Repository navigation

[Bugfix][Model] GLM-5.x: load ModelOpt NVFP4 MLA projections into BF16 fused_qkv_a_proj - #59833

Open

avifenesh wants to merge 4 commits into

vllm-project:mainfrom

avifenesh:glm5next-nvfp4-mla-proj-load

Open

[Bugfix][Model] GLM-5.x: load ModelOpt NVFP4 MLA projections into BF16 fused_qkv_a_proj#59833
avifenesh wants to merge 4 commits into
vllm-project:mainfrom
avifenesh:glm5next-nvfp4-mla-proj-load

Commits on Oct 3, 2026
- [Bugfix][Model] GLM-5.x: load ModelOpt NVFP4 MLA projections into BF16 fused_qkv_a_proj
  
  avifenesh
  and
  claude
  committed
  4a7e091
Commits on Oct 10, 2026
- Ignore unused NVFP4 input scales for BF16 MLA projections
  
  avifenesh
  and
  codex
  committed
  be8e956
Commits on Oct 20, 2026
- Merge upstream main to refresh NVFP4 MLA loader fix
  avifenesh
  committed
  78169eb
- [Bugfix][Model] Preserve FP8 and native quantized scale routing
  avifenesh
  committed
  0f2bb87