[Attention] Require explicit DCP support from attention implementations - #55780
Merged
Merged
Conversation
Default the implementation capability to false and preserve existing DCP backends through explicit opt-ins. Cover positive and negative selection, including sparse ROCm MLA and inherited implementations. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com>
AndreasKaratzas
requested review from
LucasWilkinson,
WoosukKwon,
alexm-redhat,
mgoin,
njhill,
pavanimajety,
tlrmchlsmth,
vadiklyutiy,
yewentao256,
youkaichao,
zhuohan123 and
zyongye
as code owners
September 8, 2026 01:50
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Claude Code Review
This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.
Tip: disable this comment in your organization's Code Review settings.
Member
Author
|
/ci run |
AndreasKaratzas
requested review from
MatthewBonanni,
WoosukKwon,
mgoin,
tjtanaa,
tlrmchlsmth,
yewentao256 and
zyongye
as code owners
September 8, 2026 01:50
AndreasKaratzas
requested review from
LucasWilkinson,
MatthewBonanni,
alexm-redhat,
njhill,
pavanimajety,
vadiklyutiy,
youkaichao and
zhuohan123
as code owners
September 8, 2026 01:50
|
✅ Triggered Buildkite CI #87609 for commit |
Explicitly select the multiprocessing executor for TP=2/DCP=2 selector configurations. These tests never launch workers, so CUDA executor autodetection must not require two physical GPUs before selection runs. This fixes the 11 configuration failures in Buildkite build 87609. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Member
Author
|
/ci run |
|
✅ Triggered Buildkite CI #87613 for commit |
tjtanaa
approved these changes
Sep 8, 2026
Member
Author
|
The remaining test group is a known flaky one. |
2 tasks
5 tasks
ItsRoy69
pushed a commit
to ItsRoy69/vllm
that referenced
this pull request
Sep 10, 2026
…ns (vllm-project#55780) Signed-off-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Jyotirmoy Roy <jyotirmoyroy649@gmail.com>
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
3 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.
AttentionImplBase.supports_dcptoFalseso implementations must explicitly opt into DCP.This follows the review on #54404, which raised concern that the new selection check could reject supported backends such as Triton MLA. The
Truedefault originated in #40996; #54404 preserved it and explicitly disabled DCP for sparse AITER MLA. Local testing confirmed that Triton MLA remained usable, but several unsupported implementations also inheritedTrueand failed the worker's missing-LSE check after loading weights. Explicit opt-ins move those failures into backend selection and prevent new implementations from advertising DCP implicitly.TruedeclarationFlashAttentionImplFLASH_ATTN,FLASH_ATTN_DIFFKVFlashInferImplFLASHINFERTritonMLAImplTRITON_MLAAiterMLAImplROCM_AITER_MLA,ROCM_AITER_TRITON_MLACutlassMLAImplCUTLASS_MLAFlashMLAImplFLASHMLAFlashInferMLAImplFLASHINFER_MLATokenspeedMLAImplTOKENSPEED_MLAFlashAttnMLAImplFLASH_ATTN_MLAFlashMLASparseImplFLASHMLA_SPARSEFlashInferMLASparseImplFLASHINFER_MLA_SPARSEPrepared with AI assistance (OpenAI Codex).