Repository navigation

[2/7] fix(speculative): make streaming hidden-state capture work on vLLM 0.29 - #2522

Closed

h-guo18 wants to merge 3 commits into

haoguo/dflash-rope-fp32from

haoguo/streaming-vllm029-fixes

Closed

[2/7] fix(speculative): make streaming hidden-state capture work on vLLM 0.29#2522
h-guo18 wants to merge 3 commits into
haoguo/dflash-rope-fp32from
haoguo/streaming-vllm029-fixes

Commits on Oct 30, 2026
- fix(speculative): keep the DFlash draft's RoPE frequencies in fp32 on a fresh convert
  
  h-guo18
  and
  claude
  committed
  8a7cfee
- fix(speculative): correct the streaming hidden-state connector on current vLLM
  
  h-guo18
  and
  claude
  committed
  a5132fc
- fix(speculative): find the capture layer when vLLM merges it into the attention group
  
  h-guo18
  and
  claude
  committed
  25f14f4