attn: enable Split-K for mix-batch decode rows - #485
Open
krisclarkdev wants to merge 7 commits into
Open
Conversation
Chunked-prefill serving left long-KV decode rows with num_kv_splits=1. Use get_num_splits and allocate the split temporary output so mix-batch decode can parallelize along K like pure decode. Signed-off-by: Kristopher Clark <krisclarkdev@krisbox.org> Co-authored-by: Cursor <cursoragent@cursor.com>
4 tasks
Contributor
Author
|
Required companion of vLLM FA-in-graph / XPU graphs PR: vllm#50038. Please treat kernels #485–#489 as part of the graphs stack alongside vllm #48677 (torch/image) and #49813 (softcap/MXFP8). |
This was referenced Jul 27, 2026
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Purpose
Enable Split-K on the mix-batch decode half in
csrc/flash_attn/flash_api.cpp(was hardcodednum_kv_splits=1). Pure-decode already usesget_num_splits; chunked-prefill serving left long-KV decode rows single-split.Related: group splitkv (#337), small-
q_lenSplit-K route (#455). As of 2026-07-24,mainstill hardcodes mix-batchnum_kv_splits=1.Test Plan
estimate_get_num_splits> 1)Test Result
Hardware: Intel Arc Pro B70 · Commit:
a8a3cccAccuracy
20/20 focused (float16/bf16, block64, head64/128).
Micro (attn hot-op vs base FA2)
mix_long_kvmix_decode_heavyinterleaved_shortServe E2E (Qwen2.5-7B-Instruct, bf16, enforce-eager; median of 3 after warmup; failed=0 on primary)
Claim: micro win on long mix-batch; not claiming ≥10% E2E serve tok/s (flat under these recipes).
(Optional) Documentation Update
None.