[Bugfix][PCP][DCP] Respect interleave in indexer KV gather mapping - #56715
Merged
LucasWilkinson merged 2 commits intoSep 13, 2026
Merged
LucasWilkinson merged 2 commits into
LucasWilkinson merged 2 commits into
Conversation
Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Collaborator
Author
|
/ci run |
|
✅ Triggered Buildkite CI #88671 for commit |
LucasWilkinson
enabled auto-merge (squash)
September 13, 2026 18:32
github-actions
Bot
added
the
ready
Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Collaborator
Author
|
/ci run |
|
✅ Triggered Buildkite CI #88673 for commit |
tlrmchlsmth
approved these changes
Sep 13, 2026
ItsRoy69
pushed a commit
to ItsRoy69/vllm
that referenced
this pull request
Sep 15, 2026
…llm-project#56715) Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
keneoneth
pushed a commit
to keneoneth/vllm
that referenced
this pull request
Sep 16, 2026
…llm-project#56715) Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: OpenAI Codex <noreply@openai.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
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
PCP’s index gather assumes interleave 1. With DCP4/interleave64, token 1 reads the key for token 64. Use the configured interleave when mapping gathered shards back to token positions, and keep padded gathers in bounds. No new kernel.
Separate from #55802, which removes the interleave restriction; this PR leaves that guard unchanged. Duplicate searches found no open PR fixing this PCP gather mapping.
Validation on main
b6e2aa748b: 72 ordering cases passed (CPU; device copies stubbed), covering interleave1/64, DCP2/4/8, short/boundary contexts and split request rows. Restoring the old arithmetic fails all 36 interleave64 cases. Pre-commit/mypy passed.CUDA_VISIBLE_DEVICES='' .venv/bin/python -m pytest tests/v1/attention/test_indexer_dcp_localize.py -k pcp_plan_deinterleave_restores_global_order -qCombined-head GLM-5.2-NVFP4 PCP4+TP1+EP4+DCP4 → TP4 tests passed in eager and CUDA-graph modes: GPU and CPU-offload each 28/32 GSM8K, direct TP4 26/32 in both runs. All 32 GPU/CPU pairs matched tokens/log probabilities in each run, with confirmed CPU hits; 6,155-token retrieval also passed (6,144 CPU-restored tokens; continuations differed). Zero transfer failures or pending stores. PCP requires
-cc.cudagraph_mode=PIECEWISE; its default FULL_AND_PIECEWISE fails startup. Graph run captured prefiller sizes 4–1024 and logged 100 PIECEWISE prefiller / 8,432 FULL decoder dispatches. Uses default breakable graphs, not torch.compile. Clean merge: mainb6e2aa748b, #56645f995e33f61, #5580210afd0150e, #56715899a24c572; merged #56621, managed precompiled native extensions, no optional top-k changes. Small sample, not a full accuracy benchmark. UCX errors occurred only during teardown; GPUs released.AI assistance: OpenAI Codex. Draft pending human review and validation.