Repository navigation

[Quantization] Clarify MXFP4 MoE W13 layout requirement from backends - #52240

Open

BowenBao wants to merge 3 commits into

vllm-project:mainfrom

BowenBao:bowenbao/clarify_mxfp4_moe_oracle

Open

[Quantization] Clarify MXFP4 MoE W13 layout requirement from backends#52240
BowenBao wants to merge 3 commits into
vllm-project:mainfrom
BowenBao:bowenbao/clarify_mxfp4_moe_oracle

Commits on Oct 9, 2026
- [MoE] Support interleaved w13 in convert_weight_to_mxfp4_moe_kernel_format
  
  BowenBao
  and
  claude
  committed
  8d76968
- [MoE] Convert GPT-OSS MXFP4 weights with the generic converter
  
  BowenBao
  and
  claude
  committed
  3cf8570
- [MoE] Fold the DeepSeek-V4 MXFP4 selector into select_mxfp4_moe_backend
  
  BowenBao
  and
  claude
  committed
  11b62e2