vllm
[SpecDecode] Reduce TP communication for large-vocab draft models speculative decoding
#39419
Merged
Merged
[SpecDecode] Reduce TP communication for large-vocab draft models speculative decoding #39419
benchislett
merged
14 commits into
vllm-project:main
from
EanWang211123:local_argmax_for_dflash
[feat] add local_argmax_reduction for dflash and qwen3-pard
886b2d5d
gemini-code-assist
commented
on 2026-04-09
EanWang211123
changed the title 
Merge branch 'main' into local_argmax_for_dflash
c8beb07e
[feat] add local-argmax for eagle3 llama p-eagle
8b246ce1
benchislett
commented
on 2026-04-30
benchislett
commented
on 2026-04-30
benchislett
commented
on 2026-04-30
benchislett
commented
on 2026-04-30
Merge branch 'main' into local_argmax_for_dflash
8aa572fd
Merge branch 'main' into local_argmax_for_dflash
fe1453a1
[feat] refact and add local-argmax logits
db23525c
[update] support local-argmx for qwen3-5
5587b3de
EanWang211123
changed the title
[update] remove draft_id_to_target_id init logits in llama_eagle3
1137e682
[feat] implement the logits for llama qwen
d97994a6
[fix] fix MRO conflict
bc039415
[fix] fix lint
4aa766c7
Merge branch 'main' into local_argmax_for_dflash
bf3bd55d
Merge branch 'main' into local_argmax_for_dflash
fafc1304
benchislett
commented
on 2026-06-09
benchislett
approved these changes
on 2026-06-10
Merge branch 'main' into local_argmax_for_dflash
773f68d2
Assignees
No one assigned
Labels
speculative-decoding
ready
v1
llama
qwen
deepseek
verified
Milestone
No milestone
Login to write a write a comment.
Login via GitHub