Skip to content

fix(sglang): tighten rollout argument guidance - #493

Closed
CjhHa1 wants to merge 1 commit into
Tencent-Hunyuan:mainfrom
CjhHa1:fix/sglang-arg-validation-docs
Closed

CjhHa1 wants to merge 1 commit into
Tencent-Hunyuan:mainfrom
CjhHa1:fix/sglang-arg-validation-docs

Conversation

@CjhHa1

@CjhHa1 CjhHa1 commented Sep 20, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • inline the two single-use dropped-argument helpers into the backend filter that owns the behavior
  • align the safe server-port bound with pinned SGLang 0.5.12.post1's actual grpc_port = port + 10000 derivation
  • narrow the Qwen3 guidance to a 4B dense baseline and document larger/MoE warmup, attention-backend, CUDA-graph, and temporary-port exceptions without changing recipe values

Related Issue

N/A — follow-up to #392.

Test Plan

  • SKIP=no-commit-to-branch uvx pre-commit run --all-files --show-diff-on-failure — passed
  • CPU assertion harness for UniRL-only key filtering, unknown-key warn/strict behavior, required-key precedence, the 55535 port boundary, repeated port reservation, and YAML parsing of all 9 touched recipes — passed
  • GPU rollout: Not run; production behavior is unchanged apart from expanding the valid server-port range to match the pinned SGLang derivation. Recipe edits are comments only.

The focused assertions remain one-off because the repository intentionally keeps no committed top-level test tree.

Compatibility / Risk

Default warn/strict behavior and filtered output are unchanged. The server-port candidate range expands from 1024–35535 to 1024–55535; under pinned SGLang 0.5.12.post1, an automatically derived gRPC port therefore remains within 65535. No checkpoint, data-format, recipe-value, GPU-count, or memory-requirement changes.

Reviewer Notes

PR #471 migrates SGLang to 0.5.19 and removes the derived-port cap while renaming CUDA-graph fields. If it merges first, this branch should be rebased: the 0.5.12-specific config hunk and wording become obsolete, while the helper inlining and baseline/exception documentation still apply.

AI-assisted implementation; the submitter reviewed the full diff. Duplicate-work review found #471's version-migration overlap described above, but no open PR covering the #392 helper/documentation follow-up.

Checklist

  • I reviewed the changed code and removed unrelated/generated artifacts.
  • I updated tests, docs, and configs where needed, or explained why not.

@github-actions github-actions Bot added the need review Ready and waiting for review label Sep 20, 2026
@CjhHa1

CjhHa1 commented Sep 20, 2026

Copy link
Copy Markdown
Collaborator Author

Superseded by the rebased #471: the helper inlining and documentation cleanup are now included there, while the 0.5.12-specific gRPC port correction is obsolete under the 0.5.19 migration.

@CjhHa1 CjhHa1 closed this Sep 20, 2026
@github-actions github-actions Bot removed the need review Ready and waiting for review label Sep 20, 2026
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.

1 participant