Repository navigation
Conversation
The CuTe DSL Lamport all-reduce fused with mHC post, collapse and RMSNorm (and the MoE finalize publish) is parametrized by the TP world size, but supports_mhc_all_reduce only admitted TP4. The Blackwell single-node TP recipes now default to TP2, where every sublayer boundary at <=16 tokens fell back to a separate all-reduce plus the unfused mHC post and pre. Admit TP2 and extend the kernel-vs-unfused test to TP2. Signed-off-by: Dakai An <dakaian108@gmail.com>
andakai
requested review from
AndreasKaratzas,
WoosukKwon,
mgoin,
tlrmchlsmth,
yewentao256 and
zyongye
as code owners
October 7, 2026 06:07
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
mergify
Bot
added
deepseek
2 of 6 tasks
This was referenced Oct 19, 2026
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
The fused CuTe DSL Lamport all-reduce + mHC post/collapse/RMSNorm (#57643, #58586) only runs at TP4. The Blackwell single-node recipes now default to TP2 (vllm-project/recipes#1002), so at TP2 every small-batch sublayer boundary still runs a separate all-reduce plus unfused mHC. The kernel is already parametrized by the TP world size, so this PR admits TP2 in
supports_mhc_all_reduce(TP8 stays off) and extends the tests to TP2.Evaluation
Setup: 2 × GB200, DeepSeek-V4.1-Flash
Correctness
Performance:
vllm bench serve, random dataset, ISL 1024 / OSL 512, no speculative decoding, mean of 3 seeds.Swapping the two arms' GPUs (and NUMA nodes) gives the same picture: TPOT changes by −12.1%, −13.9%, −10.8% and +0.7% at C1, C4, C16 and C64. TTFT is unchanged.
With DSpark (recipe default, k=5):
At C≥8, DSpark steps exceed the 16-token fused limit, so the fused path is not taken.
Notes
Invalid MXFP8 split-K tacticstartup crash, which is unrelated to this PR. [Bugfix]Fix FlashInfer warmup crash: autotune(tuning_buckets=...) must round M up, not down (Invalid MXFP8 split-K tactic with spec-decode) #58165 fixes it.🤖 Generated with Claude Code