Repository navigation

Add vLLM NVFP4 MLA KV-cache fake quant - #2620

Merged

sychen52 merged 2 commits into

NVIDIA:mainfrom

sychen52:glm5.3-flash-kv

Oct 3, 2026

Merged

Add vLLM NVFP4 MLA KV-cache fake quant#2620
sychen52 merged 2 commits into
NVIDIA:mainfrom
sychen52:glm5.3-flash-kv

Commits on Oct 2, 2026
- Add the kv_nvfp4_mla unit for NVFP4 MLA KV-cache fake quant in vLLM
  sychen52
  committed
  20a4944
- Quantize the MLA RoPE key to FP8 in the kv_nvfp4_mla unit
  
  sychen52
  and
  claude
  committed
  1e3a6c9