Draft:FP8 R1#6100

Closed

litaotju wants to merge 36 commits intoNVIDIA:mainfrom

lfr-0531:user/fanrongl/debug_moe

Commits on Jul 25, 2025

[Draft] DeepGEMM Blackwell integration

Barry-Delaney
authored and
lfr-0531
committed
Clean up fused_moe_deepgemm.py

Barry-Delaney
authored and
lfr-0531
committed
Moving permute space allocation to GPU

Barry-Delaney
authored and
lfr-0531
committed
optimize padding in deepgemm moe.
lfr-0531
committed
add torch compile to per_token_cast_to_fp8_e8m0 and rm the two sync.
lfr-0531
committed
Improve bmm.

yuxianq
authored and
lfr-0531
committed
Online resmooth for fp8 checkpoint on Blackwell. (#2)

yuxianq
authored and
lfr-0531
committed
Fix OOM issue for fp8 resmooth. (#4)

yuxianq
authored and
lfr-0531
committed
Enbale masked grouped GEMM (#5)

Barry-Delaney
and
lfr-0531
committed
Pin DeepGEMM's version to commit cc416ee. (#6)

yuxianq
authored and
lfr-0531
committed
Improve resmooth. (#7)

yuxianq
authored and
lfr-0531
committed
Add compile for quantization kernels (#8)

Barry-Delaney
authored and
lfr-0531
committed
Move SF transform to TRTLLM (#11)

Barry-Delaney
authored and
lfr-0531
committed
Use local barrier to avoid multi-node hang issue. (#12)

yuxianq
authored and
lfr-0531
committed
optimize the masked index copy and index gather (#13)
lfr-0531
committed
Fix adp for deepgemm moe backend (#10)

zongfeijing
authored and
lfr-0531
committed
Use DeepGEMM main branch instead.

yuxianq
authored and
lfr-0531
committed
Revert "Use DeepGEMM main branch instead."

Barry-Delaney
authored and
lfr-0531
committed
Use DeepGEMM main branch and disable ue8m0 cast. (#16)

yuxianq
authored and
lfr-0531
committed
fuse maskec index_copy and grouped fp8 quantization.
lfr-0531
committed
fix quantization accuracy issue.
lfr-0531
committed
Fuse swiglu and quant 2 (#18)

Barry-Delaney
authored and
lfr-0531
committed
Opt gather kernel (#19)

zongfeijing
authored and
lfr-0531
committed
optimize the perf of masked_index_copy_group_quant_fp8.
lfr-0531
committed
fix duplicate load.
lfr-0531
committed
fuse scaling factor transform to _masked_index_copy_group_quant_fp8.
lfr-0531
committed
fix.
lfr-0531
committed
add another for loop on the group dim.
lfr-0531
committed
Remove SFB transform from forward process (#23)

Barry-Delaney
authored and
lfr-0531
committed
change deepgeem to a new commit that with torch dependency. (#24)
lfr-0531
committed
fix format and rebase bug.
lfr-0531
committed