[Attention][DCP] Use FlashInfer native CP for MLA decode - #54012
Merged
pavanimajety merged 1 commit intoAug 27, 2026
Merged
Conversation
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Route MLA DCP decode through FlashInfer's monolithic CuTeDSL native-CP mask, pass the global causal bounds, and normalize natural-log LSE for the cross-rank merge. This enables uniform multi-token target verification for MTP, EAGLE, and DSpark while retaining the existing non-causal draft-backend capability check. Assisted-by: OpenAI Codex Signed-off-by: Summer Yang <girasoleyang@gmail.com>
GirasoleY
force-pushed
the
flashinfer-mla-native-dcp
branch
from
August 27, 2026 07:51
e52f77e to
10df552
Compare
Member
|
/ci run |
|
✅ Triggered Buildkite CI #85792 for commit |
pavanimajety
approved these changes
Aug 27, 2026
This was referenced Aug 27, 2026
khushali9
pushed a commit
to khushali9/vllm
that referenced
this pull request
Aug 29, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com> Signed-off-by: khushali9 <khushali.desai9@gmail.com>
askliar
pushed a commit
to askliar/vllm
that referenced
this pull request
Aug 30, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com>
am-cohere
pushed a commit
to am-cohere/vllm
that referenced
this pull request
Sep 1, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com>
mikeshawcode
pushed a commit
to mikeshawcode/vllm
that referenced
this pull request
Sep 1, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com> Signed-off-by: mikeshawcode <michaelwshaw2@gmail.com>
mikeshawcode
pushed a commit
to mikeshawcode/vllm
that referenced
this pull request
Sep 1, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com> Signed-off-by: mikeshawcode <michaelwshaw2@gmail.com>
LopezCastroRoberto
pushed a commit
to LopezCastroRoberto/vllm
that referenced
this pull request
Sep 2, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com>
mylibrar
pushed a commit
to tanyuqian/vllm
that referenced
this pull request
Sep 3, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com>
sheralskumar
pushed a commit
to sheralskumar/vllm
that referenced
this pull request
Sep 8, 2026
…t#54012) Signed-off-by: Summer Yang <girasoleyang@gmail.com>
4 tasks
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.
Summary
FlashInfer 0.6.17 exposes native context-parallel masking through CuTeDSL MLA decode kernel. This PR adds MLA DCP support through FlashInfer (limit to cuteDSL path only)
Relationship to existing work
This is not a duplicate of the existing speculative/DCP PRs:
I also searched open vLLM PRs for
causal_seqlens_kv_globaland the native FlashInfer MLA CP arguments and found no overlapping integration.Test
Kimi-K3 DCP8-DCP8 P/D w/ EAGLE3: