Hugging Face's logo

peculiar-ragdoll
/
Tiel-Coder-35B-A3B-MLX-oQ4e

Image-Text-to-Text
MLX
Safetensors
English
Chinese
qwen3_5_moe
omlx
apple-silicon
oq
qwen35moe
Mixture of Experts
imatrix
unsloth-dynamic
agentic-coding
token-efficient
vision
conversational
4-bit precision
Model card Files
xet
 Community
1
peculiar-ragdoll commited on
Commit
5a5e5ce
·
verified ·
1 Parent(s): 058d0c1

oQ4e re-quantized on oMLX 0.7.3rc1; imatrix calibrated on the same 3000x512-token chunks as the GGUF imatrix

Browse files
Files changed (8) hide show
  1. config.json +788 -308
  2. model-00001-of-00005.safetensors +2 -2
  3. model-00002-of-00005.safetensors +1 -1
  4. model-00003-of-00005.safetensors +1 -1
  5. model-00004-of-00005.safetensors +1 -1
  6. model-00005-of-00005.safetensors +1 -1
  7. model.safetensors.index.json +1 -1
  8. oq_imatrix_report.json +49 -49
config.json CHANGED
@@ -132,23 +132,23 @@
132
  "mode": "affine"
133
  },
134
  "language_model.model.layers.0.linear_attn.in_proj_a": {
135
- "bits": 8,
136
  "group_size": 64,
137
  "mode": "affine"
138
  },
139
  "language_model.model.layers.0.linear_attn.in_proj_b": {
140
- "bits": 8,
141
  "group_size": 64,
142
  "mode": "affine"
143
  },
144
  "language_model.model.layers.0.linear_attn.in_proj_qkv": {
145
- "bits": 5,
146
- "group_size": 128,
147
  "mode": "affine"
148
  },
149
  "language_model.model.layers.0.linear_attn.in_proj_z": {
150
- "bits": 5,
151
- "group_size": 128,
152
  "mode": "affine"
153
  },
154
  "language_model.model.layers.0.linear_attn.out_proj": {
@@ -177,18 +177,23 @@
177
  "mode": "affine"
178
  },
179
  "language_model.model.layers.1.linear_attn.in_proj_a": {
180
- "bits": 5,
181
- "group_size": 128,
182
  "mode": "affine"
183
  },
184
  "language_model.model.layers.1.linear_attn.in_proj_b": {
 
 
 
 
 
185
  "bits": 5,
186
- "group_size": 128,
187
  "mode": "affine"
188
  },
189
  "language_model.model.layers.1.linear_attn.in_proj_z": {
190
- "bits": 5,
191
- "group_size": 128,
192
  "mode": "affine"
193
  },
194
  "language_model.model.layers.1.linear_attn.out_proj": {
@@ -217,18 +222,23 @@
217
  "mode": "affine"
218
  },
219
  "language_model.model.layers.2.linear_attn.in_proj_a": {
220
- "bits": 5,
221
- "group_size": 128,
222
  "mode": "affine"
223
  },
224
  "language_model.model.layers.2.linear_attn.in_proj_b": {
 
 
 
 
 
225
  "bits": 5,
226
- "group_size": 128,
227
  "mode": "affine"
228
  },
229
  "language_model.model.layers.2.linear_attn.in_proj_z": {
230
- "bits": 5,
231
- "group_size": 128,
232
  "mode": "affine"
233
  },
234
  "language_model.model.layers.2.linear_attn.out_proj": {
@@ -276,19 +286,44 @@
276
  "group_size": 64,
277
  "mode": "affine"
278
  },
279
- "language_model.model.layers.4.linear_attn.in_proj_a": {
280
  "bits": 5,
281
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
282
  "mode": "affine"
283
  },
284
  "language_model.model.layers.4.linear_attn.in_proj_b": {
 
 
 
 
 
285
  "bits": 5,
286
- "group_size": 128,
287
  "mode": "affine"
288
  },
289
  "language_model.model.layers.4.linear_attn.in_proj_z": {
290
- "bits": 5,
291
- "group_size": 128,
292
  "mode": "affine"
293
  },
294
  "language_model.model.layers.4.linear_attn.out_proj": {
@@ -317,18 +352,23 @@
317
  "mode": "affine"
318
  },
319
  "language_model.model.layers.5.linear_attn.in_proj_a": {
320
- "bits": 5,
321
- "group_size": 128,
322
  "mode": "affine"
323
  },
324
  "language_model.model.layers.5.linear_attn.in_proj_b": {
 
 
 
 
 
325
  "bits": 5,
326
- "group_size": 128,
327
  "mode": "affine"
328
  },
329
  "language_model.model.layers.5.linear_attn.in_proj_z": {
330
- "bits": 5,
331
- "group_size": 128,
332
  "mode": "affine"
333
  },
334
  "language_model.model.layers.5.linear_attn.out_proj": {
@@ -357,18 +397,23 @@
357
  "mode": "affine"
358
  },
359
  "language_model.model.layers.6.linear_attn.in_proj_a": {
360
- "bits": 5,
361
- "group_size": 128,
362
  "mode": "affine"
363
  },
364
  "language_model.model.layers.6.linear_attn.in_proj_b": {
 
 
 
 
 
365
  "bits": 5,
366
- "group_size": 128,
367
  "mode": "affine"
368
  },
369
  "language_model.model.layers.6.linear_attn.in_proj_z": {
370
- "bits": 5,
371
- "group_size": 128,
372
  "mode": "affine"
373
  },
374
  "language_model.model.layers.6.linear_attn.out_proj": {
@@ -417,13 +462,18 @@
417
  "mode": "affine"
418
  },
419
  "language_model.model.layers.7.self_attn.k_proj": {
420
- "bits": 8,
421
  "group_size": 64,
422
  "mode": "affine"
423
  },
424
- "language_model.model.layers.7.self_attn.q_proj": {
425
  "bits": 5,
426
- "group_size": 128,
 
 
 
 
 
427
  "mode": "affine"
428
  },
429
  "language_model.model.layers.7.self_attn.v_proj": {
@@ -432,18 +482,23 @@
432
  "mode": "affine"
433
  },
434
  "language_model.model.layers.8.linear_attn.in_proj_a": {
435
- "bits": 5,
436
- "group_size": 128,
437
  "mode": "affine"
438
  },
439
  "language_model.model.layers.8.linear_attn.in_proj_b": {
 
 
 
 
 
440
  "bits": 5,
441
- "group_size": 128,
442
  "mode": "affine"
443
  },
444
  "language_model.model.layers.8.linear_attn.in_proj_z": {
445
- "bits": 5,
446
- "group_size": 128,
447
  "mode": "affine"
448
  },
449
  "language_model.model.layers.8.linear_attn.out_proj": {
@@ -472,18 +527,23 @@
472
  "mode": "affine"
473
  },
474
  "language_model.model.layers.9.linear_attn.in_proj_a": {
475
- "bits": 5,
476
- "group_size": 128,
477
  "mode": "affine"
478
  },
479
  "language_model.model.layers.9.linear_attn.in_proj_b": {
 
 
 
 
 
480
  "bits": 5,
481
- "group_size": 128,
482
  "mode": "affine"
483
  },
484
  "language_model.model.layers.9.linear_attn.in_proj_z": {
485
- "bits": 5,
486
- "group_size": 128,
487
  "mode": "affine"
488
  },
489
  "language_model.model.layers.9.linear_attn.out_proj": {
@@ -512,23 +572,23 @@
512
  "mode": "affine"
513
  },
514
  "language_model.model.layers.10.linear_attn.in_proj_a": {
515
- "bits": 8,
516
  "group_size": 64,
517
  "mode": "affine"
518
  },
519
  "language_model.model.layers.10.linear_attn.in_proj_b": {
520
- "bits": 8,
521
  "group_size": 64,
522
  "mode": "affine"
523
  },
524
  "language_model.model.layers.10.linear_attn.in_proj_qkv": {
525
- "bits": 5,
526
- "group_size": 128,
527
  "mode": "affine"
528
  },
529
  "language_model.model.layers.10.linear_attn.in_proj_z": {
530
- "bits": 5,
531
- "group_size": 128,
532
  "mode": "affine"
533
  },
534
  "language_model.model.layers.10.linear_attn.out_proj": {
@@ -576,19 +636,44 @@
576
  "group_size": 64,
577
  "mode": "affine"
578
  },
579
- "language_model.model.layers.12.linear_attn.in_proj_a": {
580
  "bits": 5,
581
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
582
  "mode": "affine"
583
  },
584
  "language_model.model.layers.12.linear_attn.in_proj_b": {
 
 
 
 
 
585
  "bits": 5,
586
- "group_size": 128,
587
  "mode": "affine"
588
  },
589
  "language_model.model.layers.12.linear_attn.in_proj_z": {
590
- "bits": 5,
591
- "group_size": 128,
592
  "mode": "affine"
593
  },
594
  "language_model.model.layers.12.linear_attn.out_proj": {
@@ -657,18 +742,23 @@
657
  "mode": "affine"
658
  },
659
  "language_model.model.layers.14.linear_attn.in_proj_a": {
660
- "bits": 5,
661
- "group_size": 128,
662
  "mode": "affine"
663
  },
664
  "language_model.model.layers.14.linear_attn.in_proj_b": {
 
 
 
 
 
665
  "bits": 5,
666
- "group_size": 128,
667
  "mode": "affine"
668
  },
669
  "language_model.model.layers.14.linear_attn.in_proj_z": {
670
- "bits": 5,
671
- "group_size": 128,
672
  "mode": "affine"
673
  },
674
  "language_model.model.layers.14.linear_attn.out_proj": {
@@ -716,19 +806,44 @@
716
  "group_size": 64,
717
  "mode": "affine"
718
  },
719
- "language_model.model.layers.16.linear_attn.in_proj_a": {
720
  "bits": 5,
721
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
722
  "mode": "affine"
723
  },
724
  "language_model.model.layers.16.linear_attn.in_proj_b": {
 
 
 
 
 
725
  "bits": 5,
726
- "group_size": 128,
727
  "mode": "affine"
728
  },
729
  "language_model.model.layers.16.linear_attn.in_proj_z": {
730
- "bits": 5,
731
- "group_size": 128,
732
  "mode": "affine"
733
  },
734
  "language_model.model.layers.16.linear_attn.out_proj": {
@@ -757,18 +872,23 @@
757
  "mode": "affine"
758
  },
759
  "language_model.model.layers.17.linear_attn.in_proj_a": {
760
- "bits": 5,
761
- "group_size": 128,
762
  "mode": "affine"
763
  },
764
  "language_model.model.layers.17.linear_attn.in_proj_b": {
 
 
 
 
 
765
  "bits": 5,
766
- "group_size": 128,
767
  "mode": "affine"
768
  },
769
  "language_model.model.layers.17.linear_attn.in_proj_z": {
770
- "bits": 5,
771
- "group_size": 128,
772
  "mode": "affine"
773
  },
774
  "language_model.model.layers.17.linear_attn.out_proj": {
@@ -797,18 +917,23 @@
797
  "mode": "affine"
798
  },
799
  "language_model.model.layers.18.linear_attn.in_proj_a": {
800
- "bits": 5,
801
- "group_size": 128,
802
  "mode": "affine"
803
  },
804
  "language_model.model.layers.18.linear_attn.in_proj_b": {
 
 
 
 
 
805
  "bits": 5,
806
- "group_size": 128,
807
  "mode": "affine"
808
  },
809
  "language_model.model.layers.18.linear_attn.in_proj_z": {
810
- "bits": 5,
811
- "group_size": 128,
812
  "mode": "affine"
813
  },
814
  "language_model.model.layers.18.linear_attn.out_proj": {
@@ -856,19 +981,44 @@
856
  "group_size": 64,
857
  "mode": "affine"
858
  },
859
- "language_model.model.layers.20.linear_attn.in_proj_a": {
860
  "bits": 5,
861
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
862
  "mode": "affine"
863
  },
864
  "language_model.model.layers.20.linear_attn.in_proj_b": {
 
 
 
 
 
865
  "bits": 5,
866
- "group_size": 128,
867
  "mode": "affine"
868
  },
869
  "language_model.model.layers.20.linear_attn.in_proj_z": {
870
- "bits": 5,
871
- "group_size": 128,
872
  "mode": "affine"
873
  },
874
  "language_model.model.layers.20.linear_attn.out_proj": {
@@ -897,18 +1047,23 @@
897
  "mode": "affine"
898
  },
899
  "language_model.model.layers.21.linear_attn.in_proj_a": {
900
- "bits": 5,
901
- "group_size": 128,
902
  "mode": "affine"
903
  },
904
  "language_model.model.layers.21.linear_attn.in_proj_b": {
 
 
 
 
 
905
  "bits": 5,
906
- "group_size": 128,
907
  "mode": "affine"
908
  },
909
  "language_model.model.layers.21.linear_attn.in_proj_z": {
910
- "bits": 5,
911
- "group_size": 128,
912
  "mode": "affine"
913
  },
914
  "language_model.model.layers.21.linear_attn.out_proj": {
@@ -937,18 +1092,23 @@
937
  "mode": "affine"
938
  },
939
  "language_model.model.layers.22.linear_attn.in_proj_a": {
940
- "bits": 5,
941
- "group_size": 128,
942
  "mode": "affine"
943
  },
944
  "language_model.model.layers.22.linear_attn.in_proj_b": {
 
 
 
 
 
945
  "bits": 5,
946
- "group_size": 128,
947
  "mode": "affine"
948
  },
949
  "language_model.model.layers.22.linear_attn.in_proj_z": {
950
- "bits": 5,
951
- "group_size": 128,
952
  "mode": "affine"
953
  },
954
  "language_model.model.layers.22.linear_attn.out_proj": {
@@ -996,19 +1156,44 @@
996
  "group_size": 64,
997
  "mode": "affine"
998
  },
999
- "language_model.model.layers.24.linear_attn.in_proj_a": {
1000
  "bits": 5,
1001
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1002
  "mode": "affine"
1003
  },
1004
  "language_model.model.layers.24.linear_attn.in_proj_b": {
 
 
 
 
 
1005
  "bits": 5,
1006
- "group_size": 128,
1007
  "mode": "affine"
1008
  },
1009
  "language_model.model.layers.24.linear_attn.in_proj_z": {
1010
- "bits": 5,
1011
- "group_size": 128,
1012
  "mode": "affine"
1013
  },
1014
  "language_model.model.layers.24.linear_attn.out_proj": {
@@ -1037,18 +1222,23 @@
1037
  "mode": "affine"
1038
  },
1039
  "language_model.model.layers.25.linear_attn.in_proj_a": {
1040
- "bits": 5,
1041
- "group_size": 128,
1042
  "mode": "affine"
1043
  },
1044
  "language_model.model.layers.25.linear_attn.in_proj_b": {
 
 
 
 
 
1045
  "bits": 5,
1046
- "group_size": 128,
1047
  "mode": "affine"
1048
  },
1049
  "language_model.model.layers.25.linear_attn.in_proj_z": {
1050
- "bits": 5,
1051
- "group_size": 128,
1052
  "mode": "affine"
1053
  },
1054
  "language_model.model.layers.25.linear_attn.out_proj": {
@@ -1077,18 +1267,23 @@
1077
  "mode": "affine"
1078
  },
1079
  "language_model.model.layers.26.linear_attn.in_proj_a": {
1080
- "bits": 5,
1081
- "group_size": 128,
1082
  "mode": "affine"
1083
  },
1084
  "language_model.model.layers.26.linear_attn.in_proj_b": {
 
 
 
 
 
1085
  "bits": 5,
1086
- "group_size": 128,
1087
  "mode": "affine"
1088
  },
1089
  "language_model.model.layers.26.linear_attn.in_proj_z": {
1090
- "bits": 5,
1091
- "group_size": 128,
1092
  "mode": "affine"
1093
  },
1094
  "language_model.model.layers.26.linear_attn.out_proj": {
@@ -1136,19 +1331,44 @@
1136
  "group_size": 64,
1137
  "mode": "affine"
1138
  },
1139
- "language_model.model.layers.28.linear_attn.in_proj_a": {
1140
  "bits": 5,
1141
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1142
  "mode": "affine"
1143
  },
1144
  "language_model.model.layers.28.linear_attn.in_proj_b": {
 
 
 
 
 
1145
  "bits": 5,
1146
- "group_size": 128,
1147
  "mode": "affine"
1148
  },
1149
  "language_model.model.layers.28.linear_attn.in_proj_z": {
1150
- "bits": 5,
1151
- "group_size": 128,
1152
  "mode": "affine"
1153
  },
1154
  "language_model.model.layers.28.linear_attn.out_proj": {
@@ -1177,18 +1397,23 @@
1177
  "mode": "affine"
1178
  },
1179
  "language_model.model.layers.29.linear_attn.in_proj_a": {
1180
- "bits": 5,
1181
- "group_size": 128,
1182
  "mode": "affine"
1183
  },
1184
  "language_model.model.layers.29.linear_attn.in_proj_b": {
 
 
 
 
 
1185
  "bits": 5,
1186
- "group_size": 128,
1187
  "mode": "affine"
1188
  },
1189
  "language_model.model.layers.29.linear_attn.in_proj_z": {
1190
- "bits": 5,
1191
- "group_size": 128,
1192
  "mode": "affine"
1193
  },
1194
  "language_model.model.layers.29.linear_attn.out_proj": {
@@ -1217,18 +1442,23 @@
1217
  "mode": "affine"
1218
  },
1219
  "language_model.model.layers.30.linear_attn.in_proj_a": {
1220
- "bits": 5,
1221
- "group_size": 128,
1222
  "mode": "affine"
1223
  },
1224
  "language_model.model.layers.30.linear_attn.in_proj_b": {
 
 
 
 
 
1225
  "bits": 5,
1226
- "group_size": 128,
1227
  "mode": "affine"
1228
  },
1229
  "language_model.model.layers.30.linear_attn.in_proj_z": {
1230
- "bits": 5,
1231
- "group_size": 128,
1232
  "mode": "affine"
1233
  },
1234
  "language_model.model.layers.30.linear_attn.out_proj": {
@@ -1277,17 +1507,17 @@
1277
  "mode": "affine"
1278
  },
1279
  "language_model.model.layers.31.self_attn.k_proj": {
1280
- "bits": 8,
1281
  "group_size": 64,
1282
  "mode": "affine"
1283
  },
1284
  "language_model.model.layers.31.self_attn.o_proj": {
1285
- "bits": 8,
1286
  "group_size": 64,
1287
  "mode": "affine"
1288
  },
1289
  "language_model.model.layers.31.self_attn.q_proj": {
1290
- "bits": 8,
1291
  "group_size": 64,
1292
  "mode": "affine"
1293
  },
@@ -1306,9 +1536,14 @@
1306
  "group_size": 64,
1307
  "mode": "affine"
1308
  },
1309
- "language_model.model.layers.32.linear_attn.in_proj_z": {
1310
  "bits": 5,
1311
- "group_size": 128,
 
 
 
 
 
1312
  "mode": "affine"
1313
  },
1314
  "language_model.model.layers.32.linear_attn.out_proj": {
@@ -1337,23 +1572,23 @@
1337
  "mode": "affine"
1338
  },
1339
  "language_model.model.layers.33.linear_attn.in_proj_a": {
1340
- "bits": 8,
1341
  "group_size": 64,
1342
  "mode": "affine"
1343
  },
1344
  "language_model.model.layers.33.linear_attn.in_proj_b": {
1345
- "bits": 8,
1346
  "group_size": 64,
1347
  "mode": "affine"
1348
  },
1349
  "language_model.model.layers.33.linear_attn.in_proj_qkv": {
1350
- "bits": 5,
1351
- "group_size": 128,
1352
  "mode": "affine"
1353
  },
1354
  "language_model.model.layers.33.linear_attn.in_proj_z": {
1355
- "bits": 5,
1356
- "group_size": 128,
1357
  "mode": "affine"
1358
  },
1359
  "language_model.model.layers.33.linear_attn.out_proj": {
@@ -1391,9 +1626,14 @@
1391
  "group_size": 64,
1392
  "mode": "affine"
1393
  },
1394
- "language_model.model.layers.34.linear_attn.in_proj_z": {
1395
  "bits": 5,
1396
- "group_size": 128,
 
 
 
 
 
1397
  "mode": "affine"
1398
  },
1399
  "language_model.model.layers.34.linear_attn.out_proj": {
@@ -1442,17 +1682,17 @@
1442
  "mode": "affine"
1443
  },
1444
  "language_model.model.layers.35.self_attn.k_proj": {
1445
- "bits": 8,
1446
  "group_size": 64,
1447
  "mode": "affine"
1448
  },
1449
  "language_model.model.layers.35.self_attn.o_proj": {
1450
- "bits": 8,
1451
  "group_size": 64,
1452
  "mode": "affine"
1453
  },
1454
  "language_model.model.layers.35.self_attn.q_proj": {
1455
- "bits": 8,
1456
  "group_size": 64,
1457
  "mode": "affine"
1458
  },
@@ -1462,12 +1702,12 @@
1462
  "mode": "affine"
1463
  },
1464
  "language_model.model.layers.36.linear_attn.in_proj_a": {
1465
- "bits": 8,
1466
  "group_size": 64,
1467
  "mode": "affine"
1468
  },
1469
  "language_model.model.layers.36.linear_attn.in_proj_b": {
1470
- "bits": 8,
1471
  "group_size": 64,
1472
  "mode": "affine"
1473
  },
@@ -1477,8 +1717,8 @@
1477
  "mode": "affine"
1478
  },
1479
  "language_model.model.layers.36.linear_attn.in_proj_z": {
1480
- "bits": 5,
1481
- "group_size": 128,
1482
  "mode": "affine"
1483
  },
1484
  "language_model.model.layers.36.linear_attn.out_proj": {
@@ -1507,22 +1747,22 @@
1507
  "mode": "affine"
1508
  },
1509
  "language_model.model.layers.37.linear_attn.in_proj_a": {
1510
- "bits": 8,
1511
  "group_size": 64,
1512
  "mode": "affine"
1513
  },
1514
  "language_model.model.layers.37.linear_attn.in_proj_b": {
1515
- "bits": 8,
1516
  "group_size": 64,
1517
  "mode": "affine"
1518
  },
1519
  "language_model.model.layers.37.linear_attn.in_proj_qkv": {
1520
- "bits": 8,
1521
  "group_size": 64,
1522
  "mode": "affine"
1523
  },
1524
  "language_model.model.layers.37.linear_attn.in_proj_z": {
1525
- "bits": 8,
1526
  "group_size": 64,
1527
  "mode": "affine"
1528
  },
@@ -1552,23 +1792,23 @@
1552
  "mode": "affine"
1553
  },
1554
  "language_model.model.layers.38.linear_attn.in_proj_a": {
1555
- "bits": 8,
1556
  "group_size": 64,
1557
  "mode": "affine"
1558
  },
1559
  "language_model.model.layers.38.linear_attn.in_proj_b": {
1560
- "bits": 8,
1561
  "group_size": 64,
1562
  "mode": "affine"
1563
  },
1564
  "language_model.model.layers.38.linear_attn.in_proj_qkv": {
1565
- "bits": 5,
1566
- "group_size": 128,
1567
  "mode": "affine"
1568
  },
1569
  "language_model.model.layers.38.linear_attn.in_proj_z": {
1570
- "bits": 5,
1571
- "group_size": 128,
1572
  "mode": "affine"
1573
  },
1574
  "language_model.model.layers.38.linear_attn.out_proj": {
@@ -1597,7 +1837,7 @@
1597
  "mode": "affine"
1598
  },
1599
  "language_model.lm_head": {
1600
- "bits": 8,
1601
  "group_size": 64,
1602
  "mode": "affine"
1603
  },
@@ -1622,17 +1862,17 @@
1622
  "mode": "affine"
1623
  },
1624
  "language_model.model.layers.39.self_attn.k_proj": {
1625
- "bits": 8,
1626
  "group_size": 64,
1627
  "mode": "affine"
1628
  },
1629
  "language_model.model.layers.39.self_attn.o_proj": {
1630
- "bits": 8,
1631
  "group_size": 64,
1632
  "mode": "affine"
1633
  },
1634
  "language_model.model.layers.39.self_attn.q_proj": {
1635
- "bits": 8,
1636
  "group_size": 64,
1637
  "mode": "affine"
1638
  },
@@ -1652,23 +1892,23 @@
1652
  "mode": "affine"
1653
  },
1654
  "language_model.model.layers.0.linear_attn.in_proj_a": {
1655
- "bits": 8,
1656
  "group_size": 64,
1657
  "mode": "affine"
1658
  },
1659
  "language_model.model.layers.0.linear_attn.in_proj_b": {
1660
- "bits": 8,
1661
  "group_size": 64,
1662
  "mode": "affine"
1663
  },
1664
  "language_model.model.layers.0.linear_attn.in_proj_qkv": {
1665
- "bits": 5,
1666
- "group_size": 128,
1667
  "mode": "affine"
1668
  },
1669
  "language_model.model.layers.0.linear_attn.in_proj_z": {
1670
- "bits": 5,
1671
- "group_size": 128,
1672
  "mode": "affine"
1673
  },
1674
  "language_model.model.layers.0.linear_attn.out_proj": {
@@ -1697,18 +1937,23 @@
1697
  "mode": "affine"
1698
  },
1699
  "language_model.model.layers.1.linear_attn.in_proj_a": {
1700
- "bits": 5,
1701
- "group_size": 128,
1702
  "mode": "affine"
1703
  },
1704
  "language_model.model.layers.1.linear_attn.in_proj_b": {
 
 
 
 
 
1705
  "bits": 5,
1706
- "group_size": 128,
1707
  "mode": "affine"
1708
  },
1709
  "language_model.model.layers.1.linear_attn.in_proj_z": {
1710
- "bits": 5,
1711
- "group_size": 128,
1712
  "mode": "affine"
1713
  },
1714
  "language_model.model.layers.1.linear_attn.out_proj": {
@@ -1737,18 +1982,23 @@
1737
  "mode": "affine"
1738
  },
1739
  "language_model.model.layers.2.linear_attn.in_proj_a": {
1740
- "bits": 5,
1741
- "group_size": 128,
1742
  "mode": "affine"
1743
  },
1744
  "language_model.model.layers.2.linear_attn.in_proj_b": {
 
 
 
 
 
1745
  "bits": 5,
1746
- "group_size": 128,
1747
  "mode": "affine"
1748
  },
1749
  "language_model.model.layers.2.linear_attn.in_proj_z": {
1750
- "bits": 5,
1751
- "group_size": 128,
1752
  "mode": "affine"
1753
  },
1754
  "language_model.model.layers.2.linear_attn.out_proj": {
@@ -1796,19 +2046,44 @@
1796
  "group_size": 64,
1797
  "mode": "affine"
1798
  },
1799
- "language_model.model.layers.4.linear_attn.in_proj_a": {
1800
  "bits": 5,
1801
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1802
  "mode": "affine"
1803
  },
1804
  "language_model.model.layers.4.linear_attn.in_proj_b": {
 
 
 
 
 
1805
  "bits": 5,
1806
- "group_size": 128,
1807
  "mode": "affine"
1808
  },
1809
  "language_model.model.layers.4.linear_attn.in_proj_z": {
1810
- "bits": 5,
1811
- "group_size": 128,
1812
  "mode": "affine"
1813
  },
1814
  "language_model.model.layers.4.linear_attn.out_proj": {
@@ -1837,18 +2112,23 @@
1837
  "mode": "affine"
1838
  },
1839
  "language_model.model.layers.5.linear_attn.in_proj_a": {
1840
- "bits": 5,
1841
- "group_size": 128,
1842
  "mode": "affine"
1843
  },
1844
  "language_model.model.layers.5.linear_attn.in_proj_b": {
 
 
 
 
 
1845
  "bits": 5,
1846
- "group_size": 128,
1847
  "mode": "affine"
1848
  },
1849
  "language_model.model.layers.5.linear_attn.in_proj_z": {
1850
- "bits": 5,
1851
- "group_size": 128,
1852
  "mode": "affine"
1853
  },
1854
  "language_model.model.layers.5.linear_attn.out_proj": {
@@ -1877,18 +2157,23 @@
1877
  "mode": "affine"
1878
  },
1879
  "language_model.model.layers.6.linear_attn.in_proj_a": {
1880
- "bits": 5,
1881
- "group_size": 128,
1882
  "mode": "affine"
1883
  },
1884
  "language_model.model.layers.6.linear_attn.in_proj_b": {
 
 
 
 
 
1885
  "bits": 5,
1886
- "group_size": 128,
1887
  "mode": "affine"
1888
  },
1889
  "language_model.model.layers.6.linear_attn.in_proj_z": {
1890
- "bits": 5,
1891
- "group_size": 128,
1892
  "mode": "affine"
1893
  },
1894
  "language_model.model.layers.6.linear_attn.out_proj": {
@@ -1937,13 +2222,18 @@
1937
  "mode": "affine"
1938
  },
1939
  "language_model.model.layers.7.self_attn.k_proj": {
1940
- "bits": 8,
1941
  "group_size": 64,
1942
  "mode": "affine"
1943
  },
1944
- "language_model.model.layers.7.self_attn.q_proj": {
1945
  "bits": 5,
1946
- "group_size": 128,
 
 
 
 
 
1947
  "mode": "affine"
1948
  },
1949
  "language_model.model.layers.7.self_attn.v_proj": {
@@ -1952,18 +2242,23 @@
1952
  "mode": "affine"
1953
  },
1954
  "language_model.model.layers.8.linear_attn.in_proj_a": {
1955
- "bits": 5,
1956
- "group_size": 128,
1957
  "mode": "affine"
1958
  },
1959
  "language_model.model.layers.8.linear_attn.in_proj_b": {
 
 
 
 
 
1960
  "bits": 5,
1961
- "group_size": 128,
1962
  "mode": "affine"
1963
  },
1964
  "language_model.model.layers.8.linear_attn.in_proj_z": {
1965
- "bits": 5,
1966
- "group_size": 128,
1967
  "mode": "affine"
1968
  },
1969
  "language_model.model.layers.8.linear_attn.out_proj": {
@@ -1992,18 +2287,23 @@
1992
  "mode": "affine"
1993
  },
1994
  "language_model.model.layers.9.linear_attn.in_proj_a": {
1995
- "bits": 5,
1996
- "group_size": 128,
1997
  "mode": "affine"
1998
  },
1999
  "language_model.model.layers.9.linear_attn.in_proj_b": {
 
 
 
 
 
2000
  "bits": 5,
2001
- "group_size": 128,
2002
  "mode": "affine"
2003
  },
2004
  "language_model.model.layers.9.linear_attn.in_proj_z": {
2005
- "bits": 5,
2006
- "group_size": 128,
2007
  "mode": "affine"
2008
  },
2009
  "language_model.model.layers.9.linear_attn.out_proj": {
@@ -2032,23 +2332,23 @@
2032
  "mode": "affine"
2033
  },
2034
  "language_model.model.layers.10.linear_attn.in_proj_a": {
2035
- "bits": 8,
2036
  "group_size": 64,
2037
  "mode": "affine"
2038
  },
2039
  "language_model.model.layers.10.linear_attn.in_proj_b": {
2040
- "bits": 8,
2041
  "group_size": 64,
2042
  "mode": "affine"
2043
  },
2044
  "language_model.model.layers.10.linear_attn.in_proj_qkv": {
2045
- "bits": 5,
2046
- "group_size": 128,
2047
  "mode": "affine"
2048
  },
2049
  "language_model.model.layers.10.linear_attn.in_proj_z": {
2050
- "bits": 5,
2051
- "group_size": 128,
2052
  "mode": "affine"
2053
  },
2054
  "language_model.model.layers.10.linear_attn.out_proj": {
@@ -2096,19 +2396,44 @@
2096
  "group_size": 64,
2097
  "mode": "affine"
2098
  },
2099
- "language_model.model.layers.12.linear_attn.in_proj_a": {
2100
  "bits": 5,
2101
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2102
  "mode": "affine"
2103
  },
2104
  "language_model.model.layers.12.linear_attn.in_proj_b": {
 
 
 
 
 
2105
  "bits": 5,
2106
- "group_size": 128,
2107
  "mode": "affine"
2108
  },
2109
  "language_model.model.layers.12.linear_attn.in_proj_z": {
2110
- "bits": 5,
2111
- "group_size": 128,
2112
  "mode": "affine"
2113
  },
2114
  "language_model.model.layers.12.linear_attn.out_proj": {
@@ -2177,18 +2502,23 @@
2177
  "mode": "affine"
2178
  },
2179
  "language_model.model.layers.14.linear_attn.in_proj_a": {
2180
- "bits": 5,
2181
- "group_size": 128,
2182
  "mode": "affine"
2183
  },
2184
  "language_model.model.layers.14.linear_attn.in_proj_b": {
 
 
 
 
 
2185
  "bits": 5,
2186
- "group_size": 128,
2187
  "mode": "affine"
2188
  },
2189
  "language_model.model.layers.14.linear_attn.in_proj_z": {
2190
- "bits": 5,
2191
- "group_size": 128,
2192
  "mode": "affine"
2193
  },
2194
  "language_model.model.layers.14.linear_attn.out_proj": {
@@ -2236,19 +2566,44 @@
2236
  "group_size": 64,
2237
  "mode": "affine"
2238
  },
2239
- "language_model.model.layers.16.linear_attn.in_proj_a": {
2240
  "bits": 5,
2241
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2242
  "mode": "affine"
2243
  },
2244
  "language_model.model.layers.16.linear_attn.in_proj_b": {
 
 
 
 
 
2245
  "bits": 5,
2246
- "group_size": 128,
2247
  "mode": "affine"
2248
  },
2249
  "language_model.model.layers.16.linear_attn.in_proj_z": {
2250
- "bits": 5,
2251
- "group_size": 128,
2252
  "mode": "affine"
2253
  },
2254
  "language_model.model.layers.16.linear_attn.out_proj": {
@@ -2277,18 +2632,23 @@
2277
  "mode": "affine"
2278
  },
2279
  "language_model.model.layers.17.linear_attn.in_proj_a": {
2280
- "bits": 5,
2281
- "group_size": 128,
2282
  "mode": "affine"
2283
  },
2284
  "language_model.model.layers.17.linear_attn.in_proj_b": {
 
 
 
 
 
2285
  "bits": 5,
2286
- "group_size": 128,
2287
  "mode": "affine"
2288
  },
2289
  "language_model.model.layers.17.linear_attn.in_proj_z": {
2290
- "bits": 5,
2291
- "group_size": 128,
2292
  "mode": "affine"
2293
  },
2294
  "language_model.model.layers.17.linear_attn.out_proj": {
@@ -2317,18 +2677,23 @@
2317
  "mode": "affine"
2318
  },
2319
  "language_model.model.layers.18.linear_attn.in_proj_a": {
2320
- "bits": 5,
2321
- "group_size": 128,
2322
  "mode": "affine"
2323
  },
2324
  "language_model.model.layers.18.linear_attn.in_proj_b": {
 
 
 
 
 
2325
  "bits": 5,
2326
- "group_size": 128,
2327
  "mode": "affine"
2328
  },
2329
  "language_model.model.layers.18.linear_attn.in_proj_z": {
2330
- "bits": 5,
2331
- "group_size": 128,
2332
  "mode": "affine"
2333
  },
2334
  "language_model.model.layers.18.linear_attn.out_proj": {
@@ -2376,19 +2741,44 @@
2376
  "group_size": 64,
2377
  "mode": "affine"
2378
  },
2379
- "language_model.model.layers.20.linear_attn.in_proj_a": {
2380
  "bits": 5,
2381
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2382
  "mode": "affine"
2383
  },
2384
  "language_model.model.layers.20.linear_attn.in_proj_b": {
 
 
 
 
 
2385
  "bits": 5,
2386
- "group_size": 128,
2387
  "mode": "affine"
2388
  },
2389
  "language_model.model.layers.20.linear_attn.in_proj_z": {
2390
- "bits": 5,
2391
- "group_size": 128,
2392
  "mode": "affine"
2393
  },
2394
  "language_model.model.layers.20.linear_attn.out_proj": {
@@ -2417,18 +2807,23 @@
2417
  "mode": "affine"
2418
  },
2419
  "language_model.model.layers.21.linear_attn.in_proj_a": {
2420
- "bits": 5,
2421
- "group_size": 128,
2422
  "mode": "affine"
2423
  },
2424
  "language_model.model.layers.21.linear_attn.in_proj_b": {
 
 
 
 
 
2425
  "bits": 5,
2426
- "group_size": 128,
2427
  "mode": "affine"
2428
  },
2429
  "language_model.model.layers.21.linear_attn.in_proj_z": {
2430
- "bits": 5,
2431
- "group_size": 128,
2432
  "mode": "affine"
2433
  },
2434
  "language_model.model.layers.21.linear_attn.out_proj": {
@@ -2457,18 +2852,23 @@
2457
  "mode": "affine"
2458
  },
2459
  "language_model.model.layers.22.linear_attn.in_proj_a": {
2460
- "bits": 5,
2461
- "group_size": 128,
2462
  "mode": "affine"
2463
  },
2464
  "language_model.model.layers.22.linear_attn.in_proj_b": {
 
 
 
 
 
2465
  "bits": 5,
2466
- "group_size": 128,
2467
  "mode": "affine"
2468
  },
2469
  "language_model.model.layers.22.linear_attn.in_proj_z": {
2470
- "bits": 5,
2471
- "group_size": 128,
2472
  "mode": "affine"
2473
  },
2474
  "language_model.model.layers.22.linear_attn.out_proj": {
@@ -2516,19 +2916,44 @@
2516
  "group_size": 64,
2517
  "mode": "affine"
2518
  },
2519
- "language_model.model.layers.24.linear_attn.in_proj_a": {
2520
  "bits": 5,
2521
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2522
  "mode": "affine"
2523
  },
2524
  "language_model.model.layers.24.linear_attn.in_proj_b": {
 
 
 
 
 
2525
  "bits": 5,
2526
- "group_size": 128,
2527
  "mode": "affine"
2528
  },
2529
  "language_model.model.layers.24.linear_attn.in_proj_z": {
2530
- "bits": 5,
2531
- "group_size": 128,
2532
  "mode": "affine"
2533
  },
2534
  "language_model.model.layers.24.linear_attn.out_proj": {
@@ -2557,18 +2982,23 @@
2557
  "mode": "affine"
2558
  },
2559
  "language_model.model.layers.25.linear_attn.in_proj_a": {
2560
- "bits": 5,
2561
- "group_size": 128,
2562
  "mode": "affine"
2563
  },
2564
  "language_model.model.layers.25.linear_attn.in_proj_b": {
 
 
 
 
 
2565
  "bits": 5,
2566
- "group_size": 128,
2567
  "mode": "affine"
2568
  },
2569
  "language_model.model.layers.25.linear_attn.in_proj_z": {
2570
- "bits": 5,
2571
- "group_size": 128,
2572
  "mode": "affine"
2573
  },
2574
  "language_model.model.layers.25.linear_attn.out_proj": {
@@ -2597,18 +3027,23 @@
2597
  "mode": "affine"
2598
  },
2599
  "language_model.model.layers.26.linear_attn.in_proj_a": {
2600
- "bits": 5,
2601
- "group_size": 128,
2602
  "mode": "affine"
2603
  },
2604
  "language_model.model.layers.26.linear_attn.in_proj_b": {
 
 
 
 
 
2605
  "bits": 5,
2606
- "group_size": 128,
2607
  "mode": "affine"
2608
  },
2609
  "language_model.model.layers.26.linear_attn.in_proj_z": {
2610
- "bits": 5,
2611
- "group_size": 128,
2612
  "mode": "affine"
2613
  },
2614
  "language_model.model.layers.26.linear_attn.out_proj": {
@@ -2656,19 +3091,44 @@
2656
  "group_size": 64,
2657
  "mode": "affine"
2658
  },
2659
- "language_model.model.layers.28.linear_attn.in_proj_a": {
2660
  "bits": 5,
2661
- "group_size": 128,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2662
  "mode": "affine"
2663
  },
2664
  "language_model.model.layers.28.linear_attn.in_proj_b": {
 
 
 
 
 
2665
  "bits": 5,
2666
- "group_size": 128,
2667
  "mode": "affine"
2668
  },
2669
  "language_model.model.layers.28.linear_attn.in_proj_z": {
2670
- "bits": 5,
2671
- "group_size": 128,
2672
  "mode": "affine"
2673
  },
2674
  "language_model.model.layers.28.linear_attn.out_proj": {
@@ -2697,18 +3157,23 @@
2697
  "mode": "affine"
2698
  },
2699
  "language_model.model.layers.29.linear_attn.in_proj_a": {
2700
- "bits": 5,
2701
- "group_size": 128,
2702
  "mode": "affine"
2703
  },
2704
  "language_model.model.layers.29.linear_attn.in_proj_b": {
 
 
 
 
 
2705
  "bits": 5,
2706
- "group_size": 128,
2707
  "mode": "affine"
2708
  },
2709
  "language_model.model.layers.29.linear_attn.in_proj_z": {
2710
- "bits": 5,
2711
- "group_size": 128,
2712
  "mode": "affine"
2713
  },
2714
  "language_model.model.layers.29.linear_attn.out_proj": {
@@ -2737,18 +3202,23 @@
2737
  "mode": "affine"
2738
  },
2739
  "language_model.model.layers.30.linear_attn.in_proj_a": {
2740
- "bits": 5,
2741
- "group_size": 128,
2742
  "mode": "affine"
2743
  },
2744
  "language_model.model.layers.30.linear_attn.in_proj_b": {
 
 
 
 
 
2745
  "bits": 5,
2746
- "group_size": 128,
2747
  "mode": "affine"
2748
  },
2749
  "language_model.model.layers.30.linear_attn.in_proj_z": {
2750
- "bits": 5,
2751
- "group_size": 128,
2752
  "mode": "affine"
2753
  },
2754
  "language_model.model.layers.30.linear_attn.out_proj": {
@@ -2797,17 +3267,17 @@
2797
  "mode": "affine"
2798
  },
2799
  "language_model.model.layers.31.self_attn.k_proj": {
2800
- "bits": 8,
2801
  "group_size": 64,
2802
  "mode": "affine"
2803
  },
2804
  "language_model.model.layers.31.self_attn.o_proj": {
2805
- "bits": 8,
2806
  "group_size": 64,
2807
  "mode": "affine"
2808
  },
2809
  "language_model.model.layers.31.self_attn.q_proj": {
2810
- "bits": 8,
2811
  "group_size": 64,
2812
  "mode": "affine"
2813
  },
@@ -2826,9 +3296,14 @@
2826
  "group_size": 64,
2827
  "mode": "affine"
2828
  },
2829
- "language_model.model.layers.32.linear_attn.in_proj_z": {
2830
  "bits": 5,
2831
- "group_size": 128,
 
 
 
 
 
2832
  "mode": "affine"
2833
  },
2834
  "language_model.model.layers.32.linear_attn.out_proj": {
@@ -2857,23 +3332,23 @@
2857
  "mode": "affine"
2858
  },
2859
  "language_model.model.layers.33.linear_attn.in_proj_a": {
2860
- "bits": 8,
2861
  "group_size": 64,
2862
  "mode": "affine"
2863
  },
2864
  "language_model.model.layers.33.linear_attn.in_proj_b": {
2865
- "bits": 8,
2866
  "group_size": 64,
2867
  "mode": "affine"
2868
  },
2869
  "language_model.model.layers.33.linear_attn.in_proj_qkv": {
2870
- "bits": 5,
2871
- "group_size": 128,
2872
  "mode": "affine"
2873
  },
2874
  "language_model.model.layers.33.linear_attn.in_proj_z": {
2875
- "bits": 5,
2876
- "group_size": 128,
2877
  "mode": "affine"
2878
  },
2879
  "language_model.model.layers.33.linear_attn.out_proj": {
@@ -2911,9 +3386,14 @@
2911
  "group_size": 64,
2912
  "mode": "affine"
2913
  },
2914
- "language_model.model.layers.34.linear_attn.in_proj_z": {
2915
  "bits": 5,
2916
- "group_size": 128,
 
 
 
 
 
2917
  "mode": "affine"
2918
  },
2919
  "language_model.model.layers.34.linear_attn.out_proj": {
@@ -2962,17 +3442,17 @@
2962
  "mode": "affine"
2963
  },
2964
  "language_model.model.layers.35.self_attn.k_proj": {
2965
- "bits": 8,
2966
  "group_size": 64,
2967
  "mode": "affine"
2968
  },
2969
  "language_model.model.layers.35.self_attn.o_proj": {
2970
- "bits": 8,
2971
  "group_size": 64,
2972
  "mode": "affine"
2973
  },
2974
  "language_model.model.layers.35.self_attn.q_proj": {
2975
- "bits": 8,
2976
  "group_size": 64,
2977
  "mode": "affine"
2978
  },
@@ -2982,12 +3462,12 @@
2982
  "mode": "affine"
2983
  },
2984
  "language_model.model.layers.36.linear_attn.in_proj_a": {
2985
- "bits": 8,
2986
  "group_size": 64,
2987
  "mode": "affine"
2988
  },
2989
  "language_model.model.layers.36.linear_attn.in_proj_b": {
2990
- "bits": 8,
2991
  "group_size": 64,
2992
  "mode": "affine"
2993
  },
@@ -2997,8 +3477,8 @@
2997
  "mode": "affine"
2998
  },
2999
  "language_model.model.layers.36.linear_attn.in_proj_z": {
3000
- "bits": 5,
3001
- "group_size": 128,
3002
  "mode": "affine"
3003
  },
3004
  "language_model.model.layers.36.linear_attn.out_proj": {
@@ -3027,22 +3507,22 @@
3027
  "mode": "affine"
3028
  },
3029
  "language_model.model.layers.37.linear_attn.in_proj_a": {
3030
- "bits": 8,
3031
  "group_size": 64,
3032
  "mode": "affine"
3033
  },
3034
  "language_model.model.layers.37.linear_attn.in_proj_b": {
3035
- "bits": 8,
3036
  "group_size": 64,
3037
  "mode": "affine"
3038
  },
3039
  "language_model.model.layers.37.linear_attn.in_proj_qkv": {
3040
- "bits": 8,
3041
  "group_size": 64,
3042
  "mode": "affine"
3043
  },
3044
  "language_model.model.layers.37.linear_attn.in_proj_z": {
3045
- "bits": 8,
3046
  "group_size": 64,
3047
  "mode": "affine"
3048
  },
@@ -3072,23 +3552,23 @@
3072
  "mode": "affine"
3073
  },
3074
  "language_model.model.layers.38.linear_attn.in_proj_a": {
3075
- "bits": 8,
3076
  "group_size": 64,
3077
  "mode": "affine"
3078
  },
3079
  "language_model.model.layers.38.linear_attn.in_proj_b": {
3080
- "bits": 8,
3081
  "group_size": 64,
3082
  "mode": "affine"
3083
  },
3084
  "language_model.model.layers.38.linear_attn.in_proj_qkv": {
3085
- "bits": 5,
3086
- "group_size": 128,
3087
  "mode": "affine"
3088
  },
3089
  "language_model.model.layers.38.linear_attn.in_proj_z": {
3090
- "bits": 5,
3091
- "group_size": 128,
3092
  "mode": "affine"
3093
  },
3094
  "language_model.model.layers.38.linear_attn.out_proj": {
@@ -3117,7 +3597,7 @@
3117
  "mode": "affine"
3118
  },
3119
  "language_model.lm_head": {
3120
- "bits": 8,
3121
  "group_size": 64,
3122
  "mode": "affine"
3123
  },
@@ -3142,17 +3622,17 @@
3142
  "mode": "affine"
3143
  },
3144
  "language_model.model.layers.39.self_attn.k_proj": {
3145
- "bits": 8,
3146
  "group_size": 64,
3147
  "mode": "affine"
3148
  },
3149
  "language_model.model.layers.39.self_attn.o_proj": {
3150
- "bits": 8,
3151
  "group_size": 64,
3152
  "mode": "affine"
3153
  },
3154
  "language_model.model.layers.39.self_attn.q_proj": {
3155
- "bits": 8,
3156
  "group_size": 64,
3157
  "mode": "affine"
3158
  },
 
132
  "mode": "affine"
133
  },
134
  "language_model.model.layers.0.linear_attn.in_proj_a": {
135
+ "bits": 6,
136
  "group_size": 64,
137
  "mode": "affine"
138
  },
139
  "language_model.model.layers.0.linear_attn.in_proj_b": {
140
+ "bits": 6,
141
  "group_size": 64,
142
  "mode": "affine"
143
  },
144
  "language_model.model.layers.0.linear_attn.in_proj_qkv": {
145
+ "bits": 6,
146
+ "group_size": 64,
147
  "mode": "affine"
148
  },
149
  "language_model.model.layers.0.linear_attn.in_proj_z": {
150
+ "bits": 6,
151
+ "group_size": 64,
152
  "mode": "affine"
153
  },
154
  "language_model.model.layers.0.linear_attn.out_proj": {
 
177
  "mode": "affine"
178
  },
179
  "language_model.model.layers.1.linear_attn.in_proj_a": {
180
+ "bits": 6,
181
+ "group_size": 64,
182
  "mode": "affine"
183
  },
184
  "language_model.model.layers.1.linear_attn.in_proj_b": {
185
+ "bits": 6,
186
+ "group_size": 64,
187
+ "mode": "affine"
188
+ },
189
+ "language_model.model.layers.1.linear_attn.in_proj_qkv": {
190
  "bits": 5,
191
+ "group_size": 64,
192
  "mode": "affine"
193
  },
194
  "language_model.model.layers.1.linear_attn.in_proj_z": {
195
+ "bits": 6,
196
+ "group_size": 64,
197
  "mode": "affine"
198
  },
199
  "language_model.model.layers.1.linear_attn.out_proj": {
 
222
  "mode": "affine"
223
  },
224
  "language_model.model.layers.2.linear_attn.in_proj_a": {
225
+ "bits": 6,
226
+ "group_size": 64,
227
  "mode": "affine"
228
  },
229
  "language_model.model.layers.2.linear_attn.in_proj_b": {
230
+ "bits": 6,
231
+ "group_size": 64,
232
+ "mode": "affine"
233
+ },
234
+ "language_model.model.layers.2.linear_attn.in_proj_qkv": {
235
  "bits": 5,
236
+ "group_size": 64,
237
  "mode": "affine"
238
  },
239
  "language_model.model.layers.2.linear_attn.in_proj_z": {
240
+ "bits": 6,
241
+ "group_size": 64,
242
  "mode": "affine"
243
  },
244
  "language_model.model.layers.2.linear_attn.out_proj": {
 
286
  "group_size": 64,
287
  "mode": "affine"
288
  },
289
+ "language_model.model.layers.3.self_attn.k_proj": {
290
  "bits": 5,
291
+ "group_size": 64,
292
+ "mode": "affine"
293
+ },
294
+ "language_model.model.layers.3.self_attn.o_proj": {
295
+ "bits": 5,
296
+ "group_size": 64,
297
+ "mode": "affine"
298
+ },
299
+ "language_model.model.layers.3.self_attn.q_proj": {
300
+ "bits": 5,
301
+ "group_size": 64,
302
+ "mode": "affine"
303
+ },
304
+ "language_model.model.layers.3.self_attn.v_proj": {
305
+ "bits": 5,
306
+ "group_size": 64,
307
+ "mode": "affine"
308
+ },
309
+ "language_model.model.layers.4.linear_attn.in_proj_a": {
310
+ "bits": 6,
311
+ "group_size": 64,
312
  "mode": "affine"
313
  },
314
  "language_model.model.layers.4.linear_attn.in_proj_b": {
315
+ "bits": 6,
316
+ "group_size": 64,
317
+ "mode": "affine"
318
+ },
319
+ "language_model.model.layers.4.linear_attn.in_proj_qkv": {
320
  "bits": 5,
321
+ "group_size": 64,
322
  "mode": "affine"
323
  },
324
  "language_model.model.layers.4.linear_attn.in_proj_z": {
325
+ "bits": 6,
326
+ "group_size": 64,
327
  "mode": "affine"
328
  },
329
  "language_model.model.layers.4.linear_attn.out_proj": {
 
352
  "mode": "affine"
353
  },
354
  "language_model.model.layers.5.linear_attn.in_proj_a": {
355
+ "bits": 6,
356
+ "group_size": 64,
357
  "mode": "affine"
358
  },
359
  "language_model.model.layers.5.linear_attn.in_proj_b": {
360
+ "bits": 6,
361
+ "group_size": 64,
362
+ "mode": "affine"
363
+ },
364
+ "language_model.model.layers.5.linear_attn.in_proj_qkv": {
365
  "bits": 5,
366
+ "group_size": 64,
367
  "mode": "affine"
368
  },
369
  "language_model.model.layers.5.linear_attn.in_proj_z": {
370
+ "bits": 6,
371
+ "group_size": 64,
372
  "mode": "affine"
373
  },
374
  "language_model.model.layers.5.linear_attn.out_proj": {
 
397
  "mode": "affine"
398
  },
399
  "language_model.model.layers.6.linear_attn.in_proj_a": {
400
+ "bits": 6,
401
+ "group_size": 64,
402
  "mode": "affine"
403
  },
404
  "language_model.model.layers.6.linear_attn.in_proj_b": {
405
+ "bits": 6,
406
+ "group_size": 64,
407
+ "mode": "affine"
408
+ },
409
+ "language_model.model.layers.6.linear_attn.in_proj_qkv": {
410
  "bits": 5,
411
+ "group_size": 64,
412
  "mode": "affine"
413
  },
414
  "language_model.model.layers.6.linear_attn.in_proj_z": {
415
+ "bits": 6,
416
+ "group_size": 64,
417
  "mode": "affine"
418
  },
419
  "language_model.model.layers.6.linear_attn.out_proj": {
 
462
  "mode": "affine"
463
  },
464
  "language_model.model.layers.7.self_attn.k_proj": {
465
+ "bits": 6,
466
  "group_size": 64,
467
  "mode": "affine"
468
  },
469
+ "language_model.model.layers.7.self_attn.o_proj": {
470
  "bits": 5,
471
+ "group_size": 64,
472
+ "mode": "affine"
473
+ },
474
+ "language_model.model.layers.7.self_attn.q_proj": {
475
+ "bits": 6,
476
+ "group_size": 64,
477
  "mode": "affine"
478
  },
479
  "language_model.model.layers.7.self_attn.v_proj": {
 
482
  "mode": "affine"
483
  },
484
  "language_model.model.layers.8.linear_attn.in_proj_a": {
485
+ "bits": 6,
486
+ "group_size": 64,
487
  "mode": "affine"
488
  },
489
  "language_model.model.layers.8.linear_attn.in_proj_b": {
490
+ "bits": 6,
491
+ "group_size": 64,
492
+ "mode": "affine"
493
+ },
494
+ "language_model.model.layers.8.linear_attn.in_proj_qkv": {
495
  "bits": 5,
496
+ "group_size": 64,
497
  "mode": "affine"
498
  },
499
  "language_model.model.layers.8.linear_attn.in_proj_z": {
500
+ "bits": 6,
501
+ "group_size": 64,
502
  "mode": "affine"
503
  },
504
  "language_model.model.layers.8.linear_attn.out_proj": {
 
527
  "mode": "affine"
528
  },
529
  "language_model.model.layers.9.linear_attn.in_proj_a": {
530
+ "bits": 6,
531
+ "group_size": 64,
532
  "mode": "affine"
533
  },
534
  "language_model.model.layers.9.linear_attn.in_proj_b": {
535
+ "bits": 6,
536
+ "group_size": 64,
537
+ "mode": "affine"
538
+ },
539
+ "language_model.model.layers.9.linear_attn.in_proj_qkv": {
540
  "bits": 5,
541
+ "group_size": 64,
542
  "mode": "affine"
543
  },
544
  "language_model.model.layers.9.linear_attn.in_proj_z": {
545
+ "bits": 6,
546
+ "group_size": 64,
547
  "mode": "affine"
548
  },
549
  "language_model.model.layers.9.linear_attn.out_proj": {
 
572
  "mode": "affine"
573
  },
574
  "language_model.model.layers.10.linear_attn.in_proj_a": {
575
+ "bits": 6,
576
  "group_size": 64,
577
  "mode": "affine"
578
  },
579
  "language_model.model.layers.10.linear_attn.in_proj_b": {
580
+ "bits": 6,
581
  "group_size": 64,
582
  "mode": "affine"
583
  },
584
  "language_model.model.layers.10.linear_attn.in_proj_qkv": {
585
+ "bits": 6,
586
+ "group_size": 64,
587
  "mode": "affine"
588
  },
589
  "language_model.model.layers.10.linear_attn.in_proj_z": {
590
+ "bits": 6,
591
+ "group_size": 64,
592
  "mode": "affine"
593
  },
594
  "language_model.model.layers.10.linear_attn.out_proj": {
 
636
  "group_size": 64,
637
  "mode": "affine"
638
  },
639
+ "language_model.model.layers.11.self_attn.k_proj": {
640
  "bits": 5,
641
+ "group_size": 64,
642
+ "mode": "affine"
643
+ },
644
+ "language_model.model.layers.11.self_attn.o_proj": {
645
+ "bits": 5,
646
+ "group_size": 64,
647
+ "mode": "affine"
648
+ },
649
+ "language_model.model.layers.11.self_attn.q_proj": {
650
+ "bits": 5,
651
+ "group_size": 64,
652
+ "mode": "affine"
653
+ },
654
+ "language_model.model.layers.11.self_attn.v_proj": {
655
+ "bits": 5,
656
+ "group_size": 64,
657
+ "mode": "affine"
658
+ },
659
+ "language_model.model.layers.12.linear_attn.in_proj_a": {
660
+ "bits": 6,
661
+ "group_size": 64,
662
  "mode": "affine"
663
  },
664
  "language_model.model.layers.12.linear_attn.in_proj_b": {
665
+ "bits": 6,
666
+ "group_size": 64,
667
+ "mode": "affine"
668
+ },
669
+ "language_model.model.layers.12.linear_attn.in_proj_qkv": {
670
  "bits": 5,
671
+ "group_size": 64,
672
  "mode": "affine"
673
  },
674
  "language_model.model.layers.12.linear_attn.in_proj_z": {
675
+ "bits": 6,
676
+ "group_size": 64,
677
  "mode": "affine"
678
  },
679
  "language_model.model.layers.12.linear_attn.out_proj": {
 
742
  "mode": "affine"
743
  },
744
  "language_model.model.layers.14.linear_attn.in_proj_a": {
745
+ "bits": 6,
746
+ "group_size": 64,
747
  "mode": "affine"
748
  },
749
  "language_model.model.layers.14.linear_attn.in_proj_b": {
750
+ "bits": 6,
751
+ "group_size": 64,
752
+ "mode": "affine"
753
+ },
754
+ "language_model.model.layers.14.linear_attn.in_proj_qkv": {
755
  "bits": 5,
756
+ "group_size": 64,
757
  "mode": "affine"
758
  },
759
  "language_model.model.layers.14.linear_attn.in_proj_z": {
760
+ "bits": 6,
761
+ "group_size": 64,
762
  "mode": "affine"
763
  },
764
  "language_model.model.layers.14.linear_attn.out_proj": {
 
806
  "group_size": 64,
807
  "mode": "affine"
808
  },
809
+ "language_model.model.layers.15.self_attn.k_proj": {
810
  "bits": 5,
811
+ "group_size": 64,
812
+ "mode": "affine"
813
+ },
814
+ "language_model.model.layers.15.self_attn.o_proj": {
815
+ "bits": 5,
816
+ "group_size": 64,
817
+ "mode": "affine"
818
+ },
819
+ "language_model.model.layers.15.self_attn.q_proj": {
820
+ "bits": 5,
821
+ "group_size": 64,
822
+ "mode": "affine"
823
+ },
824
+ "language_model.model.layers.15.self_attn.v_proj": {
825
+ "bits": 5,
826
+ "group_size": 64,
827
+ "mode": "affine"
828
+ },
829
+ "language_model.model.layers.16.linear_attn.in_proj_a": {
830
+ "bits": 6,
831
+ "group_size": 64,
832
  "mode": "affine"
833
  },
834
  "language_model.model.layers.16.linear_attn.in_proj_b": {
835
+ "bits": 6,
836
+ "group_size": 64,
837
+ "mode": "affine"
838
+ },
839
+ "language_model.model.layers.16.linear_attn.in_proj_qkv": {
840
  "bits": 5,
841
+ "group_size": 64,
842
  "mode": "affine"
843
  },
844
  "language_model.model.layers.16.linear_attn.in_proj_z": {
845
+ "bits": 6,
846
+ "group_size": 64,
847
  "mode": "affine"
848
  },
849
  "language_model.model.layers.16.linear_attn.out_proj": {
 
872
  "mode": "affine"
873
  },
874
  "language_model.model.layers.17.linear_attn.in_proj_a": {
875
+ "bits": 6,
876
+ "group_size": 64,
877
  "mode": "affine"
878
  },
879
  "language_model.model.layers.17.linear_attn.in_proj_b": {
880
+ "bits": 6,
881
+ "group_size": 64,
882
+ "mode": "affine"
883
+ },
884
+ "language_model.model.layers.17.linear_attn.in_proj_qkv": {
885
  "bits": 5,
886
+ "group_size": 64,
887
  "mode": "affine"
888
  },
889
  "language_model.model.layers.17.linear_attn.in_proj_z": {
890
+ "bits": 6,
891
+ "group_size": 64,
892
  "mode": "affine"
893
  },
894
  "language_model.model.layers.17.linear_attn.out_proj": {
 
917
  "mode": "affine"
918
  },
919
  "language_model.model.layers.18.linear_attn.in_proj_a": {
920
+ "bits": 6,
921
+ "group_size": 64,
922
  "mode": "affine"
923
  },
924
  "language_model.model.layers.18.linear_attn.in_proj_b": {
925
+ "bits": 6,
926
+ "group_size": 64,
927
+ "mode": "affine"
928
+ },
929
+ "language_model.model.layers.18.linear_attn.in_proj_qkv": {
930
  "bits": 5,
931
+ "group_size": 64,
932
  "mode": "affine"
933
  },
934
  "language_model.model.layers.18.linear_attn.in_proj_z": {
935
+ "bits": 6,
936
+ "group_size": 64,
937
  "mode": "affine"
938
  },
939
  "language_model.model.layers.18.linear_attn.out_proj": {
 
981
  "group_size": 64,
982
  "mode": "affine"
983
  },
984
+ "language_model.model.layers.19.self_attn.k_proj": {
985
  "bits": 5,
986
+ "group_size": 64,
987
+ "mode": "affine"
988
+ },
989
+ "language_model.model.layers.19.self_attn.o_proj": {
990
+ "bits": 5,
991
+ "group_size": 64,
992
+ "mode": "affine"
993
+ },
994
+ "language_model.model.layers.19.self_attn.q_proj": {
995
+ "bits": 5,
996
+ "group_size": 64,
997
+ "mode": "affine"
998
+ },
999
+ "language_model.model.layers.19.self_attn.v_proj": {
1000
+ "bits": 5,
1001
+ "group_size": 64,
1002
+ "mode": "affine"
1003
+ },
1004
+ "language_model.model.layers.20.linear_attn.in_proj_a": {
1005
+ "bits": 6,
1006
+ "group_size": 64,
1007
  "mode": "affine"
1008
  },
1009
  "language_model.model.layers.20.linear_attn.in_proj_b": {
1010
+ "bits": 6,
1011
+ "group_size": 64,
1012
+ "mode": "affine"
1013
+ },
1014
+ "language_model.model.layers.20.linear_attn.in_proj_qkv": {
1015
  "bits": 5,
1016
+ "group_size": 64,
1017
  "mode": "affine"
1018
  },
1019
  "language_model.model.layers.20.linear_attn.in_proj_z": {
1020
+ "bits": 6,
1021
+ "group_size": 64,
1022
  "mode": "affine"
1023
  },
1024
  "language_model.model.layers.20.linear_attn.out_proj": {
 
1047
  "mode": "affine"
1048
  },
1049
  "language_model.model.layers.21.linear_attn.in_proj_a": {
1050
+ "bits": 6,
1051
+ "group_size": 64,
1052
  "mode": "affine"
1053
  },
1054
  "language_model.model.layers.21.linear_attn.in_proj_b": {
1055
+ "bits": 6,
1056
+ "group_size": 64,
1057
+ "mode": "affine"
1058
+ },
1059
+ "language_model.model.layers.21.linear_attn.in_proj_qkv": {
1060
  "bits": 5,
1061
+ "group_size": 64,
1062
  "mode": "affine"
1063
  },
1064
  "language_model.model.layers.21.linear_attn.in_proj_z": {
1065
+ "bits": 6,
1066
+ "group_size": 64,
1067
  "mode": "affine"
1068
  },
1069
  "language_model.model.layers.21.linear_attn.out_proj": {
 
1092
  "mode": "affine"
1093
  },
1094
  "language_model.model.layers.22.linear_attn.in_proj_a": {
1095
+ "bits": 6,
1096
+ "group_size": 64,
1097
  "mode": "affine"
1098
  },
1099
  "language_model.model.layers.22.linear_attn.in_proj_b": {
1100
+ "bits": 6,
1101
+ "group_size": 64,
1102
+ "mode": "affine"
1103
+ },
1104
+ "language_model.model.layers.22.linear_attn.in_proj_qkv": {
1105
  "bits": 5,
1106
+ "group_size": 64,
1107
  "mode": "affine"
1108
  },
1109
  "language_model.model.layers.22.linear_attn.in_proj_z": {
1110
+ "bits": 6,
1111
+ "group_size": 64,
1112
  "mode": "affine"
1113
  },
1114
  "language_model.model.layers.22.linear_attn.out_proj": {
 
1156
  "group_size": 64,
1157
  "mode": "affine"
1158
  },
1159
+ "language_model.model.layers.23.self_attn.k_proj": {
1160
  "bits": 5,
1161
+ "group_size": 64,
1162
+ "mode": "affine"
1163
+ },
1164
+ "language_model.model.layers.23.self_attn.o_proj": {
1165
+ "bits": 5,
1166
+ "group_size": 64,
1167
+ "mode": "affine"
1168
+ },
1169
+ "language_model.model.layers.23.self_attn.q_proj": {
1170
+ "bits": 5,
1171
+ "group_size": 64,
1172
+ "mode": "affine"
1173
+ },
1174
+ "language_model.model.layers.23.self_attn.v_proj": {
1175
+ "bits": 5,
1176
+ "group_size": 64,
1177
+ "mode": "affine"
1178
+ },
1179
+ "language_model.model.layers.24.linear_attn.in_proj_a": {
1180
+ "bits": 6,
1181
+ "group_size": 64,
1182
  "mode": "affine"
1183
  },
1184
  "language_model.model.layers.24.linear_attn.in_proj_b": {
1185
+ "bits": 6,
1186
+ "group_size": 64,
1187
+ "mode": "affine"
1188
+ },
1189
+ "language_model.model.layers.24.linear_attn.in_proj_qkv": {
1190
  "bits": 5,
1191
+ "group_size": 64,
1192
  "mode": "affine"
1193
  },
1194
  "language_model.model.layers.24.linear_attn.in_proj_z": {
1195
+ "bits": 6,
1196
+ "group_size": 64,
1197
  "mode": "affine"
1198
  },
1199
  "language_model.model.layers.24.linear_attn.out_proj": {
 
1222
  "mode": "affine"
1223
  },
1224
  "language_model.model.layers.25.linear_attn.in_proj_a": {
1225
+ "bits": 6,
1226
+ "group_size": 64,
1227
  "mode": "affine"
1228
  },
1229
  "language_model.model.layers.25.linear_attn.in_proj_b": {
1230
+ "bits": 6,
1231
+ "group_size": 64,
1232
+ "mode": "affine"
1233
+ },
1234
+ "language_model.model.layers.25.linear_attn.in_proj_qkv": {
1235
  "bits": 5,
1236
+ "group_size": 64,
1237
  "mode": "affine"
1238
  },
1239
  "language_model.model.layers.25.linear_attn.in_proj_z": {
1240
+ "bits": 6,
1241
+ "group_size": 64,
1242
  "mode": "affine"
1243
  },
1244
  "language_model.model.layers.25.linear_attn.out_proj": {
 
1267
  "mode": "affine"
1268
  },
1269
  "language_model.model.layers.26.linear_attn.in_proj_a": {
1270
+ "bits": 6,
1271
+ "group_size": 64,
1272
  "mode": "affine"
1273
  },
1274
  "language_model.model.layers.26.linear_attn.in_proj_b": {
1275
+ "bits": 6,
1276
+ "group_size": 64,
1277
+ "mode": "affine"
1278
+ },
1279
+ "language_model.model.layers.26.linear_attn.in_proj_qkv": {
1280
  "bits": 5,
1281
+ "group_size": 64,
1282
  "mode": "affine"
1283
  },
1284
  "language_model.model.layers.26.linear_attn.in_proj_z": {
1285
+ "bits": 6,
1286
+ "group_size": 64,
1287
  "mode": "affine"
1288
  },
1289
  "language_model.model.layers.26.linear_attn.out_proj": {
 
1331
  "group_size": 64,
1332
  "mode": "affine"
1333
  },
1334
+ "language_model.model.layers.27.self_attn.k_proj": {
1335
  "bits": 5,
1336
+ "group_size": 64,
1337
+ "mode": "affine"
1338
+ },
1339
+ "language_model.model.layers.27.self_attn.o_proj": {
1340
+ "bits": 5,
1341
+ "group_size": 64,
1342
+ "mode": "affine"
1343
+ },
1344
+ "language_model.model.layers.27.self_attn.q_proj": {
1345
+ "bits": 5,
1346
+ "group_size": 64,
1347
+ "mode": "affine"
1348
+ },
1349
+ "language_model.model.layers.27.self_attn.v_proj": {
1350
+ "bits": 5,
1351
+ "group_size": 64,
1352
+ "mode": "affine"
1353
+ },
1354
+ "language_model.model.layers.28.linear_attn.in_proj_a": {
1355
+ "bits": 6,
1356
+ "group_size": 64,
1357
  "mode": "affine"
1358
  },
1359
  "language_model.model.layers.28.linear_attn.in_proj_b": {
1360
+ "bits": 6,
1361
+ "group_size": 64,
1362
+ "mode": "affine"
1363
+ },
1364
+ "language_model.model.layers.28.linear_attn.in_proj_qkv": {
1365
  "bits": 5,
1366
+ "group_size": 64,
1367
  "mode": "affine"
1368
  },
1369
  "language_model.model.layers.28.linear_attn.in_proj_z": {
1370
+ "bits": 6,
1371
+ "group_size": 64,
1372
  "mode": "affine"
1373
  },
1374
  "language_model.model.layers.28.linear_attn.out_proj": {
 
1397
  "mode": "affine"
1398
  },
1399
  "language_model.model.layers.29.linear_attn.in_proj_a": {
1400
+ "bits": 6,
1401
+ "group_size": 64,
1402
  "mode": "affine"
1403
  },
1404
  "language_model.model.layers.29.linear_attn.in_proj_b": {
1405
+ "bits": 6,
1406
+ "group_size": 64,
1407
+ "mode": "affine"
1408
+ },
1409
+ "language_model.model.layers.29.linear_attn.in_proj_qkv": {
1410
  "bits": 5,
1411
+ "group_size": 64,
1412
  "mode": "affine"
1413
  },
1414
  "language_model.model.layers.29.linear_attn.in_proj_z": {
1415
+ "bits": 6,
1416
+ "group_size": 64,
1417
  "mode": "affine"
1418
  },
1419
  "language_model.model.layers.29.linear_attn.out_proj": {
 
1442
  "mode": "affine"
1443
  },
1444
  "language_model.model.layers.30.linear_attn.in_proj_a": {
1445
+ "bits": 6,
1446
+ "group_size": 64,
1447
  "mode": "affine"
1448
  },
1449
  "language_model.model.layers.30.linear_attn.in_proj_b": {
1450
+ "bits": 6,
1451
+ "group_size": 64,
1452
+ "mode": "affine"
1453
+ },
1454
+ "language_model.model.layers.30.linear_attn.in_proj_qkv": {
1455
  "bits": 5,
1456
+ "group_size": 64,
1457
  "mode": "affine"
1458
  },
1459
  "language_model.model.layers.30.linear_attn.in_proj_z": {
1460
+ "bits": 6,
1461
+ "group_size": 64,
1462
  "mode": "affine"
1463
  },
1464
  "language_model.model.layers.30.linear_attn.out_proj": {
 
1507
  "mode": "affine"
1508
  },
1509
  "language_model.model.layers.31.self_attn.k_proj": {
1510
+ "bits": 6,
1511
  "group_size": 64,
1512
  "mode": "affine"
1513
  },
1514
  "language_model.model.layers.31.self_attn.o_proj": {
1515
+ "bits": 5,
1516
  "group_size": 64,
1517
  "mode": "affine"
1518
  },
1519
  "language_model.model.layers.31.self_attn.q_proj": {
1520
+ "bits": 6,
1521
  "group_size": 64,
1522
  "mode": "affine"
1523
  },
 
1536
  "group_size": 64,
1537
  "mode": "affine"
1538
  },
1539
+ "language_model.model.layers.32.linear_attn.in_proj_qkv": {
1540
  "bits": 5,
1541
+ "group_size": 64,
1542
+ "mode": "affine"
1543
+ },
1544
+ "language_model.model.layers.32.linear_attn.in_proj_z": {
1545
+ "bits": 6,
1546
+ "group_size": 64,
1547
  "mode": "affine"
1548
  },
1549
  "language_model.model.layers.32.linear_attn.out_proj": {
 
1572
  "mode": "affine"
1573
  },
1574
  "language_model.model.layers.33.linear_attn.in_proj_a": {
1575
+ "bits": 6,
1576
  "group_size": 64,
1577
  "mode": "affine"
1578
  },
1579
  "language_model.model.layers.33.linear_attn.in_proj_b": {
1580
+ "bits": 6,
1581
  "group_size": 64,
1582
  "mode": "affine"
1583
  },
1584
  "language_model.model.layers.33.linear_attn.in_proj_qkv": {
1585
+ "bits": 6,
1586
+ "group_size": 64,
1587
  "mode": "affine"
1588
  },
1589
  "language_model.model.layers.33.linear_attn.in_proj_z": {
1590
+ "bits": 6,
1591
+ "group_size": 64,
1592
  "mode": "affine"
1593
  },
1594
  "language_model.model.layers.33.linear_attn.out_proj": {
 
1626
  "group_size": 64,
1627
  "mode": "affine"
1628
  },
1629
+ "language_model.model.layers.34.linear_attn.in_proj_qkv": {
1630
  "bits": 5,
1631
+ "group_size": 64,
1632
+ "mode": "affine"
1633
+ },
1634
+ "language_model.model.layers.34.linear_attn.in_proj_z": {
1635
+ "bits": 6,
1636
+ "group_size": 64,
1637
  "mode": "affine"
1638
  },
1639
  "language_model.model.layers.34.linear_attn.out_proj": {
 
1682
  "mode": "affine"
1683
  },
1684
  "language_model.model.layers.35.self_attn.k_proj": {
1685
+ "bits": 6,
1686
  "group_size": 64,
1687
  "mode": "affine"
1688
  },
1689
  "language_model.model.layers.35.self_attn.o_proj": {
1690
+ "bits": 5,
1691
  "group_size": 64,
1692
  "mode": "affine"
1693
  },
1694
  "language_model.model.layers.35.self_attn.q_proj": {
1695
+ "bits": 6,
1696
  "group_size": 64,
1697
  "mode": "affine"
1698
  },
 
1702
  "mode": "affine"
1703
  },
1704
  "language_model.model.layers.36.linear_attn.in_proj_a": {
1705
+ "bits": 6,
1706
  "group_size": 64,
1707
  "mode": "affine"
1708
  },
1709
  "language_model.model.layers.36.linear_attn.in_proj_b": {
1710
+ "bits": 6,
1711
  "group_size": 64,
1712
  "mode": "affine"
1713
  },
 
1717
  "mode": "affine"
1718
  },
1719
  "language_model.model.layers.36.linear_attn.in_proj_z": {
1720
+ "bits": 6,
1721
+ "group_size": 64,
1722
  "mode": "affine"
1723
  },
1724
  "language_model.model.layers.36.linear_attn.out_proj": {
 
1747
  "mode": "affine"
1748
  },
1749
  "language_model.model.layers.37.linear_attn.in_proj_a": {
1750
+ "bits": 6,
1751
  "group_size": 64,
1752
  "mode": "affine"
1753
  },
1754
  "language_model.model.layers.37.linear_attn.in_proj_b": {
1755
+ "bits": 6,
1756
  "group_size": 64,
1757
  "mode": "affine"
1758
  },
1759
  "language_model.model.layers.37.linear_attn.in_proj_qkv": {
1760
+ "bits": 6,
1761
  "group_size": 64,
1762
  "mode": "affine"
1763
  },
1764
  "language_model.model.layers.37.linear_attn.in_proj_z": {
1765
+ "bits": 6,
1766
  "group_size": 64,
1767
  "mode": "affine"
1768
  },
 
1792
  "mode": "affine"
1793
  },
1794
  "language_model.model.layers.38.linear_attn.in_proj_a": {
1795
+ "bits": 6,
1796
  "group_size": 64,
1797
  "mode": "affine"
1798
  },
1799
  "language_model.model.layers.38.linear_attn.in_proj_b": {
1800
+ "bits": 6,
1801
  "group_size": 64,
1802
  "mode": "affine"
1803
  },
1804
  "language_model.model.layers.38.linear_attn.in_proj_qkv": {
1805
+ "bits": 6,
1806
+ "group_size": 64,
1807
  "mode": "affine"
1808
  },
1809
  "language_model.model.layers.38.linear_attn.in_proj_z": {
1810
+ "bits": 6,
1811
+ "group_size": 64,
1812
  "mode": "affine"
1813
  },
1814
  "language_model.model.layers.38.linear_attn.out_proj": {
 
1837
  "mode": "affine"
1838
  },
1839
  "language_model.lm_head": {
1840
+ "bits": 6,
1841
  "group_size": 64,
1842
  "mode": "affine"
1843
  },
 
1862
  "mode": "affine"
1863
  },
1864
  "language_model.model.layers.39.self_attn.k_proj": {
1865
+ "bits": 6,
1866
  "group_size": 64,
1867
  "mode": "affine"
1868
  },
1869
  "language_model.model.layers.39.self_attn.o_proj": {
1870
+ "bits": 5,
1871
  "group_size": 64,
1872
  "mode": "affine"
1873
  },
1874
  "language_model.model.layers.39.self_attn.q_proj": {
1875
+ "bits": 6,
1876
  "group_size": 64,
1877
  "mode": "affine"
1878
  },
 
1892
  "mode": "affine"
1893
  },
1894
  "language_model.model.layers.0.linear_attn.in_proj_a": {
1895
+ "bits": 6,
1896
  "group_size": 64,
1897
  "mode": "affine"
1898
  },
1899
  "language_model.model.layers.0.linear_attn.in_proj_b": {
1900
+ "bits": 6,
1901
  "group_size": 64,
1902
  "mode": "affine"
1903
  },
1904
  "language_model.model.layers.0.linear_attn.in_proj_qkv": {
1905
+ "bits": 6,
1906
+ "group_size": 64,
1907
  "mode": "affine"
1908
  },
1909
  "language_model.model.layers.0.linear_attn.in_proj_z": {
1910
+ "bits": 6,
1911
+ "group_size": 64,
1912
  "mode": "affine"
1913
  },
1914
  "language_model.model.layers.0.linear_attn.out_proj": {
 
1937
  "mode": "affine"
1938
  },
1939
  "language_model.model.layers.1.linear_attn.in_proj_a": {
1940
+ "bits": 6,
1941
+ "group_size": 64,
1942
  "mode": "affine"
1943
  },
1944
  "language_model.model.layers.1.linear_attn.in_proj_b": {
1945
+ "bits": 6,
1946
+ "group_size": 64,
1947
+ "mode": "affine"
1948
+ },
1949
+ "language_model.model.layers.1.linear_attn.in_proj_qkv": {
1950
  "bits": 5,
1951
+ "group_size": 64,
1952
  "mode": "affine"
1953
  },
1954
  "language_model.model.layers.1.linear_attn.in_proj_z": {
1955
+ "bits": 6,
1956
+ "group_size": 64,
1957
  "mode": "affine"
1958
  },
1959
  "language_model.model.layers.1.linear_attn.out_proj": {
 
1982
  "mode": "affine"
1983
  },
1984
  "language_model.model.layers.2.linear_attn.in_proj_a": {
1985
+ "bits": 6,
1986
+ "group_size": 64,
1987
  "mode": "affine"
1988
  },
1989
  "language_model.model.layers.2.linear_attn.in_proj_b": {
1990
+ "bits": 6,
1991
+ "group_size": 64,
1992
+ "mode": "affine"
1993
+ },
1994
+ "language_model.model.layers.2.linear_attn.in_proj_qkv": {
1995
  "bits": 5,
1996
+ "group_size": 64,
1997
  "mode": "affine"
1998
  },
1999
  "language_model.model.layers.2.linear_attn.in_proj_z": {
2000
+ "bits": 6,
2001
+ "group_size": 64,
2002
  "mode": "affine"
2003
  },
2004
  "language_model.model.layers.2.linear_attn.out_proj": {
 
2046
  "group_size": 64,
2047
  "mode": "affine"
2048
  },
2049
+ "language_model.model.layers.3.self_attn.k_proj": {
2050
  "bits": 5,
2051
+ "group_size": 64,
2052
+ "mode": "affine"
2053
+ },
2054
+ "language_model.model.layers.3.self_attn.o_proj": {
2055
+ "bits": 5,
2056
+ "group_size": 64,
2057
+ "mode": "affine"
2058
+ },
2059
+ "language_model.model.layers.3.self_attn.q_proj": {
2060
+ "bits": 5,
2061
+ "group_size": 64,
2062
+ "mode": "affine"
2063
+ },
2064
+ "language_model.model.layers.3.self_attn.v_proj": {
2065
+ "bits": 5,
2066
+ "group_size": 64,
2067
+ "mode": "affine"
2068
+ },
2069
+ "language_model.model.layers.4.linear_attn.in_proj_a": {
2070
+ "bits": 6,
2071
+ "group_size": 64,
2072
  "mode": "affine"
2073
  },
2074
  "language_model.model.layers.4.linear_attn.in_proj_b": {
2075
+ "bits": 6,
2076
+ "group_size": 64,
2077
+ "mode": "affine"
2078
+ },
2079
+ "language_model.model.layers.4.linear_attn.in_proj_qkv": {
2080
  "bits": 5,
2081
+ "group_size": 64,
2082
  "mode": "affine"
2083
  },
2084
  "language_model.model.layers.4.linear_attn.in_proj_z": {
2085
+ "bits": 6,
2086
+ "group_size": 64,
2087
  "mode": "affine"
2088
  },
2089
  "language_model.model.layers.4.linear_attn.out_proj": {
 
2112
  "mode": "affine"
2113
  },
2114
  "language_model.model.layers.5.linear_attn.in_proj_a": {
2115
+ "bits": 6,
2116
+ "group_size": 64,
2117
  "mode": "affine"
2118
  },
2119
  "language_model.model.layers.5.linear_attn.in_proj_b": {
2120
+ "bits": 6,
2121
+ "group_size": 64,
2122
+ "mode": "affine"
2123
+ },
2124
+ "language_model.model.layers.5.linear_attn.in_proj_qkv": {
2125
  "bits": 5,
2126
+ "group_size": 64,
2127
  "mode": "affine"
2128
  },
2129
  "language_model.model.layers.5.linear_attn.in_proj_z": {
2130
+ "bits": 6,
2131
+ "group_size": 64,
2132
  "mode": "affine"
2133
  },
2134
  "language_model.model.layers.5.linear_attn.out_proj": {
 
2157
  "mode": "affine"
2158
  },
2159
  "language_model.model.layers.6.linear_attn.in_proj_a": {
2160
+ "bits": 6,
2161
+ "group_size": 64,
2162
  "mode": "affine"
2163
  },
2164
  "language_model.model.layers.6.linear_attn.in_proj_b": {
2165
+ "bits": 6,
2166
+ "group_size": 64,
2167
+ "mode": "affine"
2168
+ },
2169
+ "language_model.model.layers.6.linear_attn.in_proj_qkv": {
2170
  "bits": 5,
2171
+ "group_size": 64,
2172
  "mode": "affine"
2173
  },
2174
  "language_model.model.layers.6.linear_attn.in_proj_z": {
2175
+ "bits": 6,
2176
+ "group_size": 64,
2177
  "mode": "affine"
2178
  },
2179
  "language_model.model.layers.6.linear_attn.out_proj": {
 
2222
  "mode": "affine"
2223
  },
2224
  "language_model.model.layers.7.self_attn.k_proj": {
2225
+ "bits": 6,
2226
  "group_size": 64,
2227
  "mode": "affine"
2228
  },
2229
+ "language_model.model.layers.7.self_attn.o_proj": {
2230
  "bits": 5,
2231
+ "group_size": 64,
2232
+ "mode": "affine"
2233
+ },
2234
+ "language_model.model.layers.7.self_attn.q_proj": {
2235
+ "bits": 6,
2236
+ "group_size": 64,
2237
  "mode": "affine"
2238
  },
2239
  "language_model.model.layers.7.self_attn.v_proj": {
 
2242
  "mode": "affine"
2243
  },
2244
  "language_model.model.layers.8.linear_attn.in_proj_a": {
2245
+ "bits": 6,
2246
+ "group_size": 64,
2247
  "mode": "affine"
2248
  },
2249
  "language_model.model.layers.8.linear_attn.in_proj_b": {
2250
+ "bits": 6,
2251
+ "group_size": 64,
2252
+ "mode": "affine"
2253
+ },
2254
+ "language_model.model.layers.8.linear_attn.in_proj_qkv": {
2255
  "bits": 5,
2256
+ "group_size": 64,
2257
  "mode": "affine"
2258
  },
2259
  "language_model.model.layers.8.linear_attn.in_proj_z": {
2260
+ "bits": 6,
2261
+ "group_size": 64,
2262
  "mode": "affine"
2263
  },
2264
  "language_model.model.layers.8.linear_attn.out_proj": {
 
2287
  "mode": "affine"
2288
  },
2289
  "language_model.model.layers.9.linear_attn.in_proj_a": {
2290
+ "bits": 6,
2291
+ "group_size": 64,
2292
  "mode": "affine"
2293
  },
2294
  "language_model.model.layers.9.linear_attn.in_proj_b": {
2295
+ "bits": 6,
2296
+ "group_size": 64,
2297
+ "mode": "affine"
2298
+ },
2299
+ "language_model.model.layers.9.linear_attn.in_proj_qkv": {
2300
  "bits": 5,
2301
+ "group_size": 64,
2302
  "mode": "affine"
2303
  },
2304
  "language_model.model.layers.9.linear_attn.in_proj_z": {
2305
+ "bits": 6,
2306
+ "group_size": 64,
2307
  "mode": "affine"
2308
  },
2309
  "language_model.model.layers.9.linear_attn.out_proj": {
 
2332
  "mode": "affine"
2333
  },
2334
  "language_model.model.layers.10.linear_attn.in_proj_a": {
2335
+ "bits": 6,
2336
  "group_size": 64,
2337
  "mode": "affine"
2338
  },
2339
  "language_model.model.layers.10.linear_attn.in_proj_b": {
2340
+ "bits": 6,
2341
  "group_size": 64,
2342
  "mode": "affine"
2343
  },
2344
  "language_model.model.layers.10.linear_attn.in_proj_qkv": {
2345
+ "bits": 6,
2346
+ "group_size": 64,
2347
  "mode": "affine"
2348
  },
2349
  "language_model.model.layers.10.linear_attn.in_proj_z": {
2350
+ "bits": 6,
2351
+ "group_size": 64,
2352
  "mode": "affine"
2353
  },
2354
  "language_model.model.layers.10.linear_attn.out_proj": {
 
2396
  "group_size": 64,
2397
  "mode": "affine"
2398
  },
2399
+ "language_model.model.layers.11.self_attn.k_proj": {
2400
  "bits": 5,
2401
+ "group_size": 64,
2402
+ "mode": "affine"
2403
+ },
2404
+ "language_model.model.layers.11.self_attn.o_proj": {
2405
+ "bits": 5,
2406
+ "group_size": 64,
2407
+ "mode": "affine"
2408
+ },
2409
+ "language_model.model.layers.11.self_attn.q_proj": {
2410
+ "bits": 5,
2411
+ "group_size": 64,
2412
+ "mode": "affine"
2413
+ },
2414
+ "language_model.model.layers.11.self_attn.v_proj": {
2415
+ "bits": 5,
2416
+ "group_size": 64,
2417
+ "mode": "affine"
2418
+ },
2419
+ "language_model.model.layers.12.linear_attn.in_proj_a": {
2420
+ "bits": 6,
2421
+ "group_size": 64,
2422
  "mode": "affine"
2423
  },
2424
  "language_model.model.layers.12.linear_attn.in_proj_b": {
2425
+ "bits": 6,
2426
+ "group_size": 64,
2427
+ "mode": "affine"
2428
+ },
2429
+ "language_model.model.layers.12.linear_attn.in_proj_qkv": {
2430
  "bits": 5,
2431
+ "group_size": 64,
2432
  "mode": "affine"
2433
  },
2434
  "language_model.model.layers.12.linear_attn.in_proj_z": {
2435
+ "bits": 6,
2436
+ "group_size": 64,
2437
  "mode": "affine"
2438
  },
2439
  "language_model.model.layers.12.linear_attn.out_proj": {
 
2502
  "mode": "affine"
2503
  },
2504
  "language_model.model.layers.14.linear_attn.in_proj_a": {
2505
+ "bits": 6,
2506
+ "group_size": 64,
2507
  "mode": "affine"
2508
  },
2509
  "language_model.model.layers.14.linear_attn.in_proj_b": {
2510
+ "bits": 6,
2511
+ "group_size": 64,
2512
+ "mode": "affine"
2513
+ },
2514
+ "language_model.model.layers.14.linear_attn.in_proj_qkv": {
2515
  "bits": 5,
2516
+ "group_size": 64,
2517
  "mode": "affine"
2518
  },
2519
  "language_model.model.layers.14.linear_attn.in_proj_z": {
2520
+ "bits": 6,
2521
+ "group_size": 64,
2522
  "mode": "affine"
2523
  },
2524
  "language_model.model.layers.14.linear_attn.out_proj": {
 
2566
  "group_size": 64,
2567
  "mode": "affine"
2568
  },
2569
+ "language_model.model.layers.15.self_attn.k_proj": {
2570
  "bits": 5,
2571
+ "group_size": 64,
2572
+ "mode": "affine"
2573
+ },
2574
+ "language_model.model.layers.15.self_attn.o_proj": {
2575
+ "bits": 5,
2576
+ "group_size": 64,
2577
+ "mode": "affine"
2578
+ },
2579
+ "language_model.model.layers.15.self_attn.q_proj": {
2580
+ "bits": 5,
2581
+ "group_size": 64,
2582
+ "mode": "affine"
2583
+ },
2584
+ "language_model.model.layers.15.self_attn.v_proj": {
2585
+ "bits": 5,
2586
+ "group_size": 64,
2587
+ "mode": "affine"
2588
+ },
2589
+ "language_model.model.layers.16.linear_attn.in_proj_a": {
2590
+ "bits": 6,
2591
+ "group_size": 64,
2592
  "mode": "affine"
2593
  },
2594
  "language_model.model.layers.16.linear_attn.in_proj_b": {
2595
+ "bits": 6,
2596
+ "group_size": 64,
2597
+ "mode": "affine"
2598
+ },
2599
+ "language_model.model.layers.16.linear_attn.in_proj_qkv": {
2600
  "bits": 5,
2601
+ "group_size": 64,
2602
  "mode": "affine"
2603
  },
2604
  "language_model.model.layers.16.linear_attn.in_proj_z": {
2605
+ "bits": 6,
2606
+ "group_size": 64,
2607
  "mode": "affine"
2608
  },
2609
  "language_model.model.layers.16.linear_attn.out_proj": {
 
2632
  "mode": "affine"
2633
  },
2634
  "language_model.model.layers.17.linear_attn.in_proj_a": {
2635
+ "bits": 6,
2636
+ "group_size": 64,
2637
  "mode": "affine"
2638
  },
2639
  "language_model.model.layers.17.linear_attn.in_proj_b": {
2640
+ "bits": 6,
2641
+ "group_size": 64,
2642
+ "mode": "affine"
2643
+ },
2644
+ "language_model.model.layers.17.linear_attn.in_proj_qkv": {
2645
  "bits": 5,
2646
+ "group_size": 64,
2647
  "mode": "affine"
2648
  },
2649
  "language_model.model.layers.17.linear_attn.in_proj_z": {
2650
+ "bits": 6,
2651
+ "group_size": 64,
2652
  "mode": "affine"
2653
  },
2654
  "language_model.model.layers.17.linear_attn.out_proj": {
 
2677
  "mode": "affine"
2678
  },
2679
  "language_model.model.layers.18.linear_attn.in_proj_a": {
2680
+ "bits": 6,
2681
+ "group_size": 64,
2682
  "mode": "affine"
2683
  },
2684
  "language_model.model.layers.18.linear_attn.in_proj_b": {
2685
+ "bits": 6,
2686
+ "group_size": 64,
2687
+ "mode": "affine"
2688
+ },
2689
+ "language_model.model.layers.18.linear_attn.in_proj_qkv": {
2690
  "bits": 5,
2691
+ "group_size": 64,
2692
  "mode": "affine"
2693
  },
2694
  "language_model.model.layers.18.linear_attn.in_proj_z": {
2695
+ "bits": 6,
2696
+ "group_size": 64,
2697
  "mode": "affine"
2698
  },
2699
  "language_model.model.layers.18.linear_attn.out_proj": {
 
2741
  "group_size": 64,
2742
  "mode": "affine"
2743
  },
2744
+ "language_model.model.layers.19.self_attn.k_proj": {
2745
  "bits": 5,
2746
+ "group_size": 64,
2747
+ "mode": "affine"
2748
+ },
2749
+ "language_model.model.layers.19.self_attn.o_proj": {
2750
+ "bits": 5,
2751
+ "group_size": 64,
2752
+ "mode": "affine"
2753
+ },
2754
+ "language_model.model.layers.19.self_attn.q_proj": {
2755
+ "bits": 5,
2756
+ "group_size": 64,
2757
+ "mode": "affine"
2758
+ },
2759
+ "language_model.model.layers.19.self_attn.v_proj": {
2760
+ "bits": 5,
2761
+ "group_size": 64,
2762
+ "mode": "affine"
2763
+ },
2764
+ "language_model.model.layers.20.linear_attn.in_proj_a": {
2765
+ "bits": 6,
2766
+ "group_size": 64,
2767
  "mode": "affine"
2768
  },
2769
  "language_model.model.layers.20.linear_attn.in_proj_b": {
2770
+ "bits": 6,
2771
+ "group_size": 64,
2772
+ "mode": "affine"
2773
+ },
2774
+ "language_model.model.layers.20.linear_attn.in_proj_qkv": {
2775
  "bits": 5,
2776
+ "group_size": 64,
2777
  "mode": "affine"
2778
  },
2779
  "language_model.model.layers.20.linear_attn.in_proj_z": {
2780
+ "bits": 6,
2781
+ "group_size": 64,
2782
  "mode": "affine"
2783
  },
2784
  "language_model.model.layers.20.linear_attn.out_proj": {
 
2807
  "mode": "affine"
2808
  },
2809
  "language_model.model.layers.21.linear_attn.in_proj_a": {
2810
+ "bits": 6,
2811
+ "group_size": 64,
2812
  "mode": "affine"
2813
  },
2814
  "language_model.model.layers.21.linear_attn.in_proj_b": {
2815
+ "bits": 6,
2816
+ "group_size": 64,
2817
+ "mode": "affine"
2818
+ },
2819
+ "language_model.model.layers.21.linear_attn.in_proj_qkv": {
2820
  "bits": 5,
2821
+ "group_size": 64,
2822
  "mode": "affine"
2823
  },
2824
  "language_model.model.layers.21.linear_attn.in_proj_z": {
2825
+ "bits": 6,
2826
+ "group_size": 64,
2827
  "mode": "affine"
2828
  },
2829
  "language_model.model.layers.21.linear_attn.out_proj": {
 
2852
  "mode": "affine"
2853
  },
2854
  "language_model.model.layers.22.linear_attn.in_proj_a": {
2855
+ "bits": 6,
2856
+ "group_size": 64,
2857
  "mode": "affine"
2858
  },
2859
  "language_model.model.layers.22.linear_attn.in_proj_b": {
2860
+ "bits": 6,
2861
+ "group_size": 64,
2862
+ "mode": "affine"
2863
+ },
2864
+ "language_model.model.layers.22.linear_attn.in_proj_qkv": {
2865
  "bits": 5,
2866
+ "group_size": 64,
2867
  "mode": "affine"
2868
  },
2869
  "language_model.model.layers.22.linear_attn.in_proj_z": {
2870
+ "bits": 6,
2871
+ "group_size": 64,
2872
  "mode": "affine"
2873
  },
2874
  "language_model.model.layers.22.linear_attn.out_proj": {
 
2916
  "group_size": 64,
2917
  "mode": "affine"
2918
  },
2919
+ "language_model.model.layers.23.self_attn.k_proj": {
2920
  "bits": 5,
2921
+ "group_size": 64,
2922
+ "mode": "affine"
2923
+ },
2924
+ "language_model.model.layers.23.self_attn.o_proj": {
2925
+ "bits": 5,
2926
+ "group_size": 64,
2927
+ "mode": "affine"
2928
+ },
2929
+ "language_model.model.layers.23.self_attn.q_proj": {
2930
+ "bits": 5,
2931
+ "group_size": 64,
2932
+ "mode": "affine"
2933
+ },
2934
+ "language_model.model.layers.23.self_attn.v_proj": {
2935
+ "bits": 5,
2936
+ "group_size": 64,
2937
+ "mode": "affine"
2938
+ },
2939
+ "language_model.model.layers.24.linear_attn.in_proj_a": {
2940
+ "bits": 6,
2941
+ "group_size": 64,
2942
  "mode": "affine"
2943
  },
2944
  "language_model.model.layers.24.linear_attn.in_proj_b": {
2945
+ "bits": 6,
2946
+ "group_size": 64,
2947
+ "mode": "affine"
2948
+ },
2949
+ "language_model.model.layers.24.linear_attn.in_proj_qkv": {
2950
  "bits": 5,
2951
+ "group_size": 64,
2952
  "mode": "affine"
2953
  },
2954
  "language_model.model.layers.24.linear_attn.in_proj_z": {
2955
+ "bits": 6,
2956
+ "group_size": 64,
2957
  "mode": "affine"
2958
  },
2959
  "language_model.model.layers.24.linear_attn.out_proj": {
 
2982
  "mode": "affine"
2983
  },
2984
  "language_model.model.layers.25.linear_attn.in_proj_a": {
2985
+ "bits": 6,
2986
+ "group_size": 64,
2987
  "mode": "affine"
2988
  },
2989
  "language_model.model.layers.25.linear_attn.in_proj_b": {
2990
+ "bits": 6,
2991
+ "group_size": 64,
2992
+ "mode": "affine"
2993
+ },
2994
+ "language_model.model.layers.25.linear_attn.in_proj_qkv": {
2995
  "bits": 5,
2996
+ "group_size": 64,
2997
  "mode": "affine"
2998
  },
2999
  "language_model.model.layers.25.linear_attn.in_proj_z": {
3000
+ "bits": 6,
3001
+ "group_size": 64,
3002
  "mode": "affine"
3003
  },
3004
  "language_model.model.layers.25.linear_attn.out_proj": {
 
3027
  "mode": "affine"
3028
  },
3029
  "language_model.model.layers.26.linear_attn.in_proj_a": {
3030
+ "bits": 6,
3031
+ "group_size": 64,
3032
  "mode": "affine"
3033
  },
3034
  "language_model.model.layers.26.linear_attn.in_proj_b": {
3035
+ "bits": 6,
3036
+ "group_size": 64,
3037
+ "mode": "affine"
3038
+ },
3039
+ "language_model.model.layers.26.linear_attn.in_proj_qkv": {
3040
  "bits": 5,
3041
+ "group_size": 64,
3042
  "mode": "affine"
3043
  },
3044
  "language_model.model.layers.26.linear_attn.in_proj_z": {
3045
+ "bits": 6,
3046
+ "group_size": 64,
3047
  "mode": "affine"
3048
  },
3049
  "language_model.model.layers.26.linear_attn.out_proj": {
 
3091
  "group_size": 64,
3092
  "mode": "affine"
3093
  },
3094
+ "language_model.model.layers.27.self_attn.k_proj": {
3095
  "bits": 5,
3096
+ "group_size": 64,
3097
+ "mode": "affine"
3098
+ },
3099
+ "language_model.model.layers.27.self_attn.o_proj": {
3100
+ "bits": 5,
3101
+ "group_size": 64,
3102
+ "mode": "affine"
3103
+ },
3104
+ "language_model.model.layers.27.self_attn.q_proj": {
3105
+ "bits": 5,
3106
+ "group_size": 64,
3107
+ "mode": "affine"
3108
+ },
3109
+ "language_model.model.layers.27.self_attn.v_proj": {
3110
+ "bits": 5,
3111
+ "group_size": 64,
3112
+ "mode": "affine"
3113
+ },
3114
+ "language_model.model.layers.28.linear_attn.in_proj_a": {
3115
+ "bits": 6,
3116
+ "group_size": 64,
3117
  "mode": "affine"
3118
  },
3119
  "language_model.model.layers.28.linear_attn.in_proj_b": {
3120
+ "bits": 6,
3121
+ "group_size": 64,
3122
+ "mode": "affine"
3123
+ },
3124
+ "language_model.model.layers.28.linear_attn.in_proj_qkv": {
3125
  "bits": 5,
3126
+ "group_size": 64,
3127
  "mode": "affine"
3128
  },
3129
  "language_model.model.layers.28.linear_attn.in_proj_z": {
3130
+ "bits": 6,
3131
+ "group_size": 64,
3132
  "mode": "affine"
3133
  },
3134
  "language_model.model.layers.28.linear_attn.out_proj": {
 
3157
  "mode": "affine"
3158
  },
3159
  "language_model.model.layers.29.linear_attn.in_proj_a": {
3160
+ "bits": 6,
3161
+ "group_size": 64,
3162
  "mode": "affine"
3163
  },
3164
  "language_model.model.layers.29.linear_attn.in_proj_b": {
3165
+ "bits": 6,
3166
+ "group_size": 64,
3167
+ "mode": "affine"
3168
+ },
3169
+ "language_model.model.layers.29.linear_attn.in_proj_qkv": {
3170
  "bits": 5,
3171
+ "group_size": 64,
3172
  "mode": "affine"
3173
  },
3174
  "language_model.model.layers.29.linear_attn.in_proj_z": {
3175
+ "bits": 6,
3176
+ "group_size": 64,
3177
  "mode": "affine"
3178
  },
3179
  "language_model.model.layers.29.linear_attn.out_proj": {
 
3202
  "mode": "affine"
3203
  },
3204
  "language_model.model.layers.30.linear_attn.in_proj_a": {
3205
+ "bits": 6,
3206
+ "group_size": 64,
3207
  "mode": "affine"
3208
  },
3209
  "language_model.model.layers.30.linear_attn.in_proj_b": {
3210
+ "bits": 6,
3211
+ "group_size": 64,
3212
+ "mode": "affine"
3213
+ },
3214
+ "language_model.model.layers.30.linear_attn.in_proj_qkv": {
3215
  "bits": 5,
3216
+ "group_size": 64,
3217
  "mode": "affine"
3218
  },
3219
  "language_model.model.layers.30.linear_attn.in_proj_z": {
3220
+ "bits": 6,
3221
+ "group_size": 64,
3222
  "mode": "affine"
3223
  },
3224
  "language_model.model.layers.30.linear_attn.out_proj": {
 
3267
  "mode": "affine"
3268
  },
3269
  "language_model.model.layers.31.self_attn.k_proj": {
3270
+ "bits": 6,
3271
  "group_size": 64,
3272
  "mode": "affine"
3273
  },
3274
  "language_model.model.layers.31.self_attn.o_proj": {
3275
+ "bits": 5,
3276
  "group_size": 64,
3277
  "mode": "affine"
3278
  },
3279
  "language_model.model.layers.31.self_attn.q_proj": {
3280
+ "bits": 6,
3281
  "group_size": 64,
3282
  "mode": "affine"
3283
  },
 
3296
  "group_size": 64,
3297
  "mode": "affine"
3298
  },
3299
+ "language_model.model.layers.32.linear_attn.in_proj_qkv": {
3300
  "bits": 5,
3301
+ "group_size": 64,
3302
+ "mode": "affine"
3303
+ },
3304
+ "language_model.model.layers.32.linear_attn.in_proj_z": {
3305
+ "bits": 6,
3306
+ "group_size": 64,
3307
  "mode": "affine"
3308
  },
3309
  "language_model.model.layers.32.linear_attn.out_proj": {
 
3332
  "mode": "affine"
3333
  },
3334
  "language_model.model.layers.33.linear_attn.in_proj_a": {
3335
+ "bits": 6,
3336
  "group_size": 64,
3337
  "mode": "affine"
3338
  },
3339
  "language_model.model.layers.33.linear_attn.in_proj_b": {
3340
+ "bits": 6,
3341
  "group_size": 64,
3342
  "mode": "affine"
3343
  },
3344
  "language_model.model.layers.33.linear_attn.in_proj_qkv": {
3345
+ "bits": 6,
3346
+ "group_size": 64,
3347
  "mode": "affine"
3348
  },
3349
  "language_model.model.layers.33.linear_attn.in_proj_z": {
3350
+ "bits": 6,
3351
+ "group_size": 64,
3352
  "mode": "affine"
3353
  },
3354
  "language_model.model.layers.33.linear_attn.out_proj": {
 
3386
  "group_size": 64,
3387
  "mode": "affine"
3388
  },
3389
+ "language_model.model.layers.34.linear_attn.in_proj_qkv": {
3390
  "bits": 5,
3391
+ "group_size": 64,
3392
+ "mode": "affine"
3393
+ },
3394
+ "language_model.model.layers.34.linear_attn.in_proj_z": {
3395
+ "bits": 6,
3396
+ "group_size": 64,
3397
  "mode": "affine"
3398
  },
3399
  "language_model.model.layers.34.linear_attn.out_proj": {
 
3442
  "mode": "affine"
3443
  },
3444
  "language_model.model.layers.35.self_attn.k_proj": {
3445
+ "bits": 6,
3446
  "group_size": 64,
3447
  "mode": "affine"
3448
  },
3449
  "language_model.model.layers.35.self_attn.o_proj": {
3450
+ "bits": 5,
3451
  "group_size": 64,
3452
  "mode": "affine"
3453
  },
3454
  "language_model.model.layers.35.self_attn.q_proj": {
3455
+ "bits": 6,
3456
  "group_size": 64,
3457
  "mode": "affine"
3458
  },
 
3462
  "mode": "affine"
3463
  },
3464
  "language_model.model.layers.36.linear_attn.in_proj_a": {
3465
+ "bits": 6,
3466
  "group_size": 64,
3467
  "mode": "affine"
3468
  },
3469
  "language_model.model.layers.36.linear_attn.in_proj_b": {
3470
+ "bits": 6,
3471
  "group_size": 64,
3472
  "mode": "affine"
3473
  },
 
3477
  "mode": "affine"
3478
  },
3479
  "language_model.model.layers.36.linear_attn.in_proj_z": {
3480
+ "bits": 6,
3481
+ "group_size": 64,
3482
  "mode": "affine"
3483
  },
3484
  "language_model.model.layers.36.linear_attn.out_proj": {
 
3507
  "mode": "affine"
3508
  },
3509
  "language_model.model.layers.37.linear_attn.in_proj_a": {
3510
+ "bits": 6,
3511
  "group_size": 64,
3512
  "mode": "affine"
3513
  },
3514
  "language_model.model.layers.37.linear_attn.in_proj_b": {
3515
+ "bits": 6,
3516
  "group_size": 64,
3517
  "mode": "affine"
3518
  },
3519
  "language_model.model.layers.37.linear_attn.in_proj_qkv": {
3520
+ "bits": 6,
3521
  "group_size": 64,
3522
  "mode": "affine"
3523
  },
3524
  "language_model.model.layers.37.linear_attn.in_proj_z": {
3525
+ "bits": 6,
3526
  "group_size": 64,
3527
  "mode": "affine"
3528
  },
 
3552
  "mode": "affine"
3553
  },
3554
  "language_model.model.layers.38.linear_attn.in_proj_a": {
3555
+ "bits": 6,
3556
  "group_size": 64,
3557
  "mode": "affine"
3558
  },
3559
  "language_model.model.layers.38.linear_attn.in_proj_b": {
3560
+ "bits": 6,
3561
  "group_size": 64,
3562
  "mode": "affine"
3563
  },
3564
  "language_model.model.layers.38.linear_attn.in_proj_qkv": {
3565
+ "bits": 6,
3566
+ "group_size": 64,
3567
  "mode": "affine"
3568
  },
3569
  "language_model.model.layers.38.linear_attn.in_proj_z": {
3570
+ "bits": 6,
3571
+ "group_size": 64,
3572
  "mode": "affine"
3573
  },
3574
  "language_model.model.layers.38.linear_attn.out_proj": {
 
3597
  "mode": "affine"
3598
  },
3599
  "language_model.lm_head": {
3600
+ "bits": 6,
3601
  "group_size": 64,
3602
  "mode": "affine"
3603
  },
 
3622
  "mode": "affine"
3623
  },
3624
  "language_model.model.layers.39.self_attn.k_proj": {
3625
+ "bits": 6,
3626
  "group_size": 64,
3627
  "mode": "affine"
3628
  },
3629
  "language_model.model.layers.39.self_attn.o_proj": {
3630
+ "bits": 5,
3631
  "group_size": 64,
3632
  "mode": "affine"
3633
  },
3634
  "language_model.model.layers.39.self_attn.q_proj": {
3635
+ "bits": 6,
3636
  "group_size": 64,
3637
  "mode": "affine"
3638
  },
model-00001-of-00005.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a85cfa9cf3a90254e3ad129543d3b0598d7c32bc23f5863afc11639508982b63
3
- size 5097602024
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b25dd783d843b29d252baae57adf4b41280f5c12f197a3cd6d522be617efe413
3
+ size 5060873189
model-00002-of-00005.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ccf6b67d3ddbac7cbe479fdb3d393340397e081f14112b5401b8c15e2f038f7a
3
  size 5133842458
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4d1d8873569f072f4b9829ca87521d41a5e4d3eb556bc5088652a02e781ce17
3
  size 5133842458
model-00003-of-00005.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9679f07dfdfc7497e184ee6059e360700905cbbec0801f08055402233a46b0cf
3
  size 5133842508
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2cc037cdc77468f08d1ca4a06e519de18fa0ab450b46d1f13d62f57c4173e821
3
  size 5133842508
model-00004-of-00005.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b8ef5e73275fd0cef1eb14948198125e66969e73e6a6cdcf9ad9abebfc252f29
3
  size 5133842502
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f855fcaba0521326f869ae4217bbece3ea4a9c698f303c5c3ba177d6fcc0e436
3
  size 5133842502
model-00005-of-00005.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:822736cf4e61adefef32f83f0c9a7a15d0ba8a55c4c51182c093eaebaf96ff34
3
  size 603981479
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a3ea673928bce7c61a97d6242b67f25cd59f4111ce9285bff8cc9021918e7fee
3
  size 603981479
model.safetensors.index.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "metadata": {
3
- "total_size": 21103110971
4
  },
5
  "weight_map": {
6
  "language_model.lm_head.biases": "model-00001-of-00005.safetensors",
 
1
  {
2
  "metadata": {
3
+ "total_size": 21066382136
4
  },
5
  "weight_map": {
6
  "language_model.lm_head.biases": "model-00001-of-00005.safetensors",
oq_imatrix_report.json CHANGED
@@ -1,28 +1,28 @@
1
  {
2
  "enabled": true,
3
- "cache_path": "/Volumes/Vault/ai/models/mlx/.oqe_imatrix/Ornith-1.5-35B-A3B-hf-ef711fabccc8-s128-l512.npz",
4
- "cache_reused": true,
5
- "entry_count": 523,
6
  "calib_dataset": "oqe_code_multilingual",
7
  "collection": {
8
  "dataset": "oqe_code_multilingual",
9
- "requested_samples": 128,
10
  "seq_length": 512,
11
  "adaptive": true,
12
- "adaptive_step_samples": 128,
13
- "adaptive_max_samples": 1024,
14
- "available_samples": 1024,
15
- "micro_batch_size": 8,
16
- "micro_batches": 16,
17
  "batch_plan": {
18
- "micro_batch_size": 8,
19
  "estimated_sample_bytes": 33554432,
20
- "capture_budget_bytes": 294268370,
21
- "system_available_bytes": 50324537344,
22
- "metal_available_bytes": 60119750106,
23
- "live_available_bytes": 50324537344,
24
- "model_bytes": 20897700290,
25
- "remaining_available_bytes": 29426837054,
26
  "fits_one_sample": true,
27
  "hidden_size": 2048,
28
  "num_experts": 256,
@@ -32,57 +32,57 @@
32
  "num_kv_shared_layers": 0,
33
  "per_layer_input_size": 0
34
  },
35
- "processed_samples": 128,
36
  "model_class": "mlx_vlm.models.qwen3_5_moe.qwen3_5_moe.Model",
37
  "layer_model_class": "mlx_vlm.models.qwen3_5_moe.language.Qwen3_5MoeModel",
38
  "layer_model_type": "qwen3_5_moe_text",
39
  "layer_count": 40,
40
  "layer_state_kind": "generic",
41
- "successful_layer_forwards": 640,
42
  "skipped_layer_forwards": 0,
43
- "mtp_head_forwards": 16,
44
- "installed_modules": 634,
45
  "capture_module_classes": {
46
- "Linear": 151,
47
- "QuantizedLinear": 360,
48
- "QuantizedSwitchLinear": 123
49
  },
50
- "switch_capture_modules": 123,
51
  "requires_expert_counts": true,
52
  "coverage_sufficient": true,
53
  "collection_sufficient": true,
54
  "coverage": {
55
  "has_expert_counts": true,
56
- "expert_modules": 123,
57
- "total_experts": 31488,
58
- "active_experts": 31488,
59
  "zero_count_experts": 0,
60
  "active_ratio": 1.0,
61
- "min_count": 1,
62
- "p05_count": 390.0,
63
- "p10_count": 555.0,
64
- "median_count": 1520.0,
65
- "max_count": 41467,
66
  "min_required_count": 16,
67
  "required_percentile": 5
68
  },
69
  "rounds": [
70
  {
71
- "processed_samples": 128,
72
  "coverage_sufficient": true,
73
  "collection_sufficient": true,
74
  "coverage": {
75
  "has_expert_counts": true,
76
- "expert_modules": 123,
77
- "total_experts": 31488,
78
- "active_experts": 31488,
79
  "zero_count_experts": 0,
80
  "active_ratio": 1.0,
81
- "min_count": 1,
82
- "p05_count": 390.0,
83
- "p10_count": 555.0,
84
- "median_count": 1520.0,
85
- "max_count": 41467,
86
  "min_required_count": 16,
87
  "required_percentile": 5
88
  }
@@ -91,16 +91,16 @@
91
  },
92
  "expert_coverage": {
93
  "has_expert_counts": true,
94
- "expert_modules": 123,
95
- "total_experts": 31488,
96
- "active_experts": 31488,
97
  "zero_count_experts": 0,
98
  "active_ratio": 1.0,
99
- "min_count": 1,
100
- "p05_count": 390.0,
101
- "p10_count": 555.0,
102
- "median_count": 1520.0,
103
- "max_count": 41467,
104
  "min_required_count": 16,
105
  "required_percentile": 5
106
  },
 
1
  {
2
  "enabled": true,
3
+ "cache_path": "Ornith-1.5-35B-A3B-hf-ef711fabccc8-s3000-l512.npz",
4
+ "cache_reused": false,
5
+ "entry_count": 510,
6
  "calib_dataset": "oqe_code_multilingual",
7
  "collection": {
8
  "dataset": "oqe_code_multilingual",
9
+ "requested_samples": 3000,
10
  "seq_length": 512,
11
  "adaptive": true,
12
+ "adaptive_step_samples": 3000,
13
+ "adaptive_max_samples": 3000,
14
+ "available_samples": 3000,
15
+ "micro_batch_size": 6,
16
+ "micro_batches": 500,
17
  "batch_plan": {
18
+ "micro_batch_size": 6,
19
  "estimated_sample_bytes": 33554432,
20
+ "capture_budget_bytes": 231565939,
21
+ "system_available_bytes": 43578408960,
22
+ "metal_available_bytes": 60123360348,
23
+ "live_available_bytes": 43578408960,
24
+ "model_bytes": 20421815039,
25
+ "remaining_available_bytes": 23156593921,
26
  "fits_one_sample": true,
27
  "hidden_size": 2048,
28
  "num_experts": 256,
 
32
  "num_kv_shared_layers": 0,
33
  "per_layer_input_size": 0
34
  },
35
+ "processed_samples": 3000,
36
  "model_class": "mlx_vlm.models.qwen3_5_moe.qwen3_5_moe.Model",
37
  "layer_model_class": "mlx_vlm.models.qwen3_5_moe.language.Qwen3_5MoeModel",
38
  "layer_model_type": "qwen3_5_moe_text",
39
  "layer_count": 40,
40
  "layer_state_kind": "generic",
41
+ "successful_layer_forwards": 20000,
42
  "skipped_layer_forwards": 0,
43
+ "mtp_head_forwards": 0,
44
+ "installed_modules": 621,
45
  "capture_module_classes": {
46
+ "Linear": 150,
47
+ "QuantizedLinear": 351,
48
+ "QuantizedSwitchLinear": 120
49
  },
50
+ "switch_capture_modules": 120,
51
  "requires_expert_counts": true,
52
  "coverage_sufficient": true,
53
  "collection_sufficient": true,
54
  "coverage": {
55
  "has_expert_counts": true,
56
+ "expert_modules": 120,
57
+ "total_experts": 30720,
58
+ "active_experts": 30720,
59
  "zero_count_experts": 0,
60
  "active_ratio": 1.0,
61
+ "min_count": 7,
62
+ "p05_count": 8999.35,
63
+ "p10_count": 12601.8,
64
+ "median_count": 35289.0,
65
+ "max_count": 915741,
66
  "min_required_count": 16,
67
  "required_percentile": 5
68
  },
69
  "rounds": [
70
  {
71
+ "processed_samples": 3000,
72
  "coverage_sufficient": true,
73
  "collection_sufficient": true,
74
  "coverage": {
75
  "has_expert_counts": true,
76
+ "expert_modules": 120,
77
+ "total_experts": 30720,
78
+ "active_experts": 30720,
79
  "zero_count_experts": 0,
80
  "active_ratio": 1.0,
81
+ "min_count": 7,
82
+ "p05_count": 8999.35,
83
+ "p10_count": 12601.8,
84
+ "median_count": 35289.0,
85
+ "max_count": 915741,
86
  "min_required_count": 16,
87
  "required_percentile": 5
88
  }
 
91
  },
92
  "expert_coverage": {
93
  "has_expert_counts": true,
94
+ "expert_modules": 120,
95
+ "total_experts": 30720,
96
+ "active_experts": 30720,
97
  "zero_count_experts": 0,
98
  "active_ratio": 1.0,
99
+ "min_count": 7,
100
+ "p05_count": 8999.35,
101
+ "p10_count": 12601.8,
102
+ "median_count": 35289.0,
103
+ "max_count": 915741,
104
  "min_required_count": 16,
105
  "required_percentile": 5
106
  },