transformers
DeepGEMM BF16 + mixed FP8/FP4 + MegaMoE + refactor
#45634
Merged

DeepGEMM BF16 + mixed FP8/FP4 + MegaMoE + refactor #45634

IlyasMoutawwakil merged 164 commits into main from deepgemm-isolation
IlyasMoutawwakil
IlyasMoutawwakil init
94a2786b
IlyasMoutawwakil Merge branch 'main' into deepgemm
8fc43d72
IlyasMoutawwakil style
357a0355
IlyasMoutawwakil full support
741b5eb7
IlyasMoutawwakil support EP better using offsets !
9fc3662d
IlyasMoutawwakil comments
84552ae9
IlyasMoutawwakil get rid of neutralize_ep_sentinels
1d9f319b
IlyasMoutawwakil remove deepgemm stuff
9b860434
IlyasMoutawwakil Merge branch 'main' into deepgemm
8a3f88bb
IlyasMoutawwakil fix
996d67d0
IlyasMoutawwakil prefix
d033a830
IlyasMoutawwakil move
e15cfe6a
IlyasMoutawwakil fix
10b6d904
IlyasMoutawwakil remove comment
d4a6b305
IlyasMoutawwakil fix unintilized outputs leaking
1d6054ff
IlyasMoutawwakil revert unnecessary changes
137393cd
IlyasMoutawwakil more unnecessary changes
774f9018
IlyasMoutawwakil revert downcast
81230fee
IlyasMoutawwakil keep it simple
9f2ff089
IlyasMoutawwakil guard deepgemm cuda version
c55b7b78
IlyasMoutawwakil fix style
20858db8
IlyasMoutawwakil update
bfea94f0
IlyasMoutawwakil IlyasMoutawwakil changed the title DeepGEMM refactor DeepGEMM BF16, isolation, refactor 138 days ago
IlyasMoutawwakil add deepgemm testing
eada47e1
HuggingFaceDocBuilderDev
IlyasMoutawwakil moe sentinel support
89d2f0bb
IlyasMoutawwakil Merge branch 'main' into deepgemm
ef72b0e4
IlyasMoutawwakil fix
60db1ca0
IlyasMoutawwakil Merge branch 'main' into deepgemm
e732af1d
IlyasMoutawwakil compilable sonicmoe
68b7b0fe
IlyasMoutawwakil mega moe kernel support attempt
faaa7aaa
IlyasMoutawwakil use package for now
b502cd62
IlyasMoutawwakil IlyasMoutawwakil changed the title DeepGEMM BF16, isolation, refactor DeepGEMM BF16 + mixed FP8/FP4 + MegaMoE + refactor 133 days ago
IlyasMoutawwakil Merge branch 'deepgemm' into deepgemm-isolation
cc753ca5
IlyasMoutawwakil skip ep router and experts pre/post processing
1c174528
IlyasMoutawwakil simpler
5a8ceaea
IlyasMoutawwakil fix
a05aa39b
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
854f4fb2
IlyasMoutawwakil fix
053c9df6
IlyasMoutawwakil fix
80a6fe5a
IlyasMoutawwakil dtensor support
ad8226ce
IlyasMoutawwakil more dtensor
a663f4d7
IlyasMoutawwakil simpler
74c3f2e3
IlyasMoutawwakil remove comment
d3cae33e
IlyasMoutawwakil Merge branch 'deepgemm' into deepgemm-isolation
c1bfa0f0
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
6c436114
IlyasMoutawwakil revert
0528e0e2
IlyasMoutawwakil bc order
2bfd0298
IlyasMoutawwakil revert extra indent
368db007
IlyasMoutawwakil revert unnecessary change
28fbb842
IlyasMoutawwakil update
6ef27abe
IlyasMoutawwakil less defensive
e436b774
IlyasMoutawwakil allow all kernels
8e8f0ee2
IlyasMoutawwakil alow all kernels
c494e35b
IlyasMoutawwakil hub only
03b04421
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
527d8b47
IlyasMoutawwakil fix
2e51b3cb
IlyasMoutawwakil fix
82c5fb59
IlyasMoutawwakil test
99fdf71d
IlyasMoutawwakil test
21398a8b
IlyasMoutawwakil sync
f21db7cd
IlyasMoutawwakil check nvcc
1fe37680
IlyasMoutawwakil probe
bbfda35e
IlyasMoutawwakil fix
68a373cd
IlyasMoutawwakil test psum
825f6d49
IlyasMoutawwakil test
028a39f1
IlyasMoutawwakil test
ee173a5f
IlyasMoutawwakil probe
900e984d
IlyasMoutawwakil fix
09a711f3
IlyasMoutawwakil test
a0d49400
IlyasMoutawwakil nan issue
653b7b3b
IlyasMoutawwakil repro
ed8af6bb
IlyasMoutawwakil repro
fb8d3388
IlyasMoutawwakil fix
562eb512
IlyasMoutawwakil simplifications
481e5e68
IlyasMoutawwakil fix
d3dbd325
IlyasMoutawwakil fix
9f168ff1
IlyasMoutawwakil fix
7274c22e
IlyasMoutawwakil fix
a800b8c3
IlyasMoutawwakil fix
804c9885
IlyasMoutawwakil fix
7f365628
IlyasMoutawwakil fix
6df9f47e
IlyasMoutawwakil empty
c3432c94
IlyasMoutawwakil simplify
77f09fe5
IlyasMoutawwakil test deepseek
7dfbedde
IlyasMoutawwakil dsv4 only
689cc296
IlyasMoutawwakil download dsv4
8de089ce
IlyasMoutawwakil fix test
cb4d6f9e
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
389eee8d
IlyasMoutawwakil push
8f29ed6a
IlyasMoutawwakil test
946e2007
IlyasMoutawwakil fix
6619647d
IlyasMoutawwakil fix ep plan
913339ac
IlyasMoutawwakil fix attempt
7cfc2b26
IlyasMoutawwakil debug
4c12f6ea
IlyasMoutawwakil attempt
92e63378
IlyasMoutawwakil debug
a9997f59
IlyasMoutawwakil fixes in modeling
197a8f62
IlyasMoutawwakil more modeling changes
30d0780d
IlyasMoutawwakil more modeling changes
3b1c470a
IlyasMoutawwakil initial megamoe works but wrong output, deepgemm fully functional
51a9425b
IlyasMoutawwakil mega moe works !
a4130db8
IlyasMoutawwakil simplification
b140f7c1
IlyasMoutawwakil comments
9ff656bc
ArthurZucker
ArthurZucker commented on 2026-05-12
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
b5b741e5
ArthurZucker
ArthurZucker commented on 2026-05-12
ArthurZucker ArthurZucker added for patch
IlyasMoutawwakil fix merge
12312581
IlyasMoutawwakil fix rank divergence
a8cc32ac
IlyasMoutawwakil use _keep_in_fp32_modules
c6486cbc
IlyasMoutawwakil simplify quantizer
bd68b6f8
ArthurZucker
ArthurZucker commented on 2026-05-12
IlyasMoutawwakil introduce indexerscorer
8630f9d1
IlyasMoutawwakil remove testing file
f3b933f6
IlyasMoutawwakil style
029a4c91
IlyasMoutawwakil weight rename for scorer
85b57686
IlyasMoutawwakil style
7f19516d
ArthurZucker
ArthurZucker approved these changes on 2026-05-12
IlyasMoutawwakil address review comments
e395e4c8
IlyasMoutawwakil style
1b94261e
IlyasMoutawwakil align with laguna
af51a6e1
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
ca934bf5
IlyasMoutawwakil IlyasMoutawwakil marked this pull request as ready for review 120 days ago
IlyasMoutawwakil stale comments and styling
f18ab28c
ArthurZucker
ArthurZucker approved these changes on 2026-05-13
ArthurZucker
ArthurZucker commented on 2026-05-13
IlyasMoutawwakil IlyasMoutawwakil force pushed from d4cbc61a to 50bcdb42 114 days ago
IlyasMoutawwakil use grouped finegrained fp8 for grouped linear
26f3ba70
IlyasMoutawwakil IlyasMoutawwakil force pushed from 50bcdb42 to 26f3ba70 114 days ago
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
4039ed69
IlyasMoutawwakil comments and docs
9f2c97b8
ArthurZucker Revert "[DeepSeek V4] Fix MoE converter substring-matching FP8 scale …
8dc75bd6
IlyasMoutawwakil Merge remote-tracking branch 'origin/main' into deepgemm-isolation
57081ccf
IlyasMoutawwakil fix quantizer and use upstream deepgemm
78fcee66
IlyasMoutawwakil IlyasMoutawwakil force pushed from 8c561b5c to 78fcee66 104 days ago
IlyasMoutawwakil remove force al kernels
2e4203fa
ArthurZucker [deepseek_v4] fix FP8 save-side substring match + fp32 gate through s…
5f9e5cdf
IlyasMoutawwakil always floats scal inv with triton
afd1b708
IlyasMoutawwakil minimal changes i quantizer
619b5e9b
IlyasMoutawwakil force stable abi build for now
3e76df5e
IlyasMoutawwakil support transposed
f0e2859e
IlyasMoutawwakil fix yarn
d0c4a6df
IlyasMoutawwakil unnecessary change
c1adcd70
sywangyi add flash base test and fix the incorrect and crash issue during enab…
143835f3
IlyasMoutawwakil merge
b47b30f0
IlyasMoutawwakil better errors across deepgemm
8bd9619e
IlyasMoutawwakil more better errors for deepgemm
751c8b5c
IlyasMoutawwakil ep tests
70c7c3b9
ArthurZucker Apply suggestions from code review
464512e1
ArthurZucker scope supports_round_trip to quantization ops only
d6fd64db
ArthurZucker Apply suggestions from code review
6e175094
ArthurZucker Update src/transformers/core_model_loading.py
32596183
ArthurZucker
ArthurZucker commented on 2026-05-29
ArthurZucker style: ruff format
2c06aefc
IlyasMoutawwakil custom ep/tp plan for megamoe
7dfe6bda
IlyasMoutawwakil process megamoe after weight loading
18891635
IlyasMoutawwakil fix megamoe process
46dc2897
ArthurZucker up
e9ccaa05
ArthurZucker the actual fix???
7d746846
ArthurZucker up
87d2ee99
ArthurZucker nit comment for futur
7be266b7
ArthurZucker style: drop unused exception variable in Fp8Dequantize._dequantize_one
5e102f16
ArthurZucker up
8631e6fb
ArthurZucker style: apply make fix-repo docstring formatting
b05bda48
ArthurZucker
ArthurZucker approved these changes on 2026-06-01
IlyasMoutawwakil Merge pr-46265 into deepgemm-isolation, drop redundant scale-merge wo…
7481aa67
IlyasMoutawwakil simplify after merge
f7698471
IlyasMoutawwakil review comments
2b8bba05
IlyasMoutawwakil Merge branch 'main' into deepgemm-isolation
beac5f22
IlyasMoutawwakil revert unncessary
f72bc20c
IlyasMoutawwakil avoid name contraction
b5f2d451
IlyasMoutawwakil fixes
014707d7
IlyasMoutawwakil dynamic tp/ep plan swap
2a77aa76
IlyasMoutawwakil fix deepgemm single error and device assertion docs
6922e3a9
IlyasMoutawwakil properly support bias in grouped linear
74711095
IlyasMoutawwakil simplify
b63ce4fd
IlyasMoutawwakil fix dist error handling and min cuda compatbility
c12e8924
IlyasMoutawwakil fix least cuda compat
1ec0147e
IlyasMoutawwakil get alignment from deepgemm and cache it
4fd793d3
github-actions
IlyasMoutawwakil IlyasMoutawwakil merged fa6c8308 into main 99 days ago
IlyasMoutawwakil IlyasMoutawwakil deleted the deepgemm-isolation branch 99 days ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone