Skip to content

Enable SGLang CSGMV split-K shrink by default on samplers - #151

Merged
kailash109 merged 2 commits into
mainfrom
codex/csgmv-split-k-default
Oct 29, 2026
Merged

kailash109 merged 2 commits into
mainfrom
codex/csgmv-split-k-default

Conversation

@kailash109

Copy link
Copy Markdown
Collaborator

Set SGLANG_ENABLE_LORA_CSGMV_SPLIT_K=1 in the rollout image so every sampler gets the split-K shrink from sgl-project/sglang#41951. The pinned fork revision already contains the backport (modal-projects/sglang#79); only the flag was off.

SGLang applies split-K only to BF16 rank-32 CSGMV shrinks with input size 4096 or 12288 and its default tile shape, and never in deterministic mode. Other models and shapes keep the original kernel. A recipe can still opt out by setting the variable to 0 in inference_env.

qwen35_9b_instruct_lora_256k set SGLANG_CSGMV_SPLIT_K, the flag's name before the upstream rename, so split-K never ran there. The DAPO example's explicit setting is now redundant. Both are removed.

Validation: deployment tests and pre-commit pass. Not GPU-benchmarked in this PR; the upstream PR has the kernel benchmarks.

Made with Cursor

Co-authored-by: Cursor <cursoragent@cursor.com>
Keep the rollout-image split-K flag. Main already enables it for LoRA
servers and replaced the DAPO routing fields this branch still carried.

Co-authored-by: Cursor <cursoragent@cursor.com>
@kailash109
kailash109 merged commit 45431f2 into main Oct 29, 2026
4 checks passed
@kailash109
kailash109 deleted the codex/csgmv-split-k-default branch October 29, 2026 18:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants