[Klaud Cold] Port #3394: dsv41flash-fp4-b300-vllm-agentic-dspark to vLLM nightly-0961bbae with FlashInfer sparse indexer + fp8 KV

[functionstackx]

Summary

Port of #3394 (by @xinli-sw) onto the srt-slurm recipe that #3428 moved this config to. #3394 edits the retired benchmarks/single_node/agentic/dsv41flash_fp4_b300_vllm_mtp.sh bash script, which the B300 DeepSeek-V4.1-Flash vLLM AgentX config no longer runs.

Recipes touched: dsv41flash-fp4-b300-vllm-agentic-dspark

Changes:

• configs/nvidia-master.yaml: image vllm/vllm-openai:deepseekv41-flash-0909 → vllm/vllm-openai:nightly-0961bbae2894d574be790d219651824eb199318e (2026-09-22 nightly).
• srt-slurm-recipes/dsv41flash/vllm/b300-fp4-mtp/agentic.yaml:
  • container bump to the same nightly.
  • attention-config: {"backend":"FLASHINFER_MLA_SPARSE_DSV41","indexer_kv_dtype":"mxfp4","indexer_sparse_logits":true} at both TP2 and TP4 (as Update B300 DeepSeek V4.1 Flash vLLM nightly / 更新 B300 DeepSeek V4.1 Flash vLLM nightly #3394's script does; its description mentions MegaAttention at TP4, but its code uses FlashInfer for both).
  • kv-cache-dtype: fp8.
  • VLLM_ENGINE_READY_TIMEOUT_S 3600 → 7200 and health_check.max_attempts 360 → 720 (10 s interval), replacing Update B300 DeepSeek V4.1 Flash vLLM nightly / 更新 B300 DeepSeek V4.1 Flash vLLM nightly #3394's launcher-side VLLM_ENGINE_READY_TIMEOUT_S=7200.
• perf-changelog.yaml: new entry.

Topology, conc lists, CUDA-graph ladders and memory limits are unchanged. #3394's EVAL_ONLY script edits and DSV41_MIN_CUDAGRAPH_CAPTURE_SIZE launcher export apply only to the legacy script path and are not needed here.

References: vLLM tracking issue, upstream recipe.

Test plan

• full-sweep-enabled sweep passes.

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[claude]

🟡 (optional) This merges a perf-changelog entry with pr-link: PRLINK_PLACEHOLDER instead of a real PR URL, breaking the changelog's own convention. Every other of the 1039 pr-link entries in perf-changelog.yaml points to a real https://github.com/SemiAnalysisAI/InferenceX/pull/NNNN URL; this is the only literal placeholder, so anyone auditing or tooling that parses pr-link to find the originating PR gets a dead value for this recipe change. Fix: replace PRLINK_PLACEHOLDER with the actual PR URL before merge (or have CI fail the append-only changelog check when a pr-link is not a valid PR URL).

Why this was flagged

The entry appended at perf-changelog.yaml:8962-8970 documents the image/attention-config/kv-cache-dtype/timeout changes but sets pr-link: PRLINK_PLACEHOLDER (perf-changelog.yaml:8970) rather than a real PR link. AGENTS.md:124 requires a perf-changelog entry for every performance-affecting change but does not sanction placeholder links, and every other of the 1039 pr-link entries in the file is a resolved github.com/.../pull/NNNN URL. Anyone using pr-link to trace which PR introduced a given recipe change (audit, regression bisection, tooling) gets an unresolvable placeholder instead of the base branch's convention of a working URL. No check in the diff or repo enforces pr-link is a real URL before merge.

Verification: normal. perf-changelog.yaml:8970 (the entry this PR appends, lines 8966-8970) sets pr-link: PRLINK_PLACEHOLDER. It is the sole non-URL pr-link in the file (1039 pr-link occurrences; only this one is not a github .../pull/NNNN URL). Beyond breaking the file convention, the repo's own gate rejects it: infx/workflows/validate_perf_changelog.py defines CANONICAL_PR_LINK =… | nit. The finding is…