Repository navigation

[Feature][MRV2] Adapt extract_hidden_states for Model Runner V2 - #14699

Merged

weiguihua2 merged 10 commits into

vllm-project:mainfrom

yjyang62:mrv2-extract-hidden-states-5b6a

Sep 9, 2026

Merged

[Feature][MRV2] Adapt extract_hidden_states for Model Runner V2#14699
weiguihua2 merged 10 commits into
vllm-project:mainfrom
yjyang62:mrv2-extract-hidden-states-5b6a

Commits on Sep 9, 2026
- feat(mrv2): adapt extract_hidden_states for vLLM 0828 KV layout
  
  yjyang62
  authored and
  cursoragent
  committed
  ce71485
- fix(mrv2): collect all extract_hidden_states aux layers
  
  yjyang62
  authored and
  cursoragent
  committed
  39bb43a
- refactor(mrv2): reuse upstream ExtractHiddenStatesSpeculator
  
  yjyang62
  authored and
  cursoragent
  committed
  ff07e8a
- refactor(mrv2): drop unnecessary extract_hidden_states adapter
  
  yjyang62
  authored and
  cursoragent
  committed
  50c0742
- refactor(mrv2): inline extract_hidden_states cache shape
  
  yjyang62
  authored and
  cursoragent
  committed
  39ed373
- refactor(mrv2): identify dump cache by HiddenStateCacheSpec only
  
  yjyang62
  authored and
  cursoragent
  committed
  867bbab
- docs: note extract_hidden_states dumps train EAGLE/EAGLE-3 drafts
  
  yjyang62
  authored and
  cursoragent
  committed
  3357517
- test: add token-in-token-out e2e for MRV1 and MRV2
  
  yjyang62
  authored and
  cursoragent
  committed
  1605956
- revert: drop standalone token-in-token-out e2e
  
  cursoragent
  and
  yjyang62
  committed
  7def8c6
- test: run skip_tokenizer_init dump e2e on text-only Qwen3
  
  cursoragent
  and
  yjyang62
  committed
  04668ae