[Perf][PCP] Fuse direct-KV publication fence - #54030
Draft
LopezCastroRoberto wants to merge 6 commits into
Draft
LopezCastroRoberto wants to merge 6 commits into
LopezCastroRoberto wants to merge 6 commits into
Conversation
… memory Port the direct-final PCP KV idea from vllm-project#49517 onto the merged vllm-project#50484 symmetric-memory substrate. Producers write final Main-KV and Indexer-K rows into every replica through peer pointers and skip the vllm-project#52046 AllGather + insert path. Signed-off-by: 云挚 <ningyunxiao.nyx@antgroup.com>
… oracle Make opt-in SymmMem allocation all-or-nothing at startup. Fence timeout now issues an unconditional PTX trap. Oracle covers fp8_ds_mla + Indexer-K on a packed backing (PCP2/PCP4). Wire teardown and drop peer-view refs on close. Signed-off-by: 云挚 <ningyunxiao.nyx@antgroup.com>
Keep this PR P2P-only. NVLS multimem.st lives on foraxe:agent/pcp-direct-kv-symm-mem-multimem. Signed-off-by: 云挚 <ningyunxiao.nyx@antgroup.com>
Scope symmetric KV publication to each PCP subgroup while allowing TP>1, and add a four-GPU TP2/PCP2 byte oracle. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: 云挚 <ningyunxiao.nyx@antgroup.com>
Keep direct symmetric-memory allocation integrated with the current common allocate_kv_cache path after rebasing onto upstream main. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: 云挚 <ningyunxiao.nyx@antgroup.com>
Move epoch advancement onto the device and combine system-scope publication and peer waiting in a single Triton kernel. This reduces every direct-KV fence from two kernel launches to one and lets the epoch advance correctly during CUDA graph replay. Add a two-GPU replay test that captures the fence, replays it repeatedly, and verifies that every replay advances the device-resident epoch. Depends-On: vllm-project#52863
mergify
Bot
added
deepseek
Contributor
|
This pull request has merge conflicts that must be resolved before it can be |
LopezCastroRoberto
added a commit
to LopezCastroRoberto/vllm
that referenced
this pull request
Sep 7, 2026
Fuse device epoch advancement, system-scope publication, and peer waiting into one graph-replay-safe direct-KV fence kernel. Depends on vllm-project#52863.
LucasWilkinson
pushed a commit
to LucasWilkinson/vllm
that referenced
this pull request
Sep 8, 2026
Fuse device epoch advancement, system-scope publication, and peer waiting into one graph-replay-safe direct-KV fence kernel. Depends on vllm-project#52863.
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
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.
Summary
Follow-up to #52863.
Motivation
#52863 launches separate publish and wait kernels for every direct-KV fence.
These fences execute once per sparse layer, making repeated launch overhead
material in PCP profiles.
Keeping the epoch on the host also prevents a captured CUDA graph from
advancing it on each replay.