Skip to content

fix(dsv4): gate h200 reasoning parser flag#1263

Open
OCWC22 wants to merge 5 commits intoSemiAnalysisAI:mainfrom
OCWC22:fix/dsv4-h200-reasoning-parser-guard
Open

fix(dsv4): gate h200 reasoning parser flag#1263
OCWC22 wants to merge 5 commits intoSemiAnalysisAI:mainfrom
OCWC22:fix/dsv4-h200-reasoning-parser-guard

Conversation

@OCWC22
Copy link
Copy Markdown

@OCWC22 OCWC22 commented May 2, 2026

Summary

  • gate the H200 DSv4 --reasoning-parser deepseek_v4 flag behind ENABLE_DSV4_REASONING_PARSER=true
  • keep the parser disabled by default because the AgentX testing branch's docs/AGENTIC_TEST_RESULTS.md records dsv4-fp8-h200-vllm as ❌ FAILED with the DeepSeek-V4 reasoning parser issue (reasoning_start_str / reasoning_end_str missing)
  • preserve an opt-in path for future vLLM parser fixes without changing the default launcher behavior

Smoke test

  • bash -n benchmarks/single_node/dsv4_fp8_h200.sh passes
  • default ENABLE_DSV4_REASONING_PARSER expansion leaves REASONING_PARSER_ARGS empty, so the launcher omits --reasoning-parser deepseek_v4 unless explicitly enabled

@OCWC22 OCWC22 requested review from a team and Copilot May 2, 2026 19:36
Copy link
Copy Markdown
Contributor

@claude claude Bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

Copy link
Copy Markdown
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR keeps the DeepSeek-V4 H200 single-node launcher stable by disabling the vLLM DeepSeek-V4 reasoning parser by default, while preserving an opt-in path via an environment flag.

Changes:

  • Introduces ENABLE_DSV4_REASONING_PARSER (default false) to control whether --reasoning-parser deepseek_v4 is passed to vllm serve.
  • Refactors the reasoning-parser CLI arg into an optional Bash array so the argument is omitted cleanly when disabled.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

2 participants