape_42797" type="Reshape" version="opset1">
2516
  <data special_zero="false" />
2517
  <input>
2518
  <port id="0" precision="FP16">
@@ -2531,7 +2531,7 @@
2531
  </port>
2532
  </output>
2533
  </layer>
2534
- <layer id="181" name="Constant_31015/fq_weights_1/convert" type="Convert" version="opset1">
2535
  <data destination_type="f32" />
2536
  <input>
2537
  <port id="0" precision="FP16">
@@ -2546,7 +2546,7 @@
2546
  </port>
2547
  </output>
2548
  </layer>
2549
- <layer id="182" name="Multiply_30999" type="MatMul" version="opset1">
2550
  <data transpose_a="false" transpose_b="true" />
2551
  <input>
2552
  <port id="0" precision="FP32">
@@ -2567,7 +2567,7 @@
2567
  </port>
2568
  </output>
2569
  </layer>
2570
- <layer id="183" name="Constant_31116" type="Const" version="opset1">
2571
  <data element_type="f32" shape="1, 1, 1280" offset="71410862" size="5120" />
2572
  <output>
2573
  <port id="0" precision="FP32">
@@ -2599,7 +2599,7 @@
2599
  </port>
2600
  </output>
2601
  </layer>
2602
- <layer id="185" name="Constant_31373" type="Const" version="opset1">
2603
  <data element_type="i64" shape="4" offset="67985482" size="32" />
2604
  <output>
2605
  <port id="0" precision="I64">
@@ -2667,7 +2667,7 @@
2667
  </port>
2668
  </output>
2669
  </layer>
2670
- <layer id="190" name="Convert_42734" type="Convert" version="opset1">
2671
  <data destination_type="f16" />
2672
  <input>
2673
  <port id="0" precision="U4">
@@ -2694,7 +2694,7 @@
2694
  </port>
2695
  </output>
2696
  </layer>
2697
- <layer id="192" name="Convert_42736" type="Convert" version="opset1">
2698
  <data destination_type="f16" />
2699
  <input>
2700
  <port id="0" precision="U4">
@@ -2765,7 +2765,7 @@
2765
  </port>
2766
  </output>
2767
  </layer>
2768
- <layer id="196" name="Constant_42741" type="Const" version="opset1">
2769
  <data element_type="i64" shape="2" offset="67980346" size="16" />
2770
  <output>
2771
  <port id="0" precision="I64">
@@ -2773,7 +2773,7 @@
2773
  </port>
2774
  </output>
2775
  </layer>
2776
- <layer id="197" name="Reshape_42742" type="Reshape" version="opset1">
2777
  <data special_zero="false" />
2778
  <input>
2779
  <port id="0" precision="FP16">
@@ -2828,7 +2828,7 @@
2828
  </port>
2829
  </output>
2830
  </layer>
2831
- <layer id="200" name="Constant_24572" type="Const" version="opset1">
2832
  <data element_type="i64" shape="1" offset="72267182" size="8" />
2833
  <output>
2834
  <port id="0" precision="I64">
@@ -2836,7 +2836,7 @@
2836
  </port>
2837
  </output>
2838
  </layer>
2839
- <layer id="201" name="Constant_24573" type="Const" version="opset1">
2840
  <data element_type="i64" shape="1" offset="72267190" size="8" />
2841
  <output>
2842
  <port id="0" precision="I64">
@@ -2844,7 +2844,7 @@
2844
  </port>
2845
  </output>
2846
  </layer>
2847
- <layer id="202" name="Constant_24574" type="Const" version="opset1">
2848
  <data element_type="i64" shape="1" offset="72267198" size="8" />
2849
  <output>
2850
  <port id="0" precision="I64">
@@ -2852,7 +2852,7 @@
2852
  </port>
2853
  </output>
2854
  </layer>
2855
- <layer id="203" name="__module.model.model.decoder.layers.0.encoder_attn/prim::ListConstruct_1" type="Concat" version="opset1">
2856
  <data axis="0" />
2857
  <input>
2858
  <port id="0" precision="I64">
@@ -2978,7 +2978,7 @@
2978
  </port>
2979
  </output>
2980
  </layer>
2981
- <layer id="210" name="Convert_42723" type="Convert" version="opset1">
2982
  <data destination_type="f16" />
2983
  <input>
2984
  <port id="0" precision="U4">
@@ -3005,7 +3005,7 @@
3005
  </port>
3006
  </output>
3007
  </layer>
3008
- <layer id="212" name="Convert_42725" type="Convert" version="opset1">
3009
  <data destination_type="f16" />
3010
  <input>
3011
  <port id="0" precision="U4">
@@ -3076,7 +3076,7 @@
3076
  </port>
3077
  </output>
3078
  </layer>
3079
- <layer id="216" name="Constant_42730" type="Const" version="opset1">
3080
  <data element_type="i64" shape="2" offset="67980346" size="16" />
3081
  <output>
3082
  <port id="0" precision="I64">
@@ -3084,7 +3084,7 @@
3084
  </port>
3085
  </output>
3086
  </layer>
3087
- <layer id="217" name="Reshape_42731" type="Reshape" version="opset1">
3088
  <data special_zero="false" />
3089
  <input>
3090
  <port id="0" precision="FP16">
@@ -3139,7 +3139,7 @@
3139
  </port>
3140
  </output>
3141
  </layer>
3142
- <layer id="220" name="Constant_31117" type="Const" version="opset1">
3143
  <data element_type="f32" shape="1, 1, 1280" offset="73118406" size="5120" />
3144
  <output>
3145
  <port id="0" precision="FP32">
@@ -3275,7 +3275,7 @@
3275
  </port>
3276
  </output>
3277
  </layer>
3278
- <layer id="228" name="Constant_31374" type="Const" version="opset1">
3279
  <data element_type="i64" shape="3" offset="73123526" size="24" />
3280
  <output>
3281
  <port id="0" precision="I64">
@@ -3314,7 +3314,7 @@
3314
  </port>
3315
  </output>
3316
  </layer>
3317
- <layer id="231" name="Convert_42800" type="Convert" version="opset1">
3318
  <data destination_type="f16" />
3319
  <input>
3320
  <port id="0" precision="U4">
@@ -3341,7 +3341,7 @@
3341
  </port>
3342
  </output>
3343
  </layer>
3344
- <layer id="233" name="Convert_42802" type="Convert" version="opset1">
3345
  <data destination_type="f16" />
3346
  <input>
3347
  <port id="0" precision="U4">
@@ -3412,7 +3412,7 @@
3412
  </port>
3413
  </output>
3414
  </layer>
3415
- <layer id="237" name="Constant_42807" type="Const" version="opset1">
3416
  <data element_type="i64" shape="2" offset="67980346" size="16" />
3417
  <output>
3418
  <port id="0" precision="I64">
@@ -3420,7 +3420,7 @@
3420
  </port>
3421
  </output>
3422
  </layer>
3423
- <layer id="238" name="Reshape_42808" type="Reshape" version="opset1">
3424
  <data special_zero="false" />
3425
  <input>
3426
  <port id="0" precision="FP16">
@@ -3475,7 +3475,7 @@
3475
  </port>
3476
  </output>
3477
  </layer>
3478
- <layer id="241" name="Constant_31118" type="Const" version="opset1">
3479
  <data element_type="f32" shape="1, 1, 1280" offset="73974750" size="5120" />
3480
  <output>
3481
  <port id="0" precision="FP32">
@@ -3557,7 +3557,7 @@
3557
  </port>
3558
  </output>
3559
  </layer>
3560
- <layer id="246" name="Constant_31119" type="Const" version="opset1">
3561
  <data element_type="f32" shape="1, 1, 1280" offset="73979870" size="5120" />
3562
  <output>
3563
  <port id="0" precision="FP32">
@@ -3589,7 +3589,7 @@
3589
  </port>
3590
  </output>
3591
  </layer>
3592
- <layer id="248" name="Constant_31120" type="Const" version="opset1">
3593
  <data element_type="f32" shape="1, 1, 1280" offset="73984990" size="5120" />
3594
  <output>
3595
  <port id="0" precision="FP32">
@@ -3631,7 +3631,7 @@
3631
  </port>
3632
  </output>
3633
  </layer>
3634
- <layer id="251" name="Convert_42657" type="Convert" version="opset1">
3635
  <data destination_type="f16" />
3636
  <input>
3637
  <port id="0" precision="U4">
@@ -3658,7 +3658,7 @@
3658
  </port>
3659
  </output>
3660
  </layer>
3661
- <layer id="253" name="Convert_42659" type="Convert" version="opset1">
3662
  <data destination_type="f16" />
3663
  <input>
3664
  <port id="0" precision="U4">
@@ -3729,7 +3729,7 @@
3729
  </port>
3730
  </output>
3731
  </layer>
3732
- <layer id="257" name="Constant_42664" type="Const" version="opset1">
3733
  <data element_type="i64" shape="2" offset="77394910" size="16" />
3734
  <output>
3735
  <port id="0" precision="I64">
@@ -3737,7 +3737,7 @@
3737
  </port>
3738
  </output>
3739
  </layer>
3740
- <layer id="258" name="Reshape_42665" type="Reshape" version="opset1">
3741
  <data special_zero="false" />
3742
  <input>
3743
  <port id="0" precision="FP16">
@@ -3792,7 +3792,7 @@
3792
  </port>
3793
  </output>
3794
  </layer>
3795
- <layer id="261" name="Constant_31121" type="Const" version="opset1">
3796
  <data element_type="f32" shape="1, 1, 5120" offset="77394926" size="20480" />
3797
  <output>
3798
  <port id="0" precision="FP32">
@@ -3851,7 +3851,7 @@
3851
  </port>
3852
  </output>
3853
  </layer>
3854
- <layer id="265" name="Convert_42668" type="Convert" version="opset1">
3855
  <data destination_type="f16" />
3856
  <input>
3857
  <port id="0" precision="U4">
@@ -3878,7 +3878,7 @@
3878
  </port>
3879
  </output>
3880
  </layer>
3881
- <layer id="267" name="Convert_42670" type="Convert" version="opset1">
3882
  <data destination_type="f16" />
3883
  <input>
3884
  <port id="0" precision="U4">
@@ -3949,7 +3949,7 @@
3949
  </port>
3950
  </output>
3951
  </layer>
3952
- <layer id="271" name="Constant_42675" type="Const" version="opset1">
3953
  <data element_type="i64" shape="2" offset="80820206" size="16" />
3954
  <output>
3955
  <port id="0" precision="I64">
@@ -3957,7 +3957,7 @@
3957
  </port>
3958
  </output>
3959
  </layer>
3960
- <layer id="272" name="Reshape_42676" type="Reshape" version="opset1">
3961
  <data special_zero="false" />
3962
  <input>
3963
  <port id="0" precision="FP16">
@@ -4012,7 +4012,7 @@
4012
  </port>
4013
  </output>
4014
  </layer>
4015
- <layer id="275" name="Constant_31122" type="Const" version="opset1">
4016
  <data element_type="f32" shape="1, 1, 1280" offset="80820222" size="5120" />
4017
  <output>
4018
  <port id="0" precision="FP32">
@@ -4094,7 +4094,7 @@
4094
  </port>
4095
  </output>
4096
  </layer>
4097
- <layer id="280" name="Constant_31123" type="Const" version="opset1">
4098
  <data element_type="f32" shape="1, 1, 1280" offset="80825342" size="5120" />
4099
  <output>
4100
  <port id="0" precision="FP32">
@@ -4126,7 +4126,7 @@
4126
  </port>
4127
  </output>
4128
  </layer>
4129
- <layer id="282" name="Constant_31124" type="Const" version="opset1">
4130
  <data element_type="f32" shape="1, 1, 1280" offset="80830462" size="5120" />
4131
  <output>
4132
  <port id="0" precision="FP32">
@@ -4158,7 +4158,7 @@
4158
  </port>
4159
  </output>
4160
  </layer>
4161
- <layer id="284" name="Constant_31018" type="Const" version="opset1">
4162
  <data element_type="u4" shape="1280, 10, 128" offset="80835582" size="819200" />
4163
  <output>
4164
  <port id="0" precision="U4">
@@ -4168,7 +4168,7 @@
4168
  </port>
4169
  </output>
4170
  </layer>
4171
- <layer id="285" name="Convert_42833" type="Convert" version="opset1">
4172
  <data destination_type="f16" />
4173
  <input>
4174
  <port id="0" precision="U4">
@@ -4185,7 +4185,7 @@
4185
  </port>
4186
  </output>
4187
  </layer>
4188
- <layer id="286" name="Constant_31018/zero_point" type="Const" version="opset1">
4189
  <data element_type="u4" shape="1280, 10, 1" offset="81654782" size="6400" />
4190
  <output>
4191
  <port id="0" precision="U4">
@@ -4195,7 +4195,7 @@
4195
  </port>
4196
  </output>
4197
  </layer>
4198
- <layer id="287" name="Convert_42835" type="Convert" version="opset1">
4199
  <data destination_type="f16" />
4200
  <input>
4201
  <port id="0" precision="U4">
@@ -4212,7 +4212,7 @@
4212
  </port>
4213
  </output>
4214
  </layer>
4215
- <layer id="288" name="Constant_31018/zero_point/subtract" type="Subtract" version="opset1">
4216
  <data auto_broadcast="numpy" />
4217
  <input>
4218
  <port id="0" precision="FP16">
@@ -4234,7 +4234,7 @@
4234
  </port>
4235
  </output>
4236
  </layer>
4237
- <layer id="289" name="Constant_31018/scale" type="Const" version="opset1">
4238
  <data element_type="f16" shape="1280, 10, 1" offset="81661182" size="25600" />
4239
  <output>
4240
  <port id="0" precision="FP16">
@@ -4244,7 +4244,7 @@
4244
  </port>
4245
  </output>
4246
  </layer>
4247
- <layer id="290" name="Constant_31018/fq_weights_1" type="Multiply" version="opset1">
4248
  <data auto_broadcast="numpy" />
4249
  <input>
4250
  <port id="0" precision="FP16">
@@ -4266,7 +4266,7 @@
4266
  </port>
4267
  </output>
4268
  </layer>
4269
- <layer id="291" name="Constant_42840" type="Const" version="opset1">
4270
  <data element_type="i64" shape="2" offset="67980346" size="16" />
4271
  <output>
4272
  <port id="0" precision="I64">
@@ -4274,7 +4274,7 @@
4274
  </port>
4275
  </output>
4276
  </layer>
4277
- <layer id="292" name="Reshape_42841" type="Reshape" version="opset1">
4278
  <data special_zero="false" />
4279
  <input>
4280
  <port id="0" precision="FP16">
@@ -4293,7 +4293,7 @@
4293
  </port>
4294
  </output>
4295
  </layer>
4296
- <layer id="293" name="Constant_31018/fq_weights_1/convert" type="Convert" version="opset1">
4297
  <data destination_type="f32" />
4298
  <input>
4299
  <port id="0" precision="FP16">
@@ -4308,7 +4308,7 @@
4308
  </port>
4309
  </output>
4310
  </layer>
4311
- <layer id="294" name="Multiply_31003" type="MatMul" version="opset1">
4312
  <data transpose_a="false" transpose_b="true" />
4313
  <input>
4314
  <port id="0" precision="FP32">
@@ -4329,7 +4329,7 @@
4329
  </port>
4330
  </output>
4331
  </layer>
4332
- <layer id="295" name="Constant_31125" type="Const" version="opset1">
4333
  <data element_type="f32" shape="1, 1, 1280" offset="81686782" size="5120" />
4334
  <output>
4335
  <port id="0" precision="FP32">
@@ -4361,7 +4361,7 @@
4361
  </port>
4362
  </output>
4363
  </layer>
4364
- <layer id="297" name="Constant_31375" type="Const" version="opset1">
4365
  <data element_type="i64" shape="4" offset="67985482" size="32" />
4366
  <output>
4367
  <port id="0" precision="I64">
@@ -4429,7 +4429,7 @@
4429
  </port>
4430
  </output>
4431
  </layer>
4432
- <layer id="302" name="Convert_42822" type="Convert" version="opset1">
4433
  <data destination_type="f16" />
4434
  <input>
4435
  <port id="0" precision="U4">
@@ -4456,7 +4456,7 @@
4456
  </port>
4457
  </output>
4458
  </layer>
4459
- <layer id="304" name="Convert_42824" type="Convert" version="opset1">
4460
  <data destination_type="f16" />
4461
  <input>
4462
  <port id="0" precision="U4">
@@ -4527,7 +4527,7 @@
4527
  </port>
4528
  </output>
4529
  </layer>
4530
- <layer id="308" name="Constant_42829" type="Const" version="opset1">
4531
  <data element_type="i64" shape="2" offset="67980346" size="16" />
4532
  <output>
4533
  <port id="0" precision="I64">
@@ -4535,7 +4535,7 @@
4535
  </port>
4536
  </output>
4537
  </layer>
4538
- <layer id="309" name="Reshape_42830" type="Reshape" version="opset1">
4539
  <data special_zero="false" />
4540
  <input>
4541
  <port id="0" precision="FP16">
@@ -4590,7 +4590,7 @@
4590
  </port>
4591
  </output>
4592
  </layer>
4593
- <layer id="312" name="Constant_31376" type="Const" version="opset1">
4594
  <data element_type="i64" shape="4" offset="67985482" size="32" />
4595
  <output>
4596
  <port id="0" precision="I64">
@@ -4727,7 +4727,7 @@
4727
  </port>
4728
  </output>
4729
  </layer>
4730
- <layer id="320" name="Convert_42811" type="Convert" version="opset1">
4731
  <data destination_type="f16" />
4732
  <input>
4733
  <port id="0" precision="U4">
@@ -4754,7 +4754,7 @@
4754
  </port>
4755
  </output>
4756
  </layer>
4757
- <layer id="322" name="Convert_42813" type="Convert" version="opset1">
4758
  <data destination_type="f16" />
4759
  <input>
4760
  <port id="0" precision="U4">
@@ -4825,7 +4825,7 @@
4825
  </port>
4826
  </output>
4827
  </layer>
4828
- <layer id="326" name="Constant_42818" type="Const" version="opset1">
4829
  <data element_type="i64" shape="2" offset="67980346" size="16" />
4830
  <output>
4831
  <port id="0" precision="I64">
@@ -4833,7 +4833,7 @@
4833
  </port>
4834
  </output>
4835
  </layer>
4836
- <layer id="327" name="Reshape_42819" type="Reshape" version="opset1">
4837
  <data special_zero="false" />
4838
  <input>
4839
  <port id="0" precision="FP16">
@@ -4888,7 +4888,7 @@
4888
  </port>
4889
  </output>
4890
  </layer>
4891
- <layer id="330" name="Constant_31126" type="Const" version="opset1">
4892
  <data element_type="f32" shape="1, 1, 1280" offset="83394302" size="5120" />
4893
  <output>
4894
  <port id="0" precision="FP32">
@@ -4920,7 +4920,7 @@
4920
  </port>
4921
  </output>
4922
  </layer>
4923
- <layer id="332" name="Constant_31377" type="Const" version="opset1">
4924
  <data element_type="i64" shape="4" offset="67985482" size="32" />
4925
  <output>
4926
  <port id="0" precision="I64">
@@ -5032,7 +5032,7 @@
5032
  </port>
5033
  </output>
5034
  </layer>
5035
- <layer id="339" name="Constant_31378" type="Const" version="opset1">
5036
  <data element_type="i64" shape="3" offset="69693078" size="24" />
5037
  <output>
5038
  <port id="0" precision="I64">
@@ -5071,7 +5071,7 @@
5071
  </port>
5072
  </output>
5073
  </layer>
5074
- <layer id="342" name="Convert_42844" type="Convert" version="opset1">
5075
  <data destination_type="f16" />
5076
  <input>
5077
  <port id="0" precision="U4">
@@ -5098,7 +5098,7 @@
5098
  </port>
5099
  </output>
5100
  </layer>
5101
- <layer id="344" name="Convert_42846" type="Convert" version="opset1">
5102
  <data destination_type="f16" />
5103
  <input>
5104
  <port id="0" precision="U4">
@@ -5169,7 +5169,7 @@
5169
  </port>
5170
  </output>
5171
  </layer>
5172
- <layer id="348" name="Constant_42851" type="Const" version="opset1">
5173
  <data element_type="i64" shape="2" offset="67980346" size="16" />
5174
  <output>
5175
  <port id="0" precision="I64">
@@ -5177,7 +5177,7 @@
5177
  </port>
5178
  </output>
5179
  </layer>
5180
- <layer id="349" name="Reshape_42852" type="Reshape" version="opset1">
5181
  <data special_zero="false" />
5182
  <input>
5183
  <port id="0" precision="FP16">
@@ -5232,7 +5232,7 @@
5232
  </port>
5233
  </output>
5234
  </layer>
5235
- <layer id="352" name="Constant_31127" type="Const" version="opset1">
5236
  <data element_type="f32" shape="1, 1, 1280" offset="84250622" size="5120" />
5237
  <output>
5238
  <port id="0" precision="FP32">
@@ -5314,7 +5314,7 @@
5314
  </port>
5315
  </output>
5316
  </layer>
5317
- <layer id="357" name="Constant_31128" type="Const" version="opset1">
5318
  <data element_type="f32" shape="1, 1, 1280" offset="84255742" size="5120" />
5319
  <output>
5320
  <port id="0" precision="FP32">
@@ -5346,7 +5346,7 @@
5346
  </port>
5347
  </output>
5348
  </layer>
5349
- <layer id="359" name="Constant_31129" type="Const" version="opset1">
5350
  <data element_type="f32" shape="1, 1, 1280" offset="84260862" size="5120" />
5351
  <output>
5352
  <port id="0" precision="FP32">
@@ -5378,7 +5378,7 @@
5378
  </port>
5379
  </output>
5380
  </layer>
5381
- <layer id="361" name="Constant_31021" type="Const" version="opset1">
5382
  <data element_type="u4" shape="1280, 10, 128" offset="84265982" size="819200" />
5383
  <output>
5384
  <port id="0" precision="U4">
@@ -5388,7 +5388,7 @@
5388
  </port>
5389
  </output>
5390
  </layer>
5391
- <layer id="362" name="Convert_42855" type="Convert" version="opset1">
5392
  <data destination_type="f16" />
5393
  <input>
5394
  <port id="0" precision="U4">
@@ -5405,7 +5405,7 @@
5405
  </port>
5406
  </output>
5407
  </layer>
5408
- <layer id="363" name="Constant_31021/zero_point" type="Const" version="opset1">
5409
  <data element_type="u4" shape="1280, 10, 1" offset="85085182" size="6400" />
5410
  <output>
5411
  <port id="0" precision="U4">
@@ -5415,7 +5415,7 @@
5415
  </port>
5416
  </output>
5417
  </layer>
5418
- <layer id="364" name="Convert_42857" type="Convert" version="opset1">
5419
  <data destination_type="f16" />
5420
  <input>
5421
  <port id="0" precision="U4">
@@ -5432,7 +5432,7 @@
5432
  </port>
5433
  </output>
5434
  </layer>
5435
- <layer id="365" name="Constant_31021/zero_point/subtract" type="Subtract" version="opset1">
5436
  <data auto_broadcast="numpy" />
5437
  <input>
5438
  <port id="0" precision="FP16">
@@ -5454,7 +5454,7 @@
5454
  </port>
5455
  </output>
5456
  </layer>
5457
- <layer id="366" name="Constant_31021/scale" type="Const" version="opset1">
5458
  <data element_type="f16" shape="1280, 10, 1" offset="85091582" size="25600" />
5459
  <output>
5460
  <port id="0" precision="FP16">
@@ -5464,7 +5464,7 @@
5464
  </port>
5465
  </output>
5466
  </layer>
5467
- <layer id="367" name="Constant_31021/fq_weights_1" type="Multiply" version="opset1">
5468
  <data auto_broadcast="numpy" />
5469
  <input>
5470
  <port id="0" precision="FP16">
@@ -5486,7 +5486,7 @@
5486
  </port>
5487
  </output>
5488
  </layer>
5489
- <layer id="368" name="Constant_42862" type="Const" version="opset1">
5490
  <data element_type="i64" shape="2" offset="67980346" size="16" />
5491
  <output>
5492
  <port id="0" precision="I64">
@@ -5494,7 +5494,7 @@
5494
  </port>
5495
  </output>
5496
  </layer>
5497
- <layer id="369" name="Reshape_42863" type="Reshape" version="opset1">
5498
  <data special_zero="false" />
5499
  <input>
5500
  <port id="0" precision="FP16">
@@ -5513,7 +5513,7 @@
5513
  </port>
5514
  </output>
5515
  </layer>
5516
- <layer id="370" name="Constant_31021/fq_weights_1/convert" type="Convert" version="opset1">
5517
  <data destination_type="f32" />
5518
  <input>
5519
  <port id="0" precision="FP16">
@@ -5528,7 +5528,7 @@
5528
  </port>
5529
  </output>
5530
  </layer>
5531
- <layer id="371" name="Multiply_31007" type="MatMul" version="opset1">
5532
  <data transpose_a="false" transpose_b="true" />
5533
  <input>
5534
  <port id="0" precision="FP32">
@@ -5549,7 +5549,7 @@
5549
  </port>
5550
  </output>
5551
  </layer>
5552
- <layer id="372" name="Constant_31130" type="Const" version="opset1">
5553
  <data element_type="f32" shape="1, 1, 1280" offset="85117182" size="5120" />
5554
  <output>
5555
  <port id="0" precision="FP32">
@@ -5581,7 +5581,7 @@
5581
  </port>
5582
  </output>
5583
  </layer>
5584
- <layer id="374" name="Constant_31379" type="Const" version="opset1">
5585
  <data element_type="i64" shape="4" offset="67985482" size="32" />
5586
  <output>
5587
  <port id="0" precision="I64">
@@ -5649,7 +5649,7 @@
5649
  </port>
5650
  </output>
5651
  </layer>
5652
- <layer id="379" name="Convert_42712" type="Convert" version="opset1">
5653
  <data destination_type="f16" />
5654
  <input>
5655
  <port id="0" precision="U4">
@@ -5676,7 +5676,7 @@
5676
  </port>
5677
  </output>
5678
  </layer>
5679
- <layer id="381" name="Convert_42714" type="Convert" version="opset1">
5680
  <data destination_type="f16" />
5681
  <input>
5682
  <port id="0" precision="U4">
@@ -5747,7 +5747,7 @@
5747
  </port>
5748
  </output>
5749
  </layer>
5750
- <layer id="385" name="Constant_42719" type="Const" version="opset1">
5751
  <data element_type="i64" shape="2" offset="67980346" size="16" />
5752
  <output>
5753
  <port id="0" precision="I64">
@@ -5755,7 +5755,7 @@
5755
  </port>
5756
  </output>
5757
  </layer>
5758
- <layer id="386" name="Reshape_42720" type="Reshape" version="opset1">
5759
  <data special_zero="false" />
5760
  <input>
5761
  <port id="0" precision="FP16">
@@ -5914,7 +5914,7 @@
5914
  </port>
5915
  </output>
5916
  </layer>
5917
- <layer id="395" name="Convert_42701" type="Convert" version="opset1">
5918
  <data destination_type="f16" />
5919
  <input>
5920
  <port id="0" precision="U4">
@@ -5941,7 +5941,7 @@
5941
  </port>
5942
  </output>
5943
  </layer>
5944
- <layer id="397" name="Convert_42703" type="Convert" version="opset1">
5945
  <data destination_type="f16" />
5946
  <input>
5947
  <port id="0" precision="U4">
@@ -6012,7 +6012,7 @@
6012
  </port>
6013
  </output>
6014
  </layer>
6015
- <layer id="401" name="Constant_42708" type="Const" version="opset1">
6016
  <data element_type="i64" shape="2" offset="67980346" size="16" />
6017
  <output>
6018
  <port id="0" precision="I64">
@@ -6020,7 +6020,7 @@
6020
  </port>
6021
  </output>
6022
  </layer>
6023
- <layer id="402" name="Reshape_42709" type="Reshape" version="opset1">
6024
  <data special_zero="false" />
6025
  <input>
6026
  <port id="0" precision="FP16">
@@ -6075,7 +6075,7 @@
6075
  </port>
6076
  </output>
6077
  </layer>
6078
- <layer id="405" name="Constant_31131" type="Const" version="opset1">
6079
  <data element_type="f32" shape="1, 1, 1280" offset="86824702" size="5120" />
6080
  <output>
6081
  <port id="0" precision="FP32">
@@ -6211,7 +6211,7 @@
6211
  </port>
6212
  </output>
6213
  </layer>
6214
- <layer id="413" name="Constant_31380" type="Const" version="opset1">
6215
  <data element_type="i64" shape="3" offset="73123526" size="24" />
6216
  <output>
6217
  <port id="0" precision="I64">
@@ -6250,7 +6250,7 @@
6250
  </port>
6251
  </output>
6252
  </layer>
6253
- <layer id="416" name="Convert_42866" type="Convert" version="opset1">
6254
  <data destination_type="f16" />
6255
  <input>
6256
  <port id="0" precision="U4">
@@ -6277,7 +6277,7 @@
6277
  </port>
6278
  </output>
6279
  </layer>
6280
- <layer id="418" name="Convert_42868" type="Convert" version="opset1">
6281
  <data destination_type="f16" />
6282
  <input>
6283
  <port id="0" precision="U4">
@@ -6348,7 +6348,7 @@
6348
  </port>
6349
  </output>
6350
  </layer>
6351
- <layer id="422" name="Constant_42873" type="Const" version="opset1">
6352
  <data element_type="i64" shape="2" offset="67980346" size="16" />
6353
  <output>
6354
  <port id="0" precision="I64">
@@ -6356,7 +6356,7 @@
6356
  </port>
6357
  </output>
6358
  </layer>
6359
- <layer id="423" name="Reshape_42874" type="Reshape" version="opset1">
6360
  <data special_zero="false" />
6361
  <input>
6362
  <port id="0" precision="FP16">
@@ -6411,7 +6411,7 @@
6411
  </port>
6412
  </output>
6413
  </layer>
6414
- <layer id="426" name="Constant_31132" type="Const" version="opset1">
6415
  <data element_type="f32" shape="1, 1, 1280" offset="87681022" size="5120" />
6416
  <output>
6417
  <port id="0" precision="FP32">
@@ -6493,7 +6493,7 @@
6493
  </port>
6494
  </output>
6495
  </layer>
6496
- <layer id="431" name="Constant_31133" type="Const" version="opset1">
6497
  <data element_type="f32" shape="1, 1, 1280" offset="87686142" size="5120" />
6498
  <output>
6499
  <port id="0" precision="FP32">
@@ -6525,7 +6525,7 @@
6525
  </port>
6526
  </output>
6527
  </layer>
6528
- <layer id="433" name="Constant_31134" type="Const" version="opset1">
6529
  <data element_type="f32" shape="1, 1, 1280" offset="87691262" size="5120" />
6530
  <output>
6531
  <port id="0" precision="FP32">
@@ -6567,7 +6567,7 @@
6567
  </port>
6568
  </output>
6569
  </layer>
6570
- <layer id="436" name="Convert_42679" type="Convert" version="opset1">
6571
  <data destination_type="f16" />
6572
  <input>
6573
  <port id="0" precision="U4">
@@ -6594,7 +6594,7 @@
6594
  </port>
6595
  </output>
6596
  </layer>
6597
- <layer id="438" name="Convert_42681" type="Convert" version="opset1">
6598
  <data destination_type="f16" />
6599
  <input>
6600
  <port id="0" precision="U4">
@@ -6665,7 +6665,7 @@
6665
  </port>
6666
  </output>
6667
  </layer>
6668
- <layer id="442" name="Constant_42686" type="Const" version="opset1">
6669
  <data element_type="i64" shape="2" offset="77394910" size="16" />
6670
  <output>
6671
  <port id="0" precision="I64">
@@ -6673,7 +6673,7 @@
6673
  </port>
6674
  </output>
6675
  </layer>
6676
- <layer id="443" name="Reshape_42687" type="Reshape" version="opset1">
6677
  <data special_zero="false" />
6678
  <input>
6679
  <port id="0" precision="FP16">
@@ -6728,7 +6728,7 @@
6728
  </port>
6729
  </output>
6730
  </layer>
6731
- <layer id="446" name="Constant_31135" type="Const" version="opset1">
6732
  <data element_type="f32" shape="1, 1, 5120" offset="91101182" size="20480" />
6733
  <output>
6734
  <port id="0" precision="FP32">
@@ -6787,7 +6787,7 @@
6787
  </port>
6788
  </output>
6789
  </layer>
6790
- <layer id="450" name="Convert_42690" type="Convert" version="opset1">
6791
  <data destination_type="f16" />
6792
  <input>
6793
  <port id="0" precision="U4">
@@ -6814,7 +6814,7 @@
6814
  </port>
6815
  </output>
6816
  </layer>
6817
- <layer id="452" name="Convert_42692" type="Convert" version="opset1">
6818
  <data destination_type="f16" />
6819
  <input>
6820
  <port id="0" precision="U4">
@@ -6885,7 +6885,7 @@
6885
  </port>
6886
  </output>
6887
  </layer>
6888
- <layer id="456" name="Constant_42697" type="Const" version="opset1">
6889
  <data element_type="i64" shape="2" offset="80820206" size="16" />
6890
  <output>
6891
  <port id="0" precision="I64">
@@ -6893,7 +6893,7 @@
6893
  </port>
6894
  </output>
6895
  </layer>
6896
- <layer id="457" name="Reshape_42698" type="Reshape" version="opset1">
6897
  <data special_zero="false" />
6898
  <input>
6899
  <port id="0" precision="FP16">
@@ -6948,7 +6948,7 @@
6948
  </port>
6949
  </output>
6950
  </layer>
6951
- <layer id="460" name="Constant_31136" type="Const" version="opset1">
6952
  <data element_type="f32" shape="1, 1, 1280" offset="94526462" size="5120" />
6953
  <output>
6954
  <port id="0" precision="FP32">
@@ -7030,7 +7030,7 @@
7030
  </port>
7031
  </output>
7032
  </layer>
7033
- <layer id="465" name="Constant_31137" type="Const" version="opset1">
7034
  <data element_type="f32" shape="1, 1, 1280" offset="94531582" size="5120" />
7035
  <output>
7036
  <port id="0" precision="FP32">
@@ -7062,7 +7062,7 @@
7062
  </port>
7063
  </output>
7064
  </layer>
7065
- <layer id="467" name="Constant_31138" type="Const" version="opset1">
7066
  <data element_type="f32" shape="1, 1, 1280" offset="94536702" size="5120" />
7067
  <output>
7068
  <port id="0" precision="FP32">
@@ -7115,7 +7115,7 @@
7115
  </port>
7116
  </output>
7117
  </layer>
7118
- <layer id="470" name="Result_25933" type="Result" version="opset1">
7119
  <input>
7120
  <port id="0" precision="FP32">
7121
  <dim>-1</dim>
@@ -7632,7 +7632,7 @@
7632
  <edge from-layer="469" from-port="2" to-layer="470" to-port="0" />
7633
  </edges>
7634
  <rt_info>
7635
- <Runtime_version value="2024.5.0-16901-32aaa2fbd96" />
7636
  <conversion_parameters>
7637
  <framework value="pytorch" />
7638
  <is_python_object value="True" />
@@ -7640,9 +7640,10 @@
7640
  <nncf>
7641
  <friendly_names_were_updated value="True" />
7642
  <weight_compression>
7643
- <advanced_parameters value="{'awq_params': {'subset_size': 32, 'percent_to_apply': 0.002, 'alpha_min': 0.0, 'alpha_max': 1.0, 'steps': 100}, 'scale_estimation_params': {'subset_size': 64, 'initial_steps': 5, 'scale_steps': 5, 'weight_penalty': -1.0}, 'gptq_params': {'damp_percent': 0.1, 'block_size': 128, 'subset_size': 128}, 'lora_correction_params': {'adapter_rank': 8, 'num_iterations': 3, 'apply_regularization': True, 'subset_size': 128, 'use_int8_adapters': True}}" />
7644
  <all_layers value="False" />
7645
  <awq value="False" />
 
7646
  <gptq value="False" />
7647
  <group_size value="128" />
7648
  <ignored_scope value="[]" />
@@ -7654,10 +7655,10 @@
7654
  </weight_compression>
7655
  </nncf>
7656
  <optimum>
7657
- <optimum_intel_version value="1.20.0.dev0+2559620" />
7658
- <optimum_version value="1.23.1" />
7659
- <pytorch_version value="2.5.0" />
7660
- <transformers_version value="4.45.2" />
7661
  </optimum>
7662
  </rt_info>
7663
  </net>
 
10
  </port>
11
  </output>
12
  </layer>
13
+ <layer id="0" name="Parameter_26207" type="Parameter" version="opset1">
14
  <data shape="?,?,?" element_type="f32" />
15
  <output>
16
  <port id="0" precision="FP32" names="encoder_hidden_states">
 
29
  </port>
30
  </output>
31
  </layer>
32
+ <layer id="3" name="Convert_42987" type="Convert" version="opset1">
33
  <data destination_type="f16" />
34
  <input>
35
  <port id="0" precision="U8">
 
53
  </port>
54
  </output>
55
  </layer>
56
+ <layer id="5" name="Convert_42990" type="Convert" version="opset1">
57
  <data destination_type="f16" />
58
  <input>
59
  <port id="0" precision="U8">
 
130
  </port>
131
  </output>
132
  </layer>
133
+ <layer id="10" name="Constant_31708" type="Const" version="opset1">
134
  <data element_type="i64" shape="2" offset="66544078" size="16" />
135
  <output>
136
  <port id="0" precision="I64">
 
207
  </port>
208
  </output>
209
  </layer>
210
+ <layer id="16" name="Convert_43218" type="Convert" version="opset1">
211
  <data destination_type="f16" />
212
  <input>
213
  <port id="0" precision="U8">
 
231
  </port>
232
  </output>
233
  </layer>
234
+ <layer id="18" name="Convert_43221" type="Convert" version="opset1">
235
  <data destination_type="f16" />
236
  <input>
237
  <port id="0" precision="U8">
 
314
  <port id="0" precision="I64" names="23" />
315
  </output>
316
  </layer>
317
+ <layer id="24" name="ShapeOf_31624" type="ShapeOf" version="opset3">
318
  <data output_type="i64" />
319
  <input>
320
  <port id="0" precision="I64">
 
328
  </port>
329
  </output>
330
  </layer>
331
+ <layer id="25" name="Constant_31625" type="Const" version="opset1">
332
  <data element_type="i64" shape="" offset="67118890" size="8" />
333
  <output>
334
  <port id="0" precision="I64" />
335
  </output>
336
  </layer>
337
+ <layer id="26" name="Constant_31626" type="Const" version="opset1">
338
  <data element_type="i64" shape="" offset="67118882" size="8" />
339
  <output>
340
  <port id="0" precision="I64" />
341
  </output>
342
  </layer>
343
+ <layer id="27" name="Gather_31627" type="Gather" version="opset8">
344
  <data batch_dims="0" />
345
  <input>
346
  <port id="0" precision="I64">
 
401
  </port>
402
  </output>
403
  </layer>
404
+ <layer id="32" name="__module.model.model.decoder.embed_positions/aten::index/Constant" type="Const" version="opset1">
405
  <data element_type="i32" shape="" offset="66544094" size="4" />
406
  <output>
407
  <port id="0" precision="I32" />
 
478
  </port>
479
  </output>
480
  </layer>
481
+ <layer id="37" name="Constant_31390" type="Const" version="opset1">
482
  <data element_type="f32" shape="1, 1, 1280" offset="67118906" size="5120" />
483
  <output>
484
  <port id="0" precision="FP32">
 
510
  </port>
511
  </output>
512
  </layer>
513
+ <layer id="39" name="Constant_31391" type="Const" version="opset1">
514
  <data element_type="f32" shape="1, 1, 1280" offset="67124026" size="5120" />
515
  <output>
516
  <port id="0" precision="FP32">
 
542
  </port>
543
  </output>
544
  </layer>
545
+ <layer id="41" name="Constant_31293" type="Const" version="opset1">
546
  <data element_type="u4" shape="1280, 10, 128" offset="67129146" size="819200" />
547
  <output>
548
  <port id="0" precision="U4">
 
552
  </port>
553
  </output>
554
  </layer>
555
+ <layer id="42" name="Convert_43107" type="Convert" version="opset1">
556
  <data destination_type="f16" />
557
  <input>
558
  <port id="0" precision="U4">
 
569
  </port>
570
  </output>
571
  </layer>
572
+ <layer id="43" name="Constant_31293/zero_point" type="Const" version="opset1">
573
  <data element_type="u4" shape="1280, 10, 1" offset="67948346" size="6400" />
574
  <output>
575
  <port id="0" precision="U4">
 
579
  </port>
580
  </output>
581
  </layer>
582
+ <layer id="44" name="Convert_43109" type="Convert" version="opset1">
583
  <data destination_type="f16" />
584
  <input>
585
  <port id="0" precision="U4">
 
596
  </port>
597
  </output>
598
  </layer>
599
+ <layer id="45" name="Constant_31293/zero_point/subtract" type="Subtract" version="opset1">
600
  <data auto_broadcast="numpy" />
601
  <input>
602
  <port id="0" precision="FP16">
 
618
  </port>
619
  </output>
620
  </layer>
621
+ <layer id="46" name="Constant_31293/scale" type="Const" version="opset1">
622
  <data element_type="f16" shape="1280, 10, 1" offset="67954746" size="25600" />
623
  <output>
624
  <port id="0" precision="FP16">
 
628
  </port>
629
  </output>
630
  </layer>
631
+ <layer id="47" name="Constant_31293/fq_weights_1" type="Multiply" version="opset1">
632
  <data auto_broadcast="numpy" />
633
  <input>
634
  <port id="0" precision="FP16">
 
650
  </port>
651
  </output>
652
  </layer>
653
+ <layer id="48" name="Constant_43114" type="Const" version="opset1">
654
  <data element_type="i64" shape="2" offset="67980346" size="16" />
655
  <output>
656
  <port id="0" precision="I64">
 
658
  </port>
659
  </output>
660
  </layer>
661
+ <layer id="49" name="Reshape_43115" type="Reshape" version="opset1">
662
  <data special_zero="false" />
663
  <input>
664
  <port id="0" precision="FP16">
 
677
  </port>
678
  </output>
679
  </layer>
680
+ <layer id="50" name="Constant_31293/fq_weights_1/convert" type="Convert" version="opset1">
681
  <data destination_type="f32" />
682
  <input>
683
  <port id="0" precision="FP16">
 
692
  </port>
693
  </output>
694
  </layer>
695
+ <layer id="51" name="Multiply_31276" type="MatMul" version="opset1">
696
  <data transpose_a="false" transpose_b="true" />
697
  <input>
698
  <port id="0" precision="FP32">
 
713
  </port>
714
  </output>
715
  </layer>
716
+ <layer id="52" name="Constant_31392" type="Const" version="opset1">
717
  <data element_type="f32" shape="1, 1, 1280" offset="67980362" size="5120" />
718
  <output>
719
  <port id="0" precision="FP32">
 
745
  </port>
746
  </output>
747
  </layer>
748
+ <layer id="54" name="Constant_31709" type="Const" version="opset1">
749
  <data element_type="i64" shape="4" offset="67985482" size="32" />
750
  <output>
751
  <port id="0" precision="I64">
 
813
  </port>
814
  </output>
815
  </layer>
816
+ <layer id="59" name="Convert_43096" type="Convert" version="opset1">
817
  <data destination_type="f16" />
818
  <input>
819
  <port id="0" precision="U4">
 
840
  </port>
841
  </output>
842
  </layer>
843
+ <layer id="61" name="Convert_43098" type="Convert" version="opset1">
844
  <data destination_type="f16" />
845
  <input>
846
  <port id="0" precision="U4">
 
911
  </port>
912
  </output>
913
  </layer>
914
+ <layer id="65" name="Constant_43103" type="Const" version="opset1">
915
  <data element_type="i64" shape="2" offset="67980346" size="16" />
916
  <output>
917
  <port id="0" precision="I64">
 
919
  </port>
920
  </output>
921
  </layer>
922
+ <layer id="66" name="Reshape_43104" type="Reshape" version="opset1">
923
  <data special_zero="false" />
924
  <input>
925
  <port id="0" precision="FP16">
 
974
  </port>
975
  </output>
976
  </layer>
977
+ <layer id="69" name="Constant_31710" type="Const" version="opset1">
978
  <data element_type="i64" shape="4" offset="67985482" size="32" />
979
  <output>
980
  <port id="0" precision="I64">
 
1063
  <port id="0" precision="I32" />
1064
  </output>
1065
  </layer>
1066
+ <layer id="75" name="17" type="Const" version="opset1">
1067
  <data element_type="f32" shape="" offset="68836730" size="4" />
1068
  <output>
1069
+ <port id="0" precision="FP32" names="17" />
1070
  </output>
1071
  </layer>
1072
+ <layer id="76" name="Constant_31648" type="Const" version="opset1">
1073
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1074
  <output>
1075
  <port id="0" precision="I64">
 
1077
  </port>
1078
  </output>
1079
  </layer>
1080
+ <layer id="77" name="Reshape_31649" type="Reshape" version="opset1">
1081
  <data special_zero="false" />
1082
  <input>
1083
  <port id="0" precision="I64" />
 
1107
  <port id="2" precision="I64" names="62,64" />
1108
  </output>
1109
  </layer>
1110
+ <layer id="80" name="Constant_24084" type="Const" version="opset1">
1111
  <data element_type="i32" shape="" offset="66544094" size="4" />
1112
  <output>
1113
  <port id="0" precision="I32" />
1114
  </output>
1115
  </layer>
1116
+ <layer id="81" name="Unsqueeze_24085" type="Unsqueeze" version="opset1">
1117
  <input>
1118
  <port id="0" precision="I64" />
1119
  <port id="1" precision="I32" />
 
1155
  </port>
1156
  </output>
1157
  </layer>
1158
+ <layer id="84" name="ShapeOf_31655" type="ShapeOf" version="opset3">
1159
  <data output_type="i32" />
1160
  <input>
1161
  <port id="0" precision="FP32">
 
1169
  </port>
1170
  </output>
1171
  </layer>
1172
+ <layer id="85" name="Constant_31656" type="Const" version="opset1">
1173
  <data element_type="i64" shape="" offset="67118890" size="8" />
1174
  <output>
1175
  <port id="0" precision="I64" />
1176
  </output>
1177
  </layer>
1178
+ <layer id="86" name="Constant_31657" type="Const" version="opset1">
1179
  <data element_type="i64" shape="" offset="67118882" size="8" />
1180
  <output>
1181
  <port id="0" precision="I64" />
1182
  </output>
1183
  </layer>
1184
+ <layer id="87" name="Gather_31658" type="Gather" version="opset8">
1185
  <data batch_dims="0" />
1186
  <input>
1187
  <port id="0" precision="I32">
 
1233
  <port id="0" precision="I32" />
1234
  </output>
1235
  </layer>
1236
+ <layer id="92" name="Convert_31661" type="Convert" version="opset1">
1237
  <data destination_type="i32" />
1238
  <input>
1239
  <port id="0" precision="I64" />
 
1375
  </port>
1376
  </output>
1377
  </layer>
1378
+ <layer id="104" name="Constant_24187" type="Const" version="opset1">
1379
  <data element_type="i64" shape="2" offset="68836734" size="16" />
1380
  <output>
1381
  <port id="0" precision="I64" names="70">
 
1492
  </port>
1493
  </output>
1494
  </layer>
1495
+ <layer id="112" name="Constant_31639" type="Const" version="opset1">
1496
  <data element_type="i64" shape="1" offset="67118882" size="8" />
1497
  <output>
1498
  <port id="0" precision="I64">
 
1500
  </port>
1501
  </output>
1502
  </layer>
1503
+ <layer id="113" name="Constant_31640" type="Const" version="opset1">
1504
  <data element_type="i64" shape="" offset="67118882" size="8" />
1505
  <output>
1506
  <port id="0" precision="I64" />
1507
  </output>
1508
  </layer>
1509
+ <layer id="114" name="Gather_31641" type="Gather" version="opset8">
1510
  <data batch_dims="0" />
1511
  <input>
1512
  <port id="0" precision="I64">
 
1523
  </port>
1524
  </output>
1525
  </layer>
1526
+ <layer id="115" name="Constant_30625" type="Const" version="opset1">
1527
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1528
  <output>
1529
  <port id="0" precision="I64">
 
1531
  </port>
1532
  </output>
1533
  </layer>
1534
+ <layer id="116" name="Constant_30627" type="Const" version="opset1">
1535
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1536
  <output>
1537
  <port id="0" precision="I64">
 
1539
  </port>
1540
  </output>
1541
  </layer>
1542
+ <layer id="117" name="Constant_30629" type="Const" version="opset1">
1543
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1544
  <output>
1545
  <port id="0" precision="I64">
 
1591
  </port>
1592
  </output>
1593
  </layer>
1594
+ <layer id="120" name="Constant_24621" type="Const" version="opset1">
1595
  <data element_type="i64" shape="1" offset="67118882" size="8" />
1596
  <output>
1597
  <port id="0" precision="I64">
 
1599
  </port>
1600
  </output>
1601
  </layer>
1602
+ <layer id="121" name="ShapeOf_31670" type="ShapeOf" version="opset3">
1603
  <data output_type="i64" />
1604
  <input>
1605
  <port id="0" precision="FP32">
 
1614
  </port>
1615
  </output>
1616
  </layer>
1617
+ <layer id="122" name="Constant_31671" type="Const" version="opset1">
1618
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1619
  <output>
1620
  <port id="0" precision="I64">
 
1622
  </port>
1623
  </output>
1624
  </layer>
1625
+ <layer id="123" name="Constant_31672" type="Const" version="opset1">
1626
  <data element_type="i64" shape="" offset="67118882" size="8" />
1627
  <output>
1628
  <port id="0" precision="I64" />
1629
  </output>
1630
  </layer>
1631
+ <layer id="124" name="Gather_31673" type="Gather" version="opset8">
1632
  <data batch_dims="0" />
1633
  <input>
1634
  <port id="0" precision="I64">
 
1645
  </port>
1646
  </output>
1647
  </layer>
1648
+ <layer id="125" name="Constant_24624" type="Const" version="opset1">
1649
  <data element_type="i64" shape="1" offset="67118890" size="8" />
1650
  <output>
1651
  <port id="0" precision="I64">
 
1745
  </port>
1746
  </output>
1747
  </layer>
1748
+ <layer id="131" name="Convert_43085" type="Convert" version="opset1">
1749
  <data destination_type="f16" />
1750
  <input>
1751
  <port id="0" precision="U4">
 
1772
  </port>
1773
  </output>
1774
  </layer>
1775
+ <layer id="133" name="Convert_43087" type="Convert" version="opset1">
1776
  <data destination_type="f16" />
1777
  <input>
1778
  <port id="0" precision="U4">
 
1843
  </port>
1844
  </output>
1845
  </layer>
1846
+ <layer id="137" name="Constant_43092" type="Const" version="opset1">
1847
  <data element_type="i64" shape="2" offset="67980346" size="16" />
1848
  <output>
1849
  <port id="0" precision="I64">
 
1851
  </port>
1852
  </output>
1853
  </layer>
1854
+ <layer id="138" name="Reshape_43093" type="Reshape" version="opset1">
1855
  <data special_zero="false" />
1856
  <input>
1857
  <port id="0" precision="FP16">
 
1906
  </port>
1907
  </output>
1908
  </layer>
1909
+ <layer id="141" name="Constant_31393" type="Const" version="opset1">
1910
  <data element_type="f32" shape="1, 1, 1280" offset="69687958" size="5120" />
1911
  <output>
1912
  <port id="0" precision="FP32">
 
1938
  </port>
1939
  </output>
1940
  </layer>
1941
+ <layer id="143" name="Constant_31711" type="Const" version="opset1">
1942
  <data element_type="i64" shape="4" offset="67985482" size="32" />
1943
  <output>
1944
  <port id="0" precision="I64">
 
2050
  </port>
2051
  </output>
2052
  </layer>
2053
+ <layer id="150" name="Constant_31712" type="Const" version="opset1">
2054
  <data element_type="i64" shape="3" offset="69693078" size="24" />
2055
  <output>
2056
  <port id="0" precision="I64">
 
2089
  </port>
2090
  </output>
2091
  </layer>
2092
+ <layer id="153" name="Convert_43118" type="Convert" version="opset1">
2093
  <data destination_type="f16" />
2094
  <input>
2095
  <port id="0" precision="U4">
 
2116
  </port>
2117
  </output>
2118
  </layer>
2119
+ <layer id="155" name="Convert_43120" type="Convert" version="opset1">
2120
  <data destination_type="f16" />
2121
  <input>
2122
  <port id="0" precision="U4">
 
2187
  </port>
2188
  </output>
2189
  </layer>
2190
+ <layer id="159" name="Constant_43125" type="Const" version="opset1">
2191
  <data element_type="i64" shape="2" offset="67980346" size="16" />
2192
  <output>
2193
  <port id="0" precision="I64">
 
2195
  </port>
2196
  </output>
2197
  </layer>
2198
+ <layer id="160" name="Reshape_43126" type="Reshape" version="opset1">
2199
  <data special_zero="false" />
2200
  <input>
2201
  <port id="0" precision="FP16">
 
2250
  </port>
2251
  </output>
2252
  </layer>
2253
+ <layer id="163" name="Constant_31394" type="Const" version="opset1">
2254
  <data element_type="f32" shape="1, 1, 1280" offset="70544302" size="5120" />
2255
  <output>
2256
  <port id="0" precision="FP32">
 
2332
  </port>
2333
  </output>
2334
  </layer>
2335
+ <layer id="168" name="Constant_31395" type="Const" version="opset1">
2336
  <data element_type="f32" shape="1, 1, 1280" offset="70549422" size="5120" />
2337
  <output>
2338
  <port id="0" precision="FP32">
 
2364
  </port>
2365
  </output>
2366
  </layer>
2367
+ <layer id="170" name="Constant_31396" type="Const" version="opset1">
2368
  <data element_type="f32" shape="1, 1, 1280" offset="70554542" size="5120" />
2369
  <output>
2370
  <port id="0" precision="FP32">
 
2396
  </port>
2397
  </output>
2398
  </layer>
2399
+ <layer id="172" name="Constant_31296" type="Const" version="opset1">
2400
  <data element_type="u4" shape="1280, 10, 128" offset="70559662" size="819200" />
2401
  <output>
2402
  <port id="0" precision="U4">
 
2406
  </port>
2407
  </output>
2408
  </layer>
2409
+ <layer id="173" name="Convert_43129" type="Convert" version="opset1">
2410
  <data destination_type="f16" />
2411
  <input>
2412
  <port id="0" precision="U4">
 
2423
  </port>
2424
  </output>
2425
  </layer>
2426
+ <layer id="174" name="Constant_31296/zero_point" type="Const" version="opset1">
2427
  <data element_type="u4" shape="1280, 10, 1" offset="71378862" size="6400" />
2428
  <output>
2429
  <port id="0" precision="U4">
 
2433
  </port>
2434
  </output>
2435
  </layer>
2436
+ <layer id="175" name="Convert_43131" type="Convert" version="opset1">
2437
  <data destination_type="f16" />
2438
  <input>
2439
  <port id="0" precision="U4">
 
2450
  </port>
2451
  </output>
2452
  </layer>
2453
+ <layer id="176" name="Constant_31296/zero_point/subtract" type="Subtract" version="opset1">
2454
  <data auto_broadcast="numpy" />
2455
  <input>
2456
  <port id="0" precision="FP16">
 
2472
  </port>
2473
  </output>
2474
  </layer>
2475
+ <layer id="177" name="Constant_31296/scale" type="Const" version="opset1">
2476
  <data element_type="f16" shape="1280, 10, 1" offset="71385262" size="25600" />
2477
  <output>
2478
  <port id="0" precision="FP16">
 
2482
  </port>
2483
  </output>
2484
  </layer>
2485
+ <layer id="178" name="Constant_31296/fq_weights_1" type="Multiply" version="opset1">
2486
  <data auto_broadcast="numpy" />
2487
  <input>
2488
  <port id="0" precision="FP16">
 
2504
  </port>
2505
  </output>
2506
  </layer>
2507
+ <layer id="179" name="Constant_43136" type="Const" version="opset1">
2508
  <data element_type="i64" shape="2" offset="67980346" size="16" />
2509
  <output>
2510
  <port id="0" precision="I64">
 
2512
  </port>
2513
  </output>
2514
  </layer>
2515
+ <layer id="180" name="Reshape_43137" type="Reshape" version="opset1">
2516
  <data special_zero="false" />
2517
  <input>
2518
  <port id="0" precision="FP16">
 
2531
  </port>
2532
  </output>
2533
  </layer>
2534
+ <layer id="181" name="Constant_31296/fq_weights_1/convert" type="Convert" version="opset1">
2535
  <data destination_type="f32" />
2536
  <input>
2537
  <port id="0" precision="FP16">
 
2546
  </port>
2547
  </output>
2548
  </layer>
2549
+ <layer id="182" name="Multiply_31280" type="MatMul" version="opset1">
2550
  <data transpose_a="false" transpose_b="true" />
2551
  <input>
2552
  <port id="0" precision="FP32">
 
2567
  </port>
2568
  </output>
2569
  </layer>
2570
+ <layer id="183" name="Constant_31397" type="Const" version="opset1">
2571
  <data element_type="f32" shape="1, 1, 1280" offset="71410862" size="5120" />
2572
  <output>
2573
  <port id="0" precision="FP32">
 
2599
  </port>
2600
  </output>
2601
  </layer>
2602
+ <layer id="185" name="Constant_31713" type="Const" version="opset1">
2603
  <data element_type="i64" shape="4" offset="67985482" size="32" />
2604
  <output>
2605
  <port id="0" precision="I64">
 
2667
  </port>
2668
  </output>
2669
  </layer>
2670
+ <layer id="190" name="Convert_43074" type="Convert" version="opset1">
2671
  <data destination_type="f16" />
2672
  <input>
2673
  <port id="0" precision="U4">
 
2694
  </port>
2695
  </output>
2696
  </layer>
2697
+ <layer id="192" name="Convert_43076" type="Convert" version="opset1">
2698
  <data destination_type="f16" />
2699
  <input>
2700
  <port id="0" precision="U4">
 
2765
  </port>
2766
  </output>
2767
  </layer>
2768
+ <layer id="196" name="Constant_43081" type="Const" version="opset1">
2769
  <data element_type="i64" shape="2" offset="67980346" size="16" />
2770
  <output>
2771
  <port id="0" precision="I64">
 
2773
  </port>
2774
  </output>
2775
  </layer>
2776
+ <layer id="197" name="Reshape_43082" type="Reshape" version="opset1">
2777
  <data special_zero="false" />
2778
  <input>
2779
  <port id="0" precision="FP16">
 
2828
  </port>
2829
  </output>
2830
  </layer>
2831
+ <layer id="200" name="Constant_24369" type="Const" version="opset1">
2832
  <data element_type="i64" shape="1" offset="72267182" size="8" />
2833
  <output>
2834
  <port id="0" precision="I64">
 
2836
  </port>
2837
  </output>
2838
  </layer>
2839
+ <layer id="201" name="Constant_24370" type="Const" version="opset1">
2840
  <data element_type="i64" shape="1" offset="72267190" size="8" />
2841
  <output>
2842
  <port id="0" precision="I64">
 
2844
  </port>
2845
  </output>
2846
  </layer>
2847
+ <layer id="202" name="Constant_24371" type="Const" version="opset1">
2848
  <data element_type="i64" shape="1" offset="72267198" size="8" />
2849
  <output>
2850
  <port id="0" precision="I64">
 
2852
  </port>
2853
  </output>
2854
  </layer>
2855
+ <layer id="203" name="__module.model.model.decoder.layers.0.self_attn/prim::ListConstruct_1" type="Concat" version="opset1">
2856
  <data axis="0" />
2857
  <input>
2858
  <port id="0" precision="I64">
 
2978
  </port>
2979
  </output>
2980
  </layer>
2981
+ <layer id="210" name="Convert_43063" type="Convert" version="opset1">
2982
  <data destination_type="f16" />
2983
  <input>
2984
  <port id="0" precision="U4">
 
3005
  </port>
3006
  </output>
3007
  </layer>
3008
+ <layer id="212" name="Convert_43065" type="Convert" version="opset1">
3009
  <data destination_type="f16" />
3010
  <input>
3011
  <port id="0" precision="U4">
 
3076
  </port>
3077
  </output>
3078
  </layer>
3079
+ <layer id="216" name="Constant_43070" type="Const" version="opset1">
3080
  <data element_type="i64" shape="2" offset="67980346" size="16" />
3081
  <output>
3082
  <port id="0" precision="I64">
 
3084
  </port>
3085
  </output>
3086
  </layer>
3087
+ <layer id="217" name="Reshape_43071" type="Reshape" version="opset1">
3088
  <data special_zero="false" />
3089
  <input>
3090
  <port id="0" precision="FP16">
 
3139
  </port>
3140
  </output>
3141
  </layer>
3142
+ <layer id="220" name="Constant_31398" type="Const" version="opset1">
3143
  <data element_type="f32" shape="1, 1, 1280" offset="73118406" size="5120" />
3144
  <output>
3145
  <port id="0" precision="FP32">
 
3275
  </port>
3276
  </output>
3277
  </layer>
3278
+ <layer id="228" name="Constant_31714" type="Const" version="opset1">
3279
  <data element_type="i64" shape="3" offset="73123526" size="24" />
3280
  <output>
3281
  <port id="0" precision="I64">
 
3314
  </port>
3315
  </output>
3316
  </layer>
3317
+ <layer id="231" name="Convert_43140" type="Convert" version="opset1">
3318
  <data destination_type="f16" />
3319
  <input>
3320
  <port id="0" precision="U4">
 
3341
  </port>
3342
  </output>
3343
  </layer>
3344
+ <layer id="233" name="Convert_43142" type="Convert" version="opset1">
3345
  <data destination_type="f16" />
3346
  <input>
3347
  <port id="0" precision="U4">
 
3412
  </port>
3413
  </output>
3414
  </layer>
3415
+ <layer id="237" name="Constant_43147" type="Const" version="opset1">
3416
  <data element_type="i64" shape="2" offset="67980346" size="16" />
3417
  <output>
3418
  <port id="0" precision="I64">
 
3420
  </port>
3421
  </output>
3422
  </layer>
3423
+ <layer id="238" name="Reshape_43148" type="Reshape" version="opset1">
3424
  <data special_zero="false" />
3425
  <input>
3426
  <port id="0" precision="FP16">
 
3475
  </port>
3476
  </output>
3477
  </layer>
3478
+ <layer id="241" name="Constant_31399" type="Const" version="opset1">
3479
  <data element_type="f32" shape="1, 1, 1280" offset="73974750" size="5120" />
3480
  <output>
3481
  <port id="0" precision="FP32">
 
3557
  </port>
3558
  </output>
3559
  </layer>
3560
+ <layer id="246" name="Constant_31400" type="Const" version="opset1">
3561
  <data element_type="f32" shape="1, 1, 1280" offset="73979870" size="5120" />
3562
  <output>
3563
  <port id="0" precision="FP32">
 
3589
  </port>
3590
  </output>
3591
  </layer>
3592
+ <layer id="248" name="Constant_31401" type="Const" version="opset1">
3593
  <data element_type="f32" shape="1, 1, 1280" offset="73984990" size="5120" />
3594
  <output>
3595
  <port id="0" precision="FP32">
 
3631
  </port>
3632
  </output>
3633
  </layer>
3634
+ <layer id="251" name="Convert_42997" type="Convert" version="opset1">
3635
  <data destination_type="f16" />
3636
  <input>
3637
  <port id="0" precision="U4">
 
3658
  </port>
3659
  </output>
3660
  </layer>
3661
+ <layer id="253" name="Convert_42999" type="Convert" version="opset1">
3662
  <data destination_type="f16" />
3663
  <input>
3664
  <port id="0" precision="U4">
 
3729
  </port>
3730
  </output>
3731
  </layer>
3732
+ <layer id="257" name="Constant_43004" type="Const" version="opset1">
3733
  <data element_type="i64" shape="2" offset="77394910" size="16" />
3734
  <output>
3735
  <port id="0" precision="I64">
 
3737
  </port>
3738
  </output>
3739
  </layer>
3740
+ <layer id="258" name="Reshape_43005" type="Reshape" version="opset1">
3741
  <data special_zero="false" />
3742
  <input>
3743
  <port id="0" precision="FP16">
 
3792
  </port>
3793
  </output>
3794
  </layer>
3795
+ <layer id="261" name="Constant_31402" type="Const" version="opset1">
3796
  <data element_type="f32" shape="1, 1, 5120" offset="77394926" size="20480" />
3797
  <output>
3798
  <port id="0" precision="FP32">
 
3851
  </port>
3852
  </output>
3853
  </layer>
3854
+ <layer id="265" name="Convert_43008" type="Convert" version="opset1">
3855
  <data destination_type="f16" />
3856
  <input>
3857
  <port id="0" precision="U4">
 
3878
  </port>
3879
  </output>
3880
  </layer>
3881
+ <layer id="267" name="Convert_43010" type="Convert" version="opset1">
3882
  <data destination_type="f16" />
3883
  <input>
3884
  <port id="0" precision="U4">
 
3949
  </port>
3950
  </output>
3951
  </layer>
3952
+ <layer id="271" name="Constant_43015" type="Const" version="opset1">
3953
  <data element_type="i64" shape="2" offset="80820206" size="16" />
3954
  <output>
3955
  <port id="0" precision="I64">
 
3957
  </port>
3958
  </output>
3959
  </layer>
3960
+ <layer id="272" name="Reshape_43016" type="Reshape" version="opset1">
3961
  <data special_zero="false" />
3962
  <input>
3963
  <port id="0" precision="FP16">
 
4012
  </port>
4013
  </output>
4014
  </layer>
4015
+ <layer id="275" name="Constant_31403" type="Const" version="opset1">
4016
  <data element_type="f32" shape="1, 1, 1280" offset="80820222" size="5120" />
4017
  <output>
4018
  <port id="0" precision="FP32">
 
4094
  </port>
4095
  </output>
4096
  </layer>
4097
+ <layer id="280" name="Constant_31404" type="Const" version="opset1">
4098
  <data element_type="f32" shape="1, 1, 1280" offset="80825342" size="5120" />
4099
  <output>
4100
  <port id="0" precision="FP32">
 
4126
  </port>
4127
  </output>
4128
  </layer>
4129
+ <layer id="282" name="Constant_31405" type="Const" version="opset1">
4130
  <data element_type="f32" shape="1, 1, 1280" offset="80830462" size="5120" />
4131
  <output>
4132
  <port id="0" precision="FP32">
 
4158
  </port>
4159
  </output>
4160
  </layer>
4161
+ <layer id="284" name="Constant_31299" type="Const" version="opset1">
4162
  <data element_type="u4" shape="1280, 10, 128" offset="80835582" size="819200" />
4163
  <output>
4164
  <port id="0" precision="U4">
 
4168
  </port>
4169
  </output>
4170
  </layer>
4171
+ <layer id="285" name="Convert_43173" type="Convert" version="opset1">
4172
  <data destination_type="f16" />
4173
  <input>
4174
  <port id="0" precision="U4">
 
4185
  </port>
4186
  </output>
4187
  </layer>
4188
+ <layer id="286" name="Constant_31299/zero_point" type="Const" version="opset1">
4189
  <data element_type="u4" shape="1280, 10, 1" offset="81654782" size="6400" />
4190
  <output>
4191
  <port id="0" precision="U4">
 
4195
  </port>
4196
  </output>
4197
  </layer>
4198
+ <layer id="287" name="Convert_43175" type="Convert" version="opset1">
4199
  <data destination_type="f16" />
4200
  <input>
4201
  <port id="0" precision="U4">
 
4212
  </port>
4213
  </output>
4214
  </layer>
4215
+ <layer id="288" name="Constant_31299/zero_point/subtract" type="Subtract" version="opset1">
4216
  <data auto_broadcast="numpy" />
4217
  <input>
4218
  <port id="0" precision="FP16">
 
4234
  </port>
4235
  </output>
4236
  </layer>
4237
+ <layer id="289" name="Constant_31299/scale" type="Const" version="opset1">
4238
  <data element_type="f16" shape="1280, 10, 1" offset="81661182" size="25600" />
4239
  <output>
4240
  <port id="0" precision="FP16">
 
4244
  </port>
4245
  </output>
4246
  </layer>
4247
+ <layer id="290" name="Constant_31299/fq_weights_1" type="Multiply" version="opset1">
4248
  <data auto_broadcast="numpy" />
4249
  <input>
4250
  <port id="0" precision="FP16">
 
4266
  </port>
4267
  </output>
4268
  </layer>
4269
+ <layer id="291" name="Constant_43180" type="Const" version="opset1">
4270
  <data element_type="i64" shape="2" offset="67980346" size="16" />
4271
  <output>
4272
  <port id="0" precision="I64">
 
4274
  </port>
4275
  </output>
4276
  </layer>
4277
+ <layer id="292" name="Reshape_43181" type="Reshape" version="opset1">
4278
  <data special_zero="false" />
4279
  <input>
4280
  <port id="0" precision="FP16">
 
4293
  </port>
4294
  </output>
4295
  </layer>
4296
+ <layer id="293" name="Constant_31299/fq_weights_1/convert" type="Convert" version="opset1">
4297
  <data destination_type="f32" />
4298
  <input>
4299
  <port id="0" precision="FP16">
 
4308
  </port>
4309
  </output>
4310
  </layer>
4311
+ <layer id="294" name="Multiply_31284" type="MatMul" version="opset1">
4312
  <data transpose_a="false" transpose_b="true" />
4313
  <input>
4314
  <port id="0" precision="FP32">
 
4329
  </port>
4330
  </output>
4331
  </layer>
4332
+ <layer id="295" name="Constant_31406" type="Const" version="opset1">
4333
  <data element_type="f32" shape="1, 1, 1280" offset="81686782" size="5120" />
4334
  <output>
4335
  <port id="0" precision="FP32">
 
4361
  </port>
4362
  </output>
4363
  </layer>
4364
+ <layer id="297" name="Constant_31715" type="Const" version="opset1">
4365
  <data element_type="i64" shape="4" offset="67985482" size="32" />
4366
  <output>
4367
  <port id="0" precision="I64">
 
4429
  </port>
4430
  </output>
4431
  </layer>
4432
+ <layer id="302" name="Convert_43162" type="Convert" version="opset1">
4433
  <data destination_type="f16" />
4434
  <input>
4435
  <port id="0" precision="U4">
 
4456
  </port>
4457
  </output>
4458
  </layer>
4459
+ <layer id="304" name="Convert_43164" type="Convert" version="opset1">
4460
  <data destination_type="f16" />
4461
  <input>
4462
  <port id="0" precision="U4">
 
4527
  </port>
4528
  </output>
4529
  </layer>
4530
+ <layer id="308" name="Constant_43169" type="Const" version="opset1">
4531
  <data element_type="i64" shape="2" offset="67980346" size="16" />
4532
  <output>
4533
  <port id="0" precision="I64">
 
4535
  </port>
4536
  </output>
4537
  </layer>
4538
+ <layer id="309" name="Reshape_43170" type="Reshape" version="opset1">
4539
  <data special_zero="false" />
4540
  <input>
4541
  <port id="0" precision="FP16">
 
4590
  </port>
4591
  </output>
4592
  </layer>
4593
+ <layer id="312" name="Constant_31716" type="Const" version="opset1">
4594
  <data element_type="i64" shape="4" offset="67985482" size="32" />
4595
  <output>
4596
  <port id="0" precision="I64">
 
4727
  </port>
4728
  </output>
4729
  </layer>
4730
+ <layer id="320" name="Convert_43151" type="Convert" version="opset1">
4731
  <data destination_type="f16" />
4732
  <input>
4733
  <port id="0" precision="U4">
 
4754
  </port>
4755
  </output>
4756
  </layer>
4757
+ <layer id="322" name="Convert_43153" type="Convert" version="opset1">
4758
  <data destination_type="f16" />
4759
  <input>
4760
  <port id="0" precision="U4">
 
4825
  </port>
4826
  </output>
4827
  </layer>
4828
+ <layer id="326" name="Constant_43158" type="Const" version="opset1">
4829
  <data element_type="i64" shape="2" offset="67980346" size="16" />
4830
  <output>
4831
  <port id="0" precision="I64">
 
4833
  </port>
4834
  </output>
4835
  </layer>
4836
+ <layer id="327" name="Reshape_43159" type="Reshape" version="opset1">
4837
  <data special_zero="false" />
4838
  <input>
4839
  <port id="0" precision="FP16">
 
4888
  </port>
4889
  </output>
4890
  </layer>
4891
+ <layer id="330" name="Constant_31407" type="Const" version="opset1">
4892
  <data element_type="f32" shape="1, 1, 1280" offset="83394302" size="5120" />
4893
  <output>
4894
  <port id="0" precision="FP32">
 
4920
  </port>
4921
  </output>
4922
  </layer>
4923
+ <layer id="332" name="Constant_31717" type="Const" version="opset1">
4924
  <data element_type="i64" shape="4" offset="67985482" size="32" />
4925
  <output>
4926
  <port id="0" precision="I64">
 
5032
  </port>
5033
  </output>
5034
  </layer>
5035
+ <layer id="339" name="Constant_31718" type="Const" version="opset1">
5036
  <data element_type="i64" shape="3" offset="69693078" size="24" />
5037
  <output>
5038
  <port id="0" precision="I64">
 
5071
  </port>
5072
  </output>
5073
  </layer>
5074
+ <layer id="342" name="Convert_43184" type="Convert" version="opset1">
5075
  <data destination_type="f16" />
5076
  <input>
5077
  <port id="0" precision="U4">
 
5098
  </port>
5099
  </output>
5100
  </layer>
5101
+ <layer id="344" name="Convert_43186" type="Convert" version="opset1">
5102
  <data destination_type="f16" />
5103
  <input>
5104
  <port id="0" precision="U4">
 
5169
  </port>
5170
  </output>
5171
  </layer>
5172
+ <layer id="348" name="Constant_43191" type="Const" version="opset1">
5173
  <data element_type="i64" shape="2" offset="67980346" size="16" />
5174
  <output>
5175
  <port id="0" precision="I64">
 
5177
  </port>
5178
  </output>
5179
  </layer>
5180
+ <layer id="349" name="Reshape_43192" type="Reshape" version="opset1">
5181
  <data special_zero="false" />
5182
  <input>
5183
  <port id="0" precision="FP16">
 
5232
  </port>
5233
  </output>
5234
  </layer>
5235
+ <layer id="352" name="Constant_31408" type="Const" version="opset1">
5236
  <data element_type="f32" shape="1, 1, 1280" offset="84250622" size="5120" />
5237
  <output>
5238
  <port id="0" precision="FP32">
 
5314
  </port>
5315
  </output>
5316
  </layer>
5317
+ <layer id="357" name="Constant_31409" type="Const" version="opset1">
5318
  <data element_type="f32" shape="1, 1, 1280" offset="84255742" size="5120" />
5319
  <output>
5320
  <port id="0" precision="FP32">
 
5346
  </port>
5347
  </output>
5348
  </layer>
5349
+ <layer id="359" name="Constant_31410" type="Const" version="opset1">
5350
  <data element_type="f32" shape="1, 1, 1280" offset="84260862" size="5120" />
5351
  <output>
5352
  <port id="0" precision="FP32">
 
5378
  </port>
5379
  </output>
5380
  </layer>
5381
+ <layer id="361" name="Constant_31302" type="Const" version="opset1">
5382
  <data element_type="u4" shape="1280, 10, 128" offset="84265982" size="819200" />
5383
  <output>
5384
  <port id="0" precision="U4">
 
5388
  </port>
5389
  </output>
5390
  </layer>
5391
+ <layer id="362" name="Convert_43195" type="Convert" version="opset1">
5392
  <data destination_type="f16" />
5393
  <input>
5394
  <port id="0" precision="U4">
 
5405
  </port>
5406
  </output>
5407
  </layer>
5408
+ <layer id="363" name="Constant_31302/zero_point" type="Const" version="opset1">
5409
  <data element_type="u4" shape="1280, 10, 1" offset="85085182" size="6400" />
5410
  <output>
5411
  <port id="0" precision="U4">
 
5415
  </port>
5416
  </output>
5417
  </layer>
5418
+ <layer id="364" name="Convert_43197" type="Convert" version="opset1">
5419
  <data destination_type="f16" />
5420
  <input>
5421
  <port id="0" precision="U4">
 
5432
  </port>
5433
  </output>
5434
  </layer>
5435
+ <layer id="365" name="Constant_31302/zero_point/subtract" type="Subtract" version="opset1">
5436
  <data auto_broadcast="numpy" />
5437
  <input>
5438
  <port id="0" precision="FP16">
 
5454
  </port>
5455
  </output>
5456
  </layer>
5457
+ <layer id="366" name="Constant_31302/scale" type="Const" version="opset1">
5458
  <data element_type="f16" shape="1280, 10, 1" offset="85091582" size="25600" />
5459
  <output>
5460
  <port id="0" precision="FP16">
 
5464
  </port>
5465
  </output>
5466
  </layer>
5467
+ <layer id="367" name="Constant_31302/fq_weights_1" type="Multiply" version="opset1">
5468
  <data auto_broadcast="numpy" />
5469
  <input>
5470
  <port id="0" precision="FP16">
 
5486
  </port>
5487
  </output>
5488
  </layer>
5489
+ <layer id="368" name="Constant_43202" type="Const" version="opset1">
5490
  <data element_type="i64" shape="2" offset="67980346" size="16" />
5491
  <output>
5492
  <port id="0" precision="I64">
 
5494
  </port>
5495
  </output>
5496
  </layer>
5497
+ <layer id="369" name="Reshape_43203" type="Reshape" version="opset1">
5498
  <data special_zero="false" />
5499
  <input>
5500
  <port id="0" precision="FP16">
 
5513
  </port>
5514
  </output>
5515
  </layer>
5516
+ <layer id="370" name="Constant_31302/fq_weights_1/convert" type="Convert" version="opset1">
5517
  <data destination_type="f32" />
5518
  <input>
5519
  <port id="0" precision="FP16">
 
5528
  </port>
5529
  </output>
5530
  </layer>
5531
+ <layer id="371" name="Multiply_31288" type="MatMul" version="opset1">
5532
  <data transpose_a="false" transpose_b="true" />
5533
  <input>
5534
  <port id="0" precision="FP32">
 
5549
  </port>
5550
  </output>
5551
  </layer>
5552
+ <layer id="372" name="Constant_31411" type="Const" version="opset1">
5553
  <data element_type="f32" shape="1, 1, 1280" offset="85117182" size="5120" />
5554
  <output>
5555
  <port id="0" precision="FP32">
 
5581
  </port>
5582
  </output>
5583
  </layer>
5584
+ <layer id="374" name="Constant_31719" type="Const" version="opset1">
5585
  <data element_type="i64" shape="4" offset="67985482" size="32" />
5586
  <output>
5587
  <port id="0" precision="I64">
 
5649
  </port>
5650
  </output>
5651
  </layer>
5652
+ <layer id="379" name="Convert_43052" type="Convert" version="opset1">
5653
  <data destination_type="f16" />
5654
  <input>
5655
  <port id="0" precision="U4">
 
5676
  </port>
5677
  </output>
5678
  </layer>
5679
+ <layer id="381" name="Convert_43054" type="Convert" version="opset1">
5680
  <data destination_type="f16" />
5681
  <input>
5682
  <port id="0" precision="U4">
 
5747
  </port>
5748
  </output>
5749
  </layer>
5750
+ <layer id="385" name="Constant_43059" type="Const" version="opset1">
5751
  <data element_type="i64" shape="2" offset="67980346" size="16" />
5752
  <output>
5753
  <port id="0" precision="I64">
 
5755
  </port>
5756
  </output>
5757
  </layer>
5758
+ <layer id="386" name="Reshape_43060" type="Reshape" version="opset1">
5759
  <data special_zero="false" />
5760
  <input>
5761
  <port id="0" precision="FP16">
 
5914
  </port>
5915
  </output>
5916
  </layer>
5917
+ <layer id="395" name="Convert_43041" type="Convert" version="opset1">
5918
  <data destination_type="f16" />
5919
  <input>
5920
  <port id="0" precision="U4">
 
5941
  </port>
5942
  </output>
5943
  </layer>
5944
+ <layer id="397" name="Convert_43043" type="Convert" version="opset1">
5945
  <data destination_type="f16" />
5946
  <input>
5947
  <port id="0" precision="U4">
 
6012
  </port>
6013
  </output>
6014
  </layer>
6015
+ <layer id="401" name="Constant_43048" type="Const" version="opset1">
6016
  <data element_type="i64" shape="2" offset="67980346" size="16" />
6017
  <output>
6018
  <port id="0" precision="I64">
 
6020
  </port>
6021
  </output>
6022
  </layer>
6023
+ <layer id="402" name="Reshape_43049" type="Reshape" version="opset1">
6024
  <data special_zero="false" />
6025
  <input>
6026
  <port id="0" precision="FP16">
 
6075
  </port>
6076
  </output>
6077
  </layer>
6078
+ <layer id="405" name="Constant_31412" type="Const" version="opset1">
6079
  <data element_type="f32" shape="1, 1, 1280" offset="86824702" size="5120" />
6080
  <output>
6081
  <port id="0" precision="FP32">
 
6211
  </port>
6212
  </output>
6213
  </layer>
6214
+ <layer id="413" name="Constant_31720" type="Const" version="opset1">
6215
  <data element_type="i64" shape="3" offset="73123526" size="24" />
6216
  <output>
6217
  <port id="0" precision="I64">
 
6250
  </port>
6251
  </output>
6252
  </layer>
6253
+ <layer id="416" name="Convert_43206" type="Convert" version="opset1">
6254
  <data destination_type="f16" />
6255
  <input>
6256
  <port id="0" precision="U4">
 
6277
  </port>
6278
  </output>
6279
  </layer>
6280
+ <layer id="418" name="Convert_43208" type="Convert" version="opset1">
6281
  <data destination_type="f16" />
6282
  <input>
6283
  <port id="0" precision="U4">
 
6348
  </port>
6349
  </output>
6350
  </layer>
6351
+ <layer id="422" name="Constant_43213" type="Const" version="opset1">
6352
  <data element_type="i64" shape="2" offset="67980346" size="16" />
6353
  <output>
6354
  <port id="0" precision="I64">
 
6356
  </port>
6357
  </output>
6358
  </layer>
6359
+ <layer id="423" name="Reshape_43214" type="Reshape" version="opset1">
6360
  <data special_zero="false" />
6361
  <input>
6362
  <port id="0" precision="FP16">
 
6411
  </port>
6412
  </output>
6413
  </layer>
6414
+ <layer id="426" name="Constant_31413" type="Const" version="opset1">
6415
  <data element_type="f32" shape="1, 1, 1280" offset="87681022" size="5120" />
6416
  <output>
6417
  <port id="0" precision="FP32">
 
6493
  </port>
6494
  </output>
6495
  </layer>
6496
+ <layer id="431" name="Constant_31414" type="Const" version="opset1">
6497
  <data element_type="f32" shape="1, 1, 1280" offset="87686142" size="5120" />
6498
  <output>
6499
  <port id="0" precision="FP32">
 
6525
  </port>
6526
  </output>
6527
  </layer>
6528
+ <layer id="433" name="Constant_31415" type="Const" version="opset1">
6529
  <data element_type="f32" shape="1, 1, 1280" offset="87691262" size="5120" />
6530
  <output>
6531
  <port id="0" precision="FP32">
 
6567
  </port>
6568
  </output>
6569
  </layer>
6570
+ <layer id="436" name="Convert_43019" type="Convert" version="opset1">
6571
  <data destination_type="f16" />
6572
  <input>
6573
  <port id="0" precision="U4">
 
6594
  </port>
6595
  </output>
6596
  </layer>
6597
+ <layer id="438" name="Convert_43021" type="Convert" version="opset1">
6598
  <data destination_type="f16" />
6599
  <input>
6600
  <port id="0" precision="U4">
 
6665
  </port>
6666
  </output>
6667
  </layer>
6668
+ <layer id="442" name="Constant_43026" type="Const" version="opset1">
6669
  <data element_type="i64" shape="2" offset="77394910" size="16" />
6670
  <output>
6671
  <port id="0" precision="I64">
 
6673
  </port>
6674
  </output>
6675
  </layer>
6676
+ <layer id="443" name="Reshape_43027" type="Reshape" version="opset1">
6677
  <data special_zero="false" />
6678
  <input>
6679
  <port id="0" precision="FP16">
 
6728
  </port>
6729
  </output>
6730
  </layer>
6731
+ <layer id="446" name="Constant_31416" type="Const" version="opset1">
6732
  <data element_type="f32" shape="1, 1, 5120" offset="91101182" size="20480" />
6733
  <output>
6734
  <port id="0" precision="FP32">
 
6787
  </port>
6788
  </output>
6789
  </layer>
6790
+ <layer id="450" name="Convert_43030" type="Convert" version="opset1">
6791
  <data destination_type="f16" />
6792
  <input>
6793
  <port id="0" precision="U4">
 
6814
  </port>
6815
  </output>
6816
  </layer>
6817
+ <layer id="452" name="Convert_43032" type="Convert" version="opset1">
6818
  <data destination_type="f16" />
6819
  <input>
6820
  <port id="0" precision="U4">
 
6885
  </port>
6886
  </output>
6887
  </layer>
6888
+ <layer id="456" name="Constant_43037" type="Const" version="opset1">
6889
  <data element_type="i64" shape="2" offset="80820206" size="16" />
6890
  <output>
6891
  <port id="0" precision="I64">
 
6893
  </port>
6894
  </output>
6895
  </layer>
6896
+ <layer id="457" name="Reshape_43038" type="Reshape" version="opset1">
6897
  <data special_zero="false" />
6898
  <input>
6899
  <port id="0" precision="FP16">
 
6948
  </port>
6949
  </output>
6950
  </layer>
6951
+ <layer id="460" name="Constant_31417" type="Const" version="opset1">
6952
  <data element_type="f32" shape="1, 1, 1280" offset="94526462" size="5120" />
6953
  <output>
6954
  <port id="0" precision="FP32">
 
7030
  </port>
7031
  </output>
7032
  </layer>
7033
+ <layer id="465" name="Constant_31418" type="Const" version="opset1">
7034
  <data element_type="f32" shape="1, 1, 1280" offset="94531582" size="5120" />
7035
  <output>
7036
  <port id="0" precision="FP32">
 
7062
  </port>
7063
  </output>
7064
  </layer>
7065
+ <layer id="467" name="Constant_31419" type="Const" version="opset1">
7066
  <data element_type="f32" shape="1, 1, 1280" offset="94536702" size="5120" />
7067
  <output>
7068
  <port id="0" precision="FP32">
 
7115
  </port>
7116
  </output>
7117
  </layer>
7118
+ <layer id="470" name="Result_26206" type="Result" version="opset1">
7119
  <input>
7120
  <port id="0" precision="FP32">
7121
  <dim>-1</dim>
 
7632
  <edge from-layer="469" from-port="2" to-layer="470" to-port="0" />
7633
  </edges>
7634
  <rt_info>
7635
+ <Runtime_version value="2024.5.0-17285-ea5c1dcfdf9-releases/2024/5" />
7636
  <conversion_parameters>
7637
  <framework value="pytorch" />
7638
  <is_python_object value="True" />
 
7640
  <nncf>
7641
  <friendly_names_were_updated value="True" />
7642
  <weight_compression>
7643
+ <advanced_parameters value="{'statistics_path': None, 'awq_params': {'subset_size': 32, 'percent_to_apply': 0.002, 'alpha_min': 0.0, 'alpha_max': 1.0, 'steps': 100}, 'scale_estimation_params': {'subset_size': 64, 'initial_steps': 5, 'scale_steps': 5, 'weight_penalty': -1.0}, 'gptq_params': {'damp_percent': 0.1, 'block_size': 128, 'subset_size': 128}, 'lora_correction_params': {'adapter_rank': 8, 'num_iterations': 3, 'apply_regularization': True, 'subset_size': 128, 'use_int8_adapters': True}}" />
7644
  <all_layers value="False" />
7645
  <awq value="False" />
7646
+ <backup_mode value="int8_asym" />
7647
  <gptq value="False" />
7648
  <group_size value="128" />
7649
  <ignored_scope value="[]" />
 
7655
  </weight_compression>
7656
  </nncf>
7657
  <optimum>
7658
+ <optimum_intel_version value="1.21.0.dev0+d357376" />
7659
+ <optimum_version value="1.23.3" />
7660
+ <pytorch_version value="2.5.1" />
7661
+ <transformers_version value="4.46.3" />
7662
  </optimum>
7663
  </rt_info>
7664
  </net>
openvino_detokenizer.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b55ac04e90f19fede391281a9e5a90169fc646e1e302fe1208c782282b51ceaa
3
  size 528306
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1c95aa7a4b53ecb3768d2a6608aa9e419059aef1c75b9b99e6b1c56328281c18
3
  size 528306
openvino_detokenizer.xml CHANGED
@@ -1,16 +1,16 @@
1
  <?xml version="1.0"?>
2
  <net name="detokenizer" version="11">
3
  <layers>
4
- <layer id="0" name="Parameter_32420" type="Parameter" version="opset1">
5
  <data shape="?,?" element_type="i64" />
6
  <output>
7
- <port id="0" precision="I64" names="Parameter_32420">
8
  <dim>-1</dim>
9
  <dim>-1</dim>
10
  </port>
11
  </output>
12
  </layer>
13
- <layer id="1" name="Convert_32436" type="Convert" version="opset1">
14
  <data destination_type="i32" />
15
  <input>
16
  <port id="0" precision="I64">
@@ -25,7 +25,7 @@
25
  </port>
26
  </output>
27
  </layer>
28
- <layer id="2" name="Constant_32381" type="Const" version="opset1">
29
  <data element_type="u8" shape="528257" offset="0" size="528257" />
30
  <output>
31
  <port id="0" precision="U8">
@@ -33,7 +33,7 @@
33
  </port>
34
  </output>
35
  </layer>
36
- <layer id="3" name="StringTensorUnpack_32382" type="StringTensorUnpack" version="extension">
37
  <data mode="begins_ends" />
38
  <input>
39
  <port id="0" precision="U8">
@@ -52,7 +52,7 @@
52
  </port>
53
  </output>
54
  </layer>
55
- <layer id="4" name="VocabDecoder_32421" type="VocabDecoder" version="extension">
56
  <data skip_tokens="50257, 50258, 50259, 50260, 50261, 50262, 50263, 50264, 50265, 50266, 50267, 50268, 50269, 50270, 50271, 50272, 50273, 50274, 50275, 50276, 50277, 50278, 50279, 50280, 50281, 50282, 50283, 50284, 50285, 50286, 50287, 50288, 50289, 50290, 50291, 50292, 50293, 50294, 50295, 50296, 50297, 50298, 50299, 50300, 50301, 50302, 50303, 50304, 50305, 50306, 50307, 50308, 50309, 50310, 50311, 50312, 50313, 50314, 50315, 50316, 50317, 50318, 50319, 50320, 50321, 50322, 50323, 50324, 50325, 50326, 50327, 50328, 50329, 50330, 50331, 50332, 50333, 50334, 50335, 50336, 50337, 50338, 50339, 50340, 50341, 50342, 50343, 50344, 50345, 50346, 50347, 50348, 50349, 50350, 50351, 50352, 50353, 50354, 50355, 50356, 50357, 50358, 50359, 50360, 50361, 50362, 50363, 50364" />
57
  <input>
58
  <port id="0" precision="I32">
@@ -87,7 +87,7 @@
87
  </port>
88
  </output>
89
  </layer>
90
- <layer id="5" name="FuzeRagged_32422" type="FuzeRagged" version="extension">
91
  <input>
92
  <port id="0" precision="I32">
93
  <dim>-1</dim>
@@ -111,7 +111,7 @@
111
  </port>
112
  </output>
113
  </layer>
114
- <layer id="6" name="Constant_32424" type="Const" version="opset1">
115
  <data element_type="u8" shape="47" offset="528257" size="47" />
116
  <output>
117
  <port id="0" precision="U8">
@@ -119,7 +119,7 @@
119
  </port>
120
  </output>
121
  </layer>
122
- <layer id="7" name="Constant_32426" type="Const" version="opset1">
123
  <data element_type="u8" shape="2" offset="528304" size="2" />
124
  <output>
125
  <port id="0" precision="U8">
@@ -127,7 +127,7 @@
127
  </port>
128
  </output>
129
  </layer>
130
- <layer id="8" name="RegexNormalization_32427" type="RegexNormalization" version="extension">
131
  <data global_replace="true" />
132
  <input>
133
  <port id="0" precision="I32">
@@ -158,7 +158,7 @@
158
  </port>
159
  </output>
160
  </layer>
161
- <layer id="9" name="StringTensorPack_32428" type="StringTensorPack" version="extension">
162
  <data mode="begins_ends" />
163
  <input>
164
  <port id="0" precision="I32">
@@ -177,7 +177,7 @@
177
  </port>
178
  </output>
179
  </layer>
180
- <layer id="10" name="Result_32429" type="Result" version="opset1">
181
  <input>
182
  <port id="0" precision="STRING">
183
  <dim>-1</dim>
@@ -207,9 +207,29 @@
207
  <edge from-layer="9" from-port="3" to-layer="10" to-port="0" />
208
  </edges>
209
  <rt_info>
 
 
 
210
  <bos_token_id value="50257" />
 
 
211
  <eos_token_id value="50257" />
 
 
 
 
212
  <original_tokenizer_class value="&lt;class 'transformers.models.whisper.tokenization_whisper_fast.WhisperTokenizerFast'>" />
213
  <pad_token_id value="50257" />
 
 
 
 
 
 
 
 
 
 
 
214
  </rt_info>
215
  </net>
 
1
  <?xml version="1.0"?>
2
  <net name="detokenizer" version="11">
3
  <layers>
4
+ <layer id="0" name="Parameter_32760" type="Parameter" version="opset1">
5
  <data shape="?,?" element_type="i64" />
6
  <output>
7
+ <port id="0" precision="I64" names="Parameter_32760">
8
  <dim>-1</dim>
9
  <dim>-1</dim>
10
  </port>
11
  </output>
12
  </layer>
13
+ <layer id="1" name="Convert_32776" type="Convert" version="opset1">
14
  <data destination_type="i32" />
15
  <input>
16
  <port id="0" precision="I64">
 
25
  </port>
26
  </output>
27
  </layer>
28
+ <layer id="2" name="Constant_32721" type="Const" version="opset1">
29
  <data element_type="u8" shape="528257" offset="0" size="528257" />
30
  <output>
31
  <port id="0" precision="U8">
 
33
  </port>
34
  </output>
35
  </layer>
36
+ <layer id="3" name="StringTensorUnpack_32722" type="StringTensorUnpack" version="extension">
37
  <data mode="begins_ends" />
38
  <input>
39
  <port id="0" precision="U8">
 
52
  </port>
53
  </output>
54
  </layer>
55
+ <layer id="4" name="VocabDecoder_32761" type="VocabDecoder" version="extension">
56
  <data skip_tokens="50257, 50258, 50259, 50260, 50261, 50262, 50263, 50264, 50265, 50266, 50267, 50268, 50269, 50270, 50271, 50272, 50273, 50274, 50275, 50276, 50277, 50278, 50279, 50280, 50281, 50282, 50283, 50284, 50285, 50286, 50287, 50288, 50289, 50290, 50291, 50292, 50293, 50294, 50295, 50296, 50297, 50298, 50299, 50300, 50301, 50302, 50303, 50304, 50305, 50306, 50307, 50308, 50309, 50310, 50311, 50312, 50313, 50314, 50315, 50316, 50317, 50318, 50319, 50320, 50321, 50322, 50323, 50324, 50325, 50326, 50327, 50328, 50329, 50330, 50331, 50332, 50333, 50334, 50335, 50336, 50337, 50338, 50339, 50340, 50341, 50342, 50343, 50344, 50345, 50346, 50347, 50348, 50349, 50350, 50351, 50352, 50353, 50354, 50355, 50356, 50357, 50358, 50359, 50360, 50361, 50362, 50363, 50364" />
57
  <input>
58
  <port id="0" precision="I32">
 
87
  </port>
88
  </output>
89
  </layer>
90
+ <layer id="5" name="FuzeRagged_32762" type="FuzeRagged" version="extension">
91
  <input>
92
  <port id="0" precision="I32">
93
  <dim>-1</dim>
 
111
  </port>
112
  </output>
113
  </layer>
114
+ <layer id="6" name="Constant_32764" type="Const" version="opset1">
115
  <data element_type="u8" shape="47" offset="528257" size="47" />
116
  <output>
117
  <port id="0" precision="U8">
 
119
  </port>
120
  </output>
121
  </layer>
122
+ <layer id="7" name="Constant_32766" type="Const" version="opset1">
123
  <data element_type="u8" shape="2" offset="528304" size="2" />
124
  <output>
125
  <port id="0" precision="U8">
 
127
  </port>
128
  </output>
129
  </layer>
130
+ <layer id="8" name="RegexNormalization_32767" type="RegexNormalization" version="extension">
131
  <data global_replace="true" />
132
  <input>
133
  <port id="0" precision="I32">
 
158
  </port>
159
  </output>
160
  </layer>
161
+ <layer id="9" name="StringTensorPack_32768" type="StringTensorPack" version="extension">
162
  <data mode="begins_ends" />
163
  <input>
164
  <port id="0" precision="I32">
 
177
  </port>
178
  </output>
179
  </layer>
180
+ <layer id="10" name="Result_32769" type="Result" version="opset1">
181
  <input>
182
  <port id="0" precision="STRING">
183
  <dim>-1</dim>
 
207
  <edge from-layer="9" from-port="3" to-layer="10" to-port="0" />
208
  </edges>
209
  <rt_info>
210
+ <add_attention_mask value="True" />
211
+ <add_prefix_space />
212
+ <add_special_tokens value="True" />
213
  <bos_token_id value="50257" />
214
+ <clean_up_tokenization_spaces />
215
+ <detokenizer_input_type value="i64" />
216
  <eos_token_id value="50257" />
217
+ <handle_special_tokens_with_re />
218
+ <number_of_inputs value="1" />
219
+ <openvino_tokenizers_version value="2024.5.0.0" />
220
+ <openvino_version value="2024.5.0" />
221
  <original_tokenizer_class value="&lt;class 'transformers.models.whisper.tokenization_whisper_fast.WhisperTokenizerFast'>" />
222
  <pad_token_id value="50257" />
223
+ <sentencepiece_version value="0.2.0" />
224
+ <skip_special_tokens value="True" />
225
+ <streaming_detokenizer value="False" />
226
+ <tiktoken_version value="0.8.0" />
227
+ <tokenizer_output_type value="i64" />
228
+ <tokenizers_version value="0.20.3" />
229
+ <transformers_version value="4.46.3" />
230
+ <use_max_padding value="False" />
231
+ <use_sentencepiece_backend value="False" />
232
+ <utf8_replace_mode />
233
+ <with_detokenizer value="True" />
234
  </rt_info>
235
  </net>
openvino_encoder_model.xml CHANGED
The diff for this file is too large to render. See raw diff
 
openvino_tokenizer.xml CHANGED
@@ -1,27 +1,27 @@
1
  <?xml version="1.0"?>
2
  <net name="tokenizer" version="11">
3
  <layers>
4
- <layer id="0" name="Parameter_32300" type="Parameter" version="opset1">
5
  <data shape="?" element_type="string" />
6
  <output>
7
- <port id="0" precision="STRING" names="Parameter_32300">
8
  <dim>-1</dim>
9
  </port>
10
  </output>
11
  </layer>
12
- <layer id="1" name="Constant_32398" type="Const" version="opset1">
13
  <data element_type="i32" shape="" offset="0" size="4" />
14
  <output>
15
  <port id="0" precision="I32" />
16
  </output>
17
  </layer>
18
- <layer id="2" name="Constant_32399" type="Const" version="opset1">
19
  <data element_type="i32" shape="" offset="4" size="4" />
20
  <output>
21
  <port id="0" precision="I32" />
22
  </output>
23
  </layer>
24
- <layer id="3" name="Constant_32400" type="Const" version="opset1">
25
  <data element_type="i32" shape="1" offset="8" size="4" />
26
  <output>
27
  <port id="0" precision="I32">
@@ -29,19 +29,19 @@
29
  </port>
30
  </output>
31
  </layer>
32
- <layer id="4" name="Constant_32401" type="Const" version="opset1">
33
  <data element_type="i32" shape="" offset="0" size="4" />
34
  <output>
35
  <port id="0" precision="I32" />
36
  </output>
37
  </layer>
38
- <layer id="5" name="Constant_32402" type="Const" version="opset1">
39
  <data element_type="i32" shape="" offset="4" size="4" />
40
  <output>
41
  <port id="0" precision="I32" />
42
  </output>
43
  </layer>
44
- <layer id="6" name="Constant_32403" type="Const" version="opset1">
45
  <data element_type="i32" shape="1" offset="12" size="4" />
46
  <output>
47
  <port id="0" precision="I32">
@@ -49,13 +49,13 @@
49
  </port>
50
  </output>
51
  </layer>
52
- <layer id="7" name="Constant_32306" type="Const" version="opset1">
53
  <data element_type="i64" shape="" offset="16" size="8" />
54
  <output>
55
  <port id="0" precision="I64" />
56
  </output>
57
  </layer>
58
- <layer id="8" name="StringTensorUnpack_32301" type="StringTensorUnpack" version="extension">
59
  <data mode="begins_ends" />
60
  <input>
61
  <port id="0" precision="STRING">
@@ -74,7 +74,7 @@
74
  </port>
75
  </output>
76
  </layer>
77
- <layer id="9" name="ShapeOf_32302" type="ShapeOf" version="opset3">
78
  <data output_type="i64" />
79
  <input>
80
  <port id="0" precision="I32">
@@ -87,19 +87,19 @@
87
  </port>
88
  </output>
89
  </layer>
90
- <layer id="10" name="Constant_32303" type="Const" version="opset1">
91
  <data element_type="i64" shape="" offset="16" size="8" />
92
  <output>
93
  <port id="0" precision="I64" />
94
  </output>
95
  </layer>
96
- <layer id="11" name="Constant_32304" type="Const" version="opset1">
97
  <data element_type="i64" shape="" offset="16" size="8" />
98
  <output>
99
  <port id="0" precision="I64" />
100
  </output>
101
  </layer>
102
- <layer id="12" name="Gather_32305" type="Gather" version="opset8">
103
  <data batch_dims="0" />
104
  <input>
105
  <port id="0" precision="I64">
@@ -112,13 +112,13 @@
112
  <port id="3" precision="I64" />
113
  </output>
114
  </layer>
115
- <layer id="13" name="Constant_32307" type="Const" version="opset1">
116
  <data element_type="i64" shape="" offset="24" size="8" />
117
  <output>
118
  <port id="0" precision="I64" />
119
  </output>
120
  </layer>
121
- <layer id="14" name="Range_32308" type="Range" version="opset4">
122
  <data output_type="i32" />
123
  <input>
124
  <port id="0" precision="I64" />
@@ -131,19 +131,19 @@
131
  </port>
132
  </output>
133
  </layer>
134
- <layer id="15" name="Constant_32309" type="Const" version="opset1">
135
  <data element_type="i64" shape="" offset="24" size="8" />
136
  <output>
137
  <port id="0" precision="I64" />
138
  </output>
139
  </layer>
140
- <layer id="16" name="Constant_32310" type="Const" version="opset1">
141
  <data element_type="i64" shape="" offset="24" size="8" />
142
  <output>
143
  <port id="0" precision="I64" />
144
  </output>
145
  </layer>
146
- <layer id="17" name="Add_32311" type="Add" version="opset1">
147
  <data auto_broadcast="numpy" />
148
  <input>
149
  <port id="0" precision="I64" />
@@ -153,13 +153,13 @@
153
  <port id="2" precision="I64" />
154
  </output>
155
  </layer>
156
- <layer id="18" name="Constant_32312" type="Const" version="opset1">
157
  <data element_type="i64" shape="" offset="24" size="8" />
158
  <output>
159
  <port id="0" precision="I64" />
160
  </output>
161
  </layer>
162
- <layer id="19" name="Range_32313" type="Range" version="opset4">
163
  <data output_type="i32" />
164
  <input>
165
  <port id="0" precision="I64" />
@@ -172,7 +172,7 @@
172
  </port>
173
  </output>
174
  </layer>
175
- <layer id="20" name="Constant_32375" type="Const" version="opset1">
176
  <data element_type="u8" shape="26491" offset="32" size="26491" />
177
  <output>
178
  <port id="0" precision="U8">
@@ -180,7 +180,7 @@
180
  </port>
181
  </output>
182
  </layer>
183
- <layer id="21" name="SpecialTokensSplit_32376" type="SpecialTokensSplit" version="extension">
184
  <input>
185
  <port id="0" precision="I32">
186
  <dim>-1</dim>
@@ -222,7 +222,7 @@
222
  </port>
223
  </output>
224
  </layer>
225
- <layer id="22" name="Constant_32378" type="Const" version="opset1">
226
  <data element_type="u8" shape="64" offset="26523" size="64" />
227
  <output>
228
  <port id="0" precision="U8">
@@ -230,7 +230,7 @@
230
  </port>
231
  </output>
232
  </layer>
233
- <layer id="23" name="RegexSplit_32379" type="RegexSplit" version="extension">
234
  <data behaviour="isolate" invert="false" max_splits="-1" />
235
  <input>
236
  <port id="0" precision="I32">
@@ -276,7 +276,7 @@
276
  </port>
277
  </output>
278
  </layer>
279
- <layer id="24" name="Constant_32381" type="Const" version="opset1">
280
  <data element_type="u8" shape="528257" offset="26587" size="528257" />
281
  <output>
282
  <port id="0" precision="U8">
@@ -284,7 +284,7 @@
284
  </port>
285
  </output>
286
  </layer>
287
- <layer id="25" name="StringTensorUnpack_32382" type="StringTensorUnpack" version="extension">
288
  <data mode="begins_ends" />
289
  <input>
290
  <port id="0" precision="U8">
@@ -303,7 +303,7 @@
303
  </port>
304
  </output>
305
  </layer>
306
- <layer id="26" name="Constant_32387" type="Const" version="opset1">
307
  <data element_type="u8" shape="369958" offset="554844" size="369958" />
308
  <output>
309
  <port id="0" precision="U8">
@@ -311,7 +311,7 @@
311
  </port>
312
  </output>
313
  </layer>
314
- <layer id="27" name="StringTensorUnpack_32388" type="StringTensorUnpack" version="extension">
315
  <data mode="begins_ends" />
316
  <input>
317
  <port id="0" precision="U8">
@@ -330,7 +330,7 @@
330
  </port>
331
  </output>
332
  </layer>
333
- <layer id="28" name="Constant_32390" type="Const" version="opset1">
334
  <data element_type="u8" shape="336859" offset="924802" size="336859" />
335
  <output>
336
  <port id="0" precision="U8">
@@ -338,7 +338,7 @@
338
  </port>
339
  </output>
340
  </layer>
341
- <layer id="29" name="StringTensorUnpack_32391" type="StringTensorUnpack" version="extension">
342
  <data mode="begins_ends" />
343
  <input>
344
  <port id="0" precision="U8">
@@ -357,7 +357,7 @@
357
  </port>
358
  </output>
359
  </layer>
360
- <layer id="30" name="Constant_32384" type="Const" version="opset1">
361
  <data element_type="u8" shape="20172" offset="1261661" size="20172" />
362
  <output>
363
  <port id="0" precision="U8">
@@ -365,7 +365,7 @@
365
  </port>
366
  </output>
367
  </layer>
368
- <layer id="31" name="StringTensorUnpack_32385" type="StringTensorUnpack" version="extension">
369
  <data mode="begins_ends" />
370
  <input>
371
  <port id="0" precision="U8">
@@ -384,7 +384,7 @@
384
  </port>
385
  </output>
386
  </layer>
387
- <layer id="32" name="Constant_32392" type="Const" version="opset1">
388
  <data element_type="i32" shape="1609" offset="1281833" size="6436" />
389
  <output>
390
  <port id="0" precision="I32">
@@ -392,7 +392,7 @@
392
  </port>
393
  </output>
394
  </layer>
395
- <layer id="33" name="BPETokenizer_32393" type="BPETokenizer" version="extension">
396
  <data unk_token="" fuse_unk="false" suffix_indicator="" end_suffix="" byte_fallback="false" cache_capacity="20000" />
397
  <input>
398
  <port id="0" precision="I32">
@@ -462,7 +462,7 @@
462
  </port>
463
  </output>
464
  </layer>
465
- <layer id="34" name="Subtract_32394" type="Subtract" version="opset1">
466
  <data auto_broadcast="numpy" />
467
  <input>
468
  <port id="0" precision="I32">
@@ -478,13 +478,13 @@
478
  </port>
479
  </output>
480
  </layer>
481
- <layer id="35" name="Constant_32395" type="Const" version="opset1">
482
  <data element_type="i32" shape="" offset="1288269" size="4" />
483
  <output>
484
  <port id="0" precision="I32" />
485
  </output>
486
  </layer>
487
- <layer id="36" name="Minimum_32396" type="Minimum" version="opset1">
488
  <data auto_broadcast="numpy" />
489
  <input>
490
  <port id="0" precision="I32">
@@ -498,7 +498,7 @@
498
  </port>
499
  </output>
500
  </layer>
501
- <layer id="37" name="Add_32397" type="Add" version="opset1">
502
  <data auto_broadcast="numpy" />
503
  <input>
504
  <port id="0" precision="I32">
@@ -514,19 +514,19 @@
514
  </port>
515
  </output>
516
  </layer>
517
- <layer id="38" name="Constant_32404" type="Const" version="opset1">
518
  <data element_type="i32" shape="" offset="0" size="4" />
519
  <output>
520
  <port id="0" precision="I32" />
521
  </output>
522
  </layer>
523
- <layer id="39" name="Constant_32405" type="Const" version="opset1">
524
  <data element_type="i32" shape="" offset="4" size="4" />
525
  <output>
526
  <port id="0" precision="I32" />
527
  </output>
528
  </layer>
529
- <layer id="40" name="Constant_32406" type="Const" version="opset1">
530
  <data element_type="i32" shape="1" offset="1288273" size="4" />
531
  <output>
532
  <port id="0" precision="I32">
@@ -534,7 +534,7 @@
534
  </port>
535
  </output>
536
  </layer>
537
- <layer id="41" name="Constant_32407" type="Const" version="opset1">
538
  <data element_type="i32" shape="4" offset="1288277" size="16" />
539
  <output>
540
  <port id="0" precision="I32">
@@ -542,7 +542,7 @@
542
  </port>
543
  </output>
544
  </layer>
545
- <layer id="42" name="CombineSegments_32408" type="CombineSegments" version="extension">
546
  <input>
547
  <port id="0" precision="I32" />
548
  <port id="1" precision="I32" />
@@ -593,7 +593,7 @@
593
  </port>
594
  </output>
595
  </layer>
596
- <layer id="43" name="Subtract_32409" type="Subtract" version="opset1">
597
  <data auto_broadcast="numpy" />
598
  <input>
599
  <port id="0" precision="I32">
@@ -609,13 +609,13 @@
609
  </port>
610
  </output>
611
  </layer>
612
- <layer id="44" name="Constant_32410" type="Const" version="opset1">
613
  <data element_type="i32" shape="" offset="0" size="4" />
614
  <output>
615
  <port id="0" precision="I32" />
616
  </output>
617
  </layer>
618
- <layer id="45" name="ReduceMax_32411" type="ReduceMax" version="opset1">
619
  <data keep_dims="false" />
620
  <input>
621
  <port id="0" precision="I32">
@@ -627,13 +627,13 @@
627
  <port id="2" precision="I32" />
628
  </output>
629
  </layer>
630
- <layer id="46" name="Constant_32412" type="Const" version="opset1">
631
  <data element_type="i32" shape="" offset="1288273" size="4" />
632
  <output>
633
  <port id="0" precision="I32" />
634
  </output>
635
  </layer>
636
- <layer id="47" name="RaggedToDense_32413" type="RaggedToDense" version="extension">
637
  <data pad_right="true" />
638
  <input>
639
  <port id="0" precision="I32">
@@ -659,7 +659,7 @@
659
  </port>
660
  </output>
661
  </layer>
662
- <layer id="48" name="Convert_32414" type="Convert" version="opset1">
663
  <data destination_type="i32" />
664
  <input>
665
  <port id="0" precision="BOOL">
@@ -674,7 +674,7 @@
674
  </port>
675
  </output>
676
  </layer>
677
- <layer id="49" name="Convert_32414" type="Convert" version="opset1">
678
  <data destination_type="i64" />
679
  <input>
680
  <port id="0" precision="I32">
@@ -689,7 +689,7 @@
689
  </port>
690
  </output>
691
  </layer>
692
- <layer id="51" name="RaggedToDense_32413.0" type="Convert" version="opset1">
693
  <data destination_type="i64" />
694
  <input>
695
  <port id="0" precision="I32">
@@ -704,7 +704,7 @@
704
  </port>
705
  </output>
706
  </layer>
707
- <layer id="52" name="Result_32417" type="Result" version="opset1">
708
  <input>
709
  <port id="0" precision="I64">
710
  <dim>-1</dim>
@@ -712,7 +712,7 @@
712
  </port>
713
  </input>
714
  </layer>
715
- <layer id="50" name="Result_32419" type="Result" version="opset1">
716
  <input>
717
  <port id="0" precision="I64">
718
  <dim>-1</dim>
@@ -805,9 +805,29 @@
805
  <edge from-layer="51" from-port="1" to-layer="52" to-port="0" />
806
  </edges>
807
  <rt_info>
 
 
 
808
  <bos_token_id value="50257" />
 
 
809
  <eos_token_id value="50257" />
 
 
 
 
810
  <original_tokenizer_class value="&lt;class 'transformers.models.whisper.tokenization_whisper_fast.WhisperTokenizerFast'>" />
811
  <pad_token_id value="50257" />
 
 
 
 
 
 
 
 
 
 
 
812
  </rt_info>
813
  </net>
 
1
  <?xml version="1.0"?>
2
  <net name="tokenizer" version="11">
3
  <layers>
4
+ <layer id="0" name="Parameter_32640" type="Parameter" version="opset1">
5
  <data shape="?" element_type="string" />
6
  <output>
7
+ <port id="0" precision="STRING" names="Parameter_32640">
8
  <dim>-1</dim>
9
  </port>
10
  </output>
11
  </layer>
12
+ <layer id="1" name="Constant_32738" type="Const" version="opset1">
13
  <data element_type="i32" shape="" offset="0" size="4" />
14
  <output>
15
  <port id="0" precision="I32" />
16
  </output>
17
  </layer>
18
+ <layer id="2" name="Constant_32739" type="Const" version="opset1">
19
  <data element_type="i32" shape="" offset="4" size="4" />
20
  <output>
21
  <port id="0" precision="I32" />
22
  </output>
23
  </layer>
24
+ <layer id="3" name="Constant_32740" type="Const" version="opset1">
25
  <data element_type="i32" shape="1" offset="8" size="4" />
26
  <output>
27
  <port id="0" precision="I32">
 
29
  </port>
30
  </output>
31
  </layer>
32
+ <layer id="4" name="Constant_32741" type="Const" version="opset1">
33
  <data element_type="i32" shape="" offset="0" size="4" />
34
  <output>
35
  <port id="0" precision="I32" />
36
  </output>
37
  </layer>
38
+ <layer id="5" name="Constant_32742" type="Const" version="opset1">
39
  <data element_type="i32" shape="" offset="4" size="4" />
40
  <output>
41
  <port id="0" precision="I32" />
42
  </output>
43
  </layer>
44
+ <layer id="6" name="Constant_32743" type="Const" version="opset1">
45
  <data element_type="i32" shape="1" offset="12" size="4" />
46
  <output>
47
  <port id="0" precision="I32">
 
49
  </port>
50
  </output>
51
  </layer>
52
+ <layer id="7" name="Constant_32646" type="Const" version="opset1">
53
  <data element_type="i64" shape="" offset="16" size="8" />
54
  <output>
55
  <port id="0" precision="I64" />
56
  </output>
57
  </layer>
58
+ <layer id="8" name="StringTensorUnpack_32641" type="StringTensorUnpack" version="extension">
59
  <data mode="begins_ends" />
60
  <input>
61
  <port id="0" precision="STRING">
 
74
  </port>
75
  </output>
76
  </layer>
77
+ <layer id="9" name="ShapeOf_32642" type="ShapeOf" version="opset3">
78
  <data output_type="i64" />
79
  <input>
80
  <port id="0" precision="I32">
 
87
  </port>
88
  </output>
89
  </layer>
90
+ <layer id="10" name="Constant_32643" type="Const" version="opset1">
91
  <data element_type="i64" shape="" offset="16" size="8" />
92
  <output>
93
  <port id="0" precision="I64" />
94
  </output>
95
  </layer>
96
+ <layer id="11" name="Constant_32644" type="Const" version="opset1">
97
  <data element_type="i64" shape="" offset="16" size="8" />
98
  <output>
99
  <port id="0" precision="I64" />
100
  </output>
101
  </layer>
102
+ <layer id="12" name="Gather_32645" type="Gather" version="opset8">
103
  <data batch_dims="0" />
104
  <input>
105
  <port id="0" precision="I64">
 
112
  <port id="3" precision="I64" />
113
  </output>
114
  </layer>
115
+ <layer id="13" name="Constant_32647" type="Const" version="opset1">
116
  <data element_type="i64" shape="" offset="24" size="8" />
117
  <output>
118
  <port id="0" precision="I64" />
119
  </output>
120
  </layer>
121
+ <layer id="14" name="Range_32648" type="Range" version="opset4">
122
  <data output_type="i32" />
123
  <input>
124
  <port id="0" precision="I64" />
 
131
  </port>
132
  </output>
133
  </layer>
134
+ <layer id="15" name="Constant_32649" type="Const" version="opset1">
135
  <data element_type="i64" shape="" offset="24" size="8" />
136
  <output>
137
  <port id="0" precision="I64" />
138
  </output>
139
  </layer>
140
+ <layer id="16" name="Constant_32650" type="Const" version="opset1">
141
  <data element_type="i64" shape="" offset="24" size="8" />
142
  <output>
143
  <port id="0" precision="I64" />
144
  </output>
145
  </layer>
146
+ <layer id="17" name="Add_32651" type="Add" version="opset1">
147
  <data auto_broadcast="numpy" />
148
  <input>
149
  <port id="0" precision="I64" />
 
153
  <port id="2" precision="I64" />
154
  </output>
155
  </layer>
156
+ <layer id="18" name="Constant_32652" type="Const" version="opset1">
157
  <data element_type="i64" shape="" offset="24" size="8" />
158
  <output>
159
  <port id="0" precision="I64" />
160
  </output>
161
  </layer>
162
+ <layer id="19" name="Range_32653" type="Range" version="opset4">
163
  <data output_type="i32" />
164
  <input>
165
  <port id="0" precision="I64" />
 
172
  </port>
173
  </output>
174
  </layer>
175
+ <layer id="20" name="Constant_32715" type="Const" version="opset1">
176
  <data element_type="u8" shape="26491" offset="32" size="26491" />
177
  <output>
178
  <port id="0" precision="U8">
 
180
  </port>
181
  </output>
182
  </layer>
183
+ <layer id="21" name="SpecialTokensSplit_32716" type="SpecialTokensSplit" version="extension">
184
  <input>
185
  <port id="0" precision="I32">
186
  <dim>-1</dim>
 
222
  </port>
223
  </output>
224
  </layer>
225
+ <layer id="22" name="Constant_32718" type="Const" version="opset1">
226
  <data element_type="u8" shape="64" offset="26523" size="64" />
227
  <output>
228
  <port id="0" precision="U8">
 
230
  </port>
231
  </output>
232
  </layer>
233
+ <layer id="23" name="RegexSplit_32719" type="RegexSplit" version="extension">
234
  <data behaviour="isolate" invert="false" max_splits="-1" />
235
  <input>
236
  <port id="0" precision="I32">
 
276
  </port>
277
  </output>
278
  </layer>
279
+ <layer id="24" name="Constant_32721" type="Const" version="opset1">
280
  <data element_type="u8" shape="528257" offset="26587" size="528257" />
281
  <output>
282
  <port id="0" precision="U8">
 
284
  </port>
285
  </output>
286
  </layer>
287
+ <layer id="25" name="StringTensorUnpack_32722" type="StringTensorUnpack" version="extension">
288
  <data mode="begins_ends" />
289
  <input>
290
  <port id="0" precision="U8">
 
303
  </port>
304
  </output>
305
  </layer>
306
+ <layer id="26" name="Constant_32727" type="Const" version="opset1">
307
  <data element_type="u8" shape="369958" offset="554844" size="369958" />
308
  <output>
309
  <port id="0" precision="U8">
 
311
  </port>
312
  </output>
313
  </layer>
314
+ <layer id="27" name="StringTensorUnpack_32728" type="StringTensorUnpack" version="extension">
315
  <data mode="begins_ends" />
316
  <input>
317
  <port id="0" precision="U8">
 
330
  </port>
331
  </output>
332
  </layer>
333
+ <layer id="28" name="Constant_32730" type="Const" version="opset1">
334
  <data element_type="u8" shape="336859" offset="924802" size="336859" />
335
  <output>
336
  <port id="0" precision="U8">
 
338
  </port>
339
  </output>
340
  </layer>
341
+ <layer id="29" name="StringTensorUnpack_32731" type="StringTensorUnpack" version="extension">
342
  <data mode="begins_ends" />
343
  <input>
344
  <port id="0" precision="U8">
 
357
  </port>
358
  </output>
359
  </layer>
360
+ <layer id="30" name="Constant_32724" type="Const" version="opset1">
361
  <data element_type="u8" shape="20172" offset="1261661" size="20172" />
362
  <output>
363
  <port id="0" precision="U8">
 
365
  </port>
366
  </output>
367
  </layer>
368
+ <layer id="31" name="StringTensorUnpack_32725" type="StringTensorUnpack" version="extension">
369
  <data mode="begins_ends" />
370
  <input>
371
  <port id="0" precision="U8">
 
384
  </port>
385
  </output>
386
  </layer>
387
+ <layer id="32" name="Constant_32732" type="Const" version="opset1">
388
  <data element_type="i32" shape="1609" offset="1281833" size="6436" />
389
  <output>
390
  <port id="0" precision="I32">
 
392
  </port>
393
  </output>
394
  </layer>
395
+ <layer id="33" name="BPETokenizer_32733" type="BPETokenizer" version="extension">
396
  <data unk_token="" fuse_unk="false" suffix_indicator="" end_suffix="" byte_fallback="false" cache_capacity="20000" />
397
  <input>
398
  <port id="0" precision="I32">
 
462
  </port>
463
  </output>
464
  </layer>
465
+ <layer id="34" name="Subtract_32734" type="Subtract" version="opset1">
466
  <data auto_broadcast="numpy" />
467
  <input>
468
  <port id="0" precision="I32">
 
478
  </port>
479
  </output>
480
  </layer>
481
+ <layer id="35" name="Constant_32735" type="Const" version="opset1">
482
  <data element_type="i32" shape="" offset="1288269" size="4" />
483
  <output>
484
  <port id="0" precision="I32" />
485
  </output>
486
  </layer>
487
+ <layer id="36" name="Minimum_32736" type="Minimum" version="opset1">
488
  <data auto_broadcast="numpy" />
489
  <input>
490
  <port id="0" precision="I32">
 
498
  </port>
499
  </output>
500
  </layer>
501
+ <layer id="37" name="Add_32737" type="Add" version="opset1">
502
  <data auto_broadcast="numpy" />
503
  <input>
504
  <port id="0" precision="I32">
 
514
  </port>
515
  </output>
516
  </layer>
517
+ <layer id="38" name="Constant_32744" type="Const" version="opset1">
518
  <data element_type="i32" shape="" offset="0" size="4" />
519
  <output>
520
  <port id="0" precision="I32" />
521
  </output>
522
  </layer>
523
+ <layer id="39" name="Constant_32745" type="Const" version="opset1">
524
  <data element_type="i32" shape="" offset="4" size="4" />
525
  <output>
526
  <port id="0" precision="I32" />
527
  </output>
528
  </layer>
529
+ <layer id="40" name="Constant_32746" type="Const" version="opset1">
530
  <data element_type="i32" shape="1" offset="1288273" size="4" />
531
  <output>
532
  <port id="0" precision="I32">
 
534
  </port>
535
  </output>
536
  </layer>
537
+ <layer id="41" name="Constant_32747" type="Const" version="opset1">
538
  <data element_type="i32" shape="4" offset="1288277" size="16" />
539
  <output>
540
  <port id="0" precision="I32">
 
542
  </port>
543
  </output>
544
  </layer>
545
+ <layer id="42" name="CombineSegments_32748" type="CombineSegments" version="extension">
546
  <input>
547
  <port id="0" precision="I32" />
548
  <port id="1" precision="I32" />
 
593
  </port>
594
  </output>
595
  </layer>
596
+ <layer id="43" name="Subtract_32749" type="Subtract" version="opset1">
597
  <data auto_broadcast="numpy" />
598
  <input>
599
  <port id="0" precision="I32">
 
609
  </port>
610
  </output>
611
  </layer>
612
+ <layer id="44" name="Constant_32750" type="Const" version="opset1">
613
  <data element_type="i32" shape="" offset="0" size="4" />
614
  <output>
615
  <port id="0" precision="I32" />
616
  </output>
617
  </layer>
618
+ <layer id="45" name="ReduceMax_32751" type="ReduceMax" version="opset1">
619
  <data keep_dims="false" />
620
  <input>
621
  <port id="0" precision="I32">
 
627
  <port id="2" precision="I32" />
628
  </output>
629
  </layer>
630
+ <layer id="46" name="Constant_32752" type="Const" version="opset1">
631
  <data element_type="i32" shape="" offset="1288273" size="4" />
632
  <output>
633
  <port id="0" precision="I32" />
634
  </output>
635
  </layer>
636
+ <layer id="47" name="RaggedToDense_32753" type="RaggedToDense" version="extension">
637
  <data pad_right="true" />
638
  <input>
639
  <port id="0" precision="I32">
 
659
  </port>
660
  </output>
661
  </layer>
662
+ <layer id="48" name="Convert_32754" type="Convert" version="opset1">
663
  <data destination_type="i32" />
664
  <input>
665
  <port id="0" precision="BOOL">
 
674
  </port>
675
  </output>
676
  </layer>
677
+ <layer id="49" name="Convert_32754" type="Convert" version="opset1">
678
  <data destination_type="i64" />
679
  <input>
680
  <port id="0" precision="I32">
 
689
  </port>
690
  </output>
691
  </layer>
692
+ <layer id="51" name="RaggedToDense_32753.0" type="Convert" version="opset1">
693
  <data destination_type="i64" />
694
  <input>
695
  <port id="0" precision="I32">
 
704
  </port>
705
  </output>
706
  </layer>
707
+ <layer id="52" name="Result_32757" type="Result" version="opset1">
708
  <input>
709
  <port id="0" precision="I64">
710
  <dim>-1</dim>
 
712
  </port>
713
  </input>
714
  </layer>
715
+ <layer id="50" name="Result_32759" type="Result" version="opset1">
716
  <input>
717
  <port id="0" precision="I64">
718
  <dim>-1</dim>
 
805
  <edge from-layer="51" from-port="1" to-layer="52" to-port="0" />
806
  </edges>
807
  <rt_info>
808
+ <add_attention_mask value="True" />
809
+ <add_prefix_space />
810
+ <add_special_tokens value="True" />
811
  <bos_token_id value="50257" />
812
+ <clean_up_tokenization_spaces />
813
+ <detokenizer_input_type value="i64" />
814
  <eos_token_id value="50257" />
815
+ <handle_special_tokens_with_re />
816
+ <number_of_inputs value="1" />
817
+ <openvino_tokenizers_version value="2024.5.0.0" />
818
+ <openvino_version value="2024.5.0" />
819
  <original_tokenizer_class value="&lt;class 'transformers.models.whisper.tokenization_whisper_fast.WhisperTokenizerFast'>" />
820
  <pad_token_id value="50257" />
821
+ <sentencepiece_version value="0.2.0" />
822
+ <skip_special_tokens value="True" />
823
+ <streaming_detokenizer value="False" />
824
+ <tiktoken_version value="0.8.0" />
825
+ <tokenizer_output_type value="i64" />
826
+ <tokenizers_version value="0.20.3" />
827
+ <transformers_version value="4.46.3" />
828
+ <use_max_padding value="False" />
829
+ <use_sentencepiece_backend value="False" />
830
+ <utf8_replace_mode />
831
+ <with_detokenizer value="True" />
832
  </rt_info>
833
  </net>