Conversation
mergify
Bot
added
deepseek
haosdent
force-pushed
the
fix-48568
branch
2 times, most recently
from
July 14, 2026 06:32
3a064ac to
19d5dcb
Compare
Contributor
Author
|
@AndreasKaratzas can you help to take a look, this is to fix MTP issues when I run it in AMD Mi300x |
haosdent
changed the title
haosdent
marked this pull request as ready for review
July 14, 2026 06:39
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Contributor
|
This pull request has merge conflicts that must be resolved before it can be |
Contributor
|
Seems like item 1 has been covered in #48036 |
haosdent
requested review from
aarnphm,
benchislett,
mgoin and
russellb
as code owners
July 29, 2026 00:17
Warmup only ran compute_logits' vocab-parallel all-gather at one row per request, but spec decode verifies max_num_reqs*(1+num_spec_tokens) rows. On ROCm/RCCL the larger message connects extra channels lazily on first use, so when that first use is a live decode step the connect fails with "no transport for peer on channel N" and the engine hangs (vllm-project#48568). Warm compute_logits at the spec-decode row count during init so those channels connect before serving. Signed-off-by: haosdent <haosdent@gmail.com>
haosdent
changed the title
haosdent
changed the title
Contributor
Author
|
@vllmellm I have rebase on the latest main |
Contributor
Author
|
Thx @cjackal , have rebased. |
njhill
added
the
mrv1-only
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
4 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Purpose
Fixes DeepSeek/GLM-5.2 MTP speculative decoding hanging at startup (#48568).
Warmup runs
compute_logits' vocab-parallel all-gather at one row per request, but spec decode verifiesmax_num_reqs*(1+num_speculative_tokens)rows. On ROCm/RCCL the larger message connects extra channels lazily on first use, so when that first use is a live decode step the connect fails (no transport for peer on channel N) and the engine hangs. Warmcompute_logitsat the spec-decode row count during init.Test Plan
Serve GLM-5.2-FP8 with MTP speculative decoding on 8x MI300X (gfx942):
Test Result
Server starts and serves normally with MTP speculative decoding, producing coherent completions. Without this change the engine hangs in the logits all-gather on the first speculative-decode step.