Skip to content

[DeepSeek V4] Fix stale override wording in ROCm FlashMLA sparse-prefill disable - #30289

Closed
michaelzhang-ai wants to merge 1 commit into
mainfrom
michaelzhang-ai/dsv4-flashmla-warning-wording
Closed

michaelzhang-ai wants to merge 1 commit into
mainfrom
michaelzhang-ai/dsv4-flashmla-warning-wording

Conversation

@michaelzhang-ai

@michaelzhang-ai michaelzhang-ai commented Jul 6, 2026 •

Copy link
Copy Markdown
Collaborator

Motivation

#30237 made the ROCm/HIP disable of SGLANG_OPT_FLASHMLA_SPARSE_PREFILL unconditional (dropped the is_set() check), which is correct — but the warning message and preceding comment still say the env var can be set explicitly to override, which is no longer true. This was also raised by the gemini review bot on #30237.

Modifications

python/sglang/srt/arg_groups/deepseek_v4_hook.py: update the comment and logger.warning text to reflect that the flag is unconditionally disabled on ROCm/HIP until the sparse prefill kernel is validated there. No behavior change.

Accuracy Tests

Comment/log-string only; no functional change.

Checklist

  • Docs/log-string only, no behavior change.

CI States

Latest PR Test (Base): ❌ Run #28817248929
Latest PR Test (Extra): ❌ Run #28817248875

…ill disable

#30237 made the ROCm/HIP disable of SGLANG_OPT_FLASHMLA_SPARSE_PREFILL
unconditional (dropped the is_set() check), but the warning message and
comment still claimed the env var could be set to override it. Update the
wording to match the actual behavior.
@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

@michaelzhang-ai

Copy link
Copy Markdown
Collaborator Author

Context: this is a docs/log-string-only follow-up to #30237 (no behavior change).

Related disagg validation run (image ...-mi35x-20260706, contains #30237): https://github.com/sgl-project/sglang/actions/runs/28817206537

Note that run shows the DeepSeek-V4-Flash disagg configs still failing — but the cause is a MORI/RDMA prefill crash (RegisterRdmaMemoryRegion failed ... errno:22), not the sparse-prefill path this wording refers to. Details on #29982. This PR only corrects the now-stale "set it explicitly to override" text.

@Jiminator
Jiminator deleted the michaelzhang-ai/dsv4-flashmla-warning-wording branch September 14, 2026 04:44
@alexnails
alexnails restored the michaelzhang-ai/dsv4-flashmla-warning-wording branch September 14, 2026 05:43
@cursor
cursor Bot deleted the michaelzhang-ai/dsv4-flashmla-warning-wording branch September 17, 2026 07:45
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant