Repository navigation
Conversation
Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
Author
|
With |
drakosha
added a commit
to drakosha/FlashMLA
that referenced
this pull request
Nov 3, 2026
vllm-project#19 allocated the split-KV accumulators only when the scheduler has more than one partition. The V4.1 sync (vllm-project#20) brought back the upstream interface, which allocates them whenever split-KV scheduling is enabled, and on SM90 that is always. With a single partition no request is split: the SM90 kernel writes `out` / `lse` directly and the combine has nothing to merge, so the (b + num_sm_parts) x s_q x h_q x d_v fp32 buffers are never read. In vLLM's mixed-batch layout (b=1, s_q=num_tokens) they grow to 8 GiB at 32k tokens and reintroduce the runtime OOM of vllm-project/vllm#53413 (vllm-project/vllm#59305). Skip the accumulators and the combine launch on SM90 when there is a single partition, and bring back the workspace test from vllm-project#19 for SM90. SM100 is left to vllm-project#42: its kernels in the new tree build the o_accum tensor map and offset its pointers unconditionally, so there split-KV has to be turned off before scheduling instead. Signed-off-by: Mikhail Kostryukov <mike@triptrack.net> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fix a OOM when running Mistral Large 4 on SM100 devices. (On SM100, FlashMLA is the default kernel for Mistral Large 4)
This PR is effectively the same fix as #19 in the sense that we do not allocate
o_accumandlse_accum(on SM100), which is reverted by #20.Due to how this fixes the OOM, this fix only works for fp8 kv cache (AFAICT nvfp4 would take the split path). Also this PR does not fix SM90 regression from #20, it should be fixes by #29 instead.