Repository navigation

Fix/compressed tensors nvfp4 hopper - #40022

Merged

BBuf merged 6 commits into

sgl-project:mainfrom

jlqibm:fix/compressed-tensors-nvfp4-hopper

Oct 13, 2026

Merged

Fix/compressed tensors nvfp4 hopper#40022
BBuf merged 6 commits into
sgl-project:mainfrom
jlqibm:fix/compressed-tensors-nvfp4-hopper

Commits on Oct 8, 2026
- [compressed-tensors] Serve NVFP4 weight-only via FP4 Marlin on pre-Blackwell
  
  jlqibm
  and
  claude
  committed
  c7fd88a
- [compressed-tensors] Test the NVFP4 weight-only downgrade warning
  
  jlqibm
  and
  claude
  committed
  ee710ae
- [compressed-tensors] Address review: honor --fp4-gemm-backend, reject mismatched NVFP4 global scales
  
  jlqibm
  and
  claude
  committed
  257f3a0
- [test] Fix import order in the FP4 linear numeric test (isort)
  
  jlqibm
  and
  claude
  committed
  d906a27
- [test] Consolidate the FP4 compressed-tensors tests into one CPU and one GPU file
  
  jlqibm
  and
  claude
  committed
  b5824c1
Commits on Oct 10, 2026
- [test] Register the FP4 numeric tests on 1-gpu-large (H100)
  
  jlqibm
  and
  claude
  committed
  2a5b4f2