Repository navigation

Draft:FP8 R1 - #6100

Closed

litaotju wants to merge 36 commits into

NVIDIA:mainfrom

lfr-0531:user/fanrongl/debug_moe

Closed

Draft:FP8 R1#6100
litaotju wants to merge 36 commits into
NVIDIA:mainfrom
lfr-0531:user/fanrongl/debug_moe

Commits on Jul 25, 2025
- [Draft] DeepGEMM Blackwell integration
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  09b012e
- Clean up fused_moe_deepgemm.py
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  ec400ab
- Moving permute space allocation to GPU
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  d9a85ac
- optimize padding in deepgemm moe.
  lfr-0531
  committed
  7c4045c
- add torch compile to per_token_cast_to_fp8_e8m0 and rm the two sync.
  lfr-0531
  committed
  20b2592
- Improve bmm.
  
  yuxianq
  authored and
  lfr-0531
  committed
  c74a31a
- Online resmooth for fp8 checkpoint on Blackwell. (#2)
  
  yuxianq
  authored and
  lfr-0531
  committed
  d3e1797
- Fix OOM issue for fp8 resmooth. (#4)
  
  yuxianq
  authored and
  lfr-0531
  committed
  d83cc25
- Enbale masked grouped GEMM (#5)
  
  Barry-Delaney
  and
  lfr-0531
  committed
  e1e96fd
- Pin DeepGEMM's version to commit cc416ee. (#6)
  
  yuxianq
  authored and
  lfr-0531
  committed
  09b0465
- Improve resmooth. (#7)
  
  yuxianq
  authored and
  lfr-0531
  committed
  35b4e23
- Add compile for quantization kernels (#8)
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  dce291f
- Move SF transform to TRTLLM (#11)
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  b3ab47d
- Use local barrier to avoid multi-node hang issue. (#12)
  
  yuxianq
  authored and
  lfr-0531
  committed
  65d05d6
- optimize the masked index copy and index gather (#13)
  lfr-0531
  committed
  d65bdac
- Fix adp for deepgemm moe backend (#10)
  
  zongfeijing
  authored and
  lfr-0531
  committed
  0af69ac
- Use DeepGEMM main branch instead.
  
  yuxianq
  authored and
  lfr-0531
  committed
  6f431f6
- Revert "Use DeepGEMM main branch instead."
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  481fd50
- Use DeepGEMM main branch and disable ue8m0 cast. (#16)
  
  yuxianq
  authored and
  lfr-0531
  committed
  ab7175f
- fuse maskec index_copy and grouped fp8 quantization.
  lfr-0531
  committed
  97a21fd
- fix quantization accuracy issue.
  lfr-0531
  committed
  f668fa7
- Fuse swiglu and quant 2 (#18)
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  c6b8985
- Opt gather kernel (#19)
  
  zongfeijing
  authored and
  lfr-0531
  committed
  11053b7
- optimize the perf of masked_index_copy_group_quant_fp8.
  lfr-0531
  committed
  0173836
- fix duplicate load.
  lfr-0531
  committed
  bd94e37
- fuse scaling factor transform to _masked_index_copy_group_quant_fp8.
  lfr-0531
  committed
  f1d3115
- fix.
  lfr-0531
  committed
  acd4381
- add another for loop on the group dim.
  lfr-0531
  committed
  2d5beab
- Remove SFB transform from forward process (#23)
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  5653eea
- change deepgeem to a new commit that with torch dependency. (#24)
  lfr-0531
  committed
  49dcb98
- fix format and rebase bug.
  lfr-0531
  committed
  9997006
Commits on Jul 28, 2025
- fix dummy requests when estimate kv cache with attention DP enabled to avoid OOM (#25)
  lfr-0531
  committed
  d8ae02c
- Fuse quantize and transform e8m0 scales (#26)
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  fb3e467
- Revert "Fuse quantize and transform e8m0 scales (#26)" (#27)
  
  Barry-Delaney
  authored and
  lfr-0531
  committed
  9107cfa
- Fix CI install error for DeepGEMM. (#28)
  
  yuxianq
  authored and
  lfr-0531
  committed
  59b3957
- Reapply "Fuse quantize and transform e8m0 scales (#26)" (#27) (#29)
  Barry-Delaney
  authored
  3c413be

Commits

Commits on Jul 25, 2025

Commits on Jul 28, 2025