← Back to Quantization

QAT

Quantization
Used in
123 PRs
Best BPB
0.1653
Avg BPB
1.2017

Submissions

PR #30by JackYoung27
1.2663
PR #79by Marvbuster
1.8698
PR #117by trovatochris
1.1702
PR #126by Athenox14
1.7510
PR #130by mohosy
1.6372
PR #145by mrdavtan
1.2052
PR #160by ChaseWNorton
1.1623
PR #185by dttdrv
1.3043
PR #193by KHUCHAN
1.2917
PR #218by bopmite
1.1248
PR #237by takoyakisoft
1.8389
PR #263by Dannybc123
1.5382
PR #286by chris-buckley
1.1628
PR #292by xuafeng
1.3274
PR #292by xuafeng
1.3274
PR #316by SkywardSyntax
1.2035
PR #345by anandks2006
1.8522
PR #356by sjp611
1.8338
PR #375by charmquark1984
1.1257
PR #397by translatingthename
1.1364
PR #400by chanwoo-park-official
1.1296
PR #410by EthanYangTW
1.1216
PR #414by signalrush
1.1233
PR #415by EthanYangTW
1.1216
PR #415by EthanYangTW
1.1216
PR #429by AbhisekBasu1
1.1231
PR #432by jadechip
1.5295
PR #457by carlesonielfa
1.1839
PR #469by cmcdnd
1.1418
PR #478by gowtham0992
1.1268
PR #483by tmustier
1.1346
PR #503by EthanYangTW
1.1195
PR #508by newjordan
1.1215
PR #527by Shuvam-Banerji-Seal
1.4078
PR #528by EthanYangTW
1.1195
PR #529by EthanYangTW
1.1195
PR #534by rarce
1.1804
PR #545by EthanYangTW
1.1179
PR #589by RoyiRa
1.1178
PR #607by Neopolita
1.4750
PR #617by ryanadamsai
1.1228
PR #674by newjordan
1.0461
PR #680by bro4all
1.1483
PR #681by Alfaxad
1.4775
PR #692by EthanYangTW
1.1186
PR #707by Shuvam-Banerji-Seal
1.4078
PR #712by Shuvam-Banerji-Seal
1.4078
PR #714by Upsalla
1.1187
PR #719by Shuvam-Banerji-Seal
1.5252
PR #725by Shuvam-Banerji-Seal
1.5252
PR #730by janwww
1.1570
PR #751by Shuvam-Banerji-Seal
1.5252
PR #785by SirSaltySalmon
1.5364
PR #790by danialht
1.1172
PR #836by autocode-rayes
1.1219
PR #852by Prush69
1.1189
PR #872by gowtham0992
1.0467
PR #882by IshiPareek
1.3762
PR #885by lolrazh
0.9958
PR #890by sofiabod
0.4405
PR #891by robbiebusinessacc
1.1428
PR #903by CiprianFlorin-Ifrim
1.2064
PR #918by haikosys
0.1653
PR #918by haikosys
0.1653
PR #918by haikosys
0.1653
PR #920by CiprianFlorin-Ifrim
1.1539
PR #923by CiprianFlorin-Ifrim
1.1090
PR #928by autocode-rayes
1.1211
PR #952by FlashyFlash3011
1.1144
PR #972by Idan3011
0.3922
PR #975by Abhishek8108
1.1216
PR #989by alexanderaperry-arch
1.1402
PR #996by Idan3011
1.1478
PR #1002by SoHarshh
1.1650
PR #1007by dillon-blake
1.2252
PR #1009by SoHarshh
1.1574
PR #1037by TimPietruskyRunPod
1.1179
PR #1040by JoeProAI
1.1336
PR #1072by vimeto
1.1170
PR #1081by michaelwinczuk
1.1220
PR #1088by serdardoesml
1.2542
PR #1096by vimeto
1.3342
PR #1096by vimeto
1.3342
PR #1107by mradassaad
1.5633
PR #1108by DbBested
1.1502
PR #1111by MichaelMcCulloch
0.2532
PR #1140by newjordan
1.1874
PR #1172by dexhunter
1.1015
PR #1179by dexhunter
1.1105
PR #1200by Mister2005
1.6768
PR #1201by Mister2005
1.6371
PR #1227by himanshudongre
1.4841
PR #1246by deborahnelson8788726
0.9650
PR #1254by Elarwei001
1.1070
PR #1256by oidebrett
1.1444
PR #1260by dexhunter
1.0929
PR #1270by VirajDeshwal
1.1088
PR #1273by DushyantChetiwal
1.2196
PR #1275by ranausmanai
1.1492
PR #1289by MatoTeziTanka
1.0819
PR #1302by vlivashkin
1.1078
PR #1305by DariusFeher
1.2070
PR #1318by renqianluo
1.0095
PR #1346by shasank0001
1.2283
PR #1347by shasank0001
1.3038
PR #1357by mollahasani
1.2200
PR #1366by yunoshev
1.1371
PR #1417by BruhTheMomentum
1.3039
PR #1418by Park-Tae-Hwan
1.4192
PR #1431by Idan3011
1.1266
PR #1433by mtybadger
1.2067
PR #1436by DevWizard-Vandan
1.5546
PR #1574by KRGulaj
1.3587
PR #1579by Tonyy1977
1.1372
PR #1602by SPThole
1.0744
PR #1620by shiawyonglim
1.6644
PR #1627by mike-ferguson
1.3246
PR #1630by KevinChunye
1.1412
PR #1683by yunoshev
1.1280
PR #1683by yunoshev
1.1280
PR #1718by himanshudongre
1.0788
PR #1757by aiejvn
1.5194
PR #1760by BrandtChristian
1.1863

Hyperparameters Across PRs

pr_numberbitsscope
30post-quantization gap
796all
117weights
1262all weights
1308large matrices (>65K params)
1458per-row weights
160submission artifact / timed run support, but not activated before stop
1858model weights
1938export-matched int8 path
218all
237all
2638weights during training
286final phase only
2925MLP layers
2926attention layers
3167all
3458all
3566all
3754full-run
3976all
4006mlp, attn
4106attention; int5 for MLP layers
4146model weights
4156attention
4155MLP
429all
432attention
4578all
4695all
4786all
4836all
5036all
5086weights
5276
5286all
5296all
534late training
5455all weights
5896all
6078per-row
617all
6746all
6806all
6816all
6926all weights
7076all
7126all
7146all
7196all
7256all
7308fp_params / model artifact
751all
785late QAT
790mixed int6 GPTQ with early QAT
8366all
8524all
872all
882all
885all
890all
8916int6
9034large weights
9182MLP up
9183attn/MLP down
9184embeddings
9208FP8 path / model artifact
923all
9286all
9526all
9726all
9756all
9896all
9966all
10024MLP and bigram; INT6 attention
1007all
10094MLP + bigram
1037all
10405all
1072all
1081all
10888model weights
10966shared block weights
10966shared weights
11076Mamba projections and standard CastedLinear layers
11086all
11116all
11408model
1172late
1179all
12006Q/K/V/O and MLP up/down bank slices
12016bank slices (Q/K/V/O and MLP up/down)
12275all
1246all large weight matrices
12546all
12566all
1260all
1270model
12731.58all
12758all
1289model weights
13026all
1305all
1318all
1346MLP
1347MLP
13576all
1366MLP int5, attention int6
1417all weights
14184model weights
14315MLP layers
1433codebook weights
14366all
15746weights
15796large weight matrices
1602all
16206all
16278training export
1630all
16834MLP
16835attention
17186matrices only
17576all
17606all layers