Skip to content

chore(recipes): remove 20 unused srt-slurm YAMLs / 删除 20 个未使用的 srt-slurm 配方 - #3465

Merged
functionstackx merged 2 commits into
mainfrom
cleanup/unused-srt-recipes
Sep 26, 2026
Merged

functionstackx merged 2 commits into
mainfrom
cleanup/unused-srt-recipes

Conversation

@Oseltamivir

Copy link
Copy Markdown
Collaborator

Summary

Remove 20 unreferenced srt-slurm recipe YAMLs, totaling 2,632 deleted lines:

Scope Files
Kimi-K2.5 / TRT-LLM / B200 FP4 / 8k1k 11
MiniMax-M3 / vLLM / GB200 FP4 / AgentX 6
Qwen3.5 / SGLang / GB200 FP4 / AgentX 1
Kimi-K3 / vLLM / MI355X FP4 MTP / AgentX 1
MiniMax-M3 / vLLM / MI355X FP4 MTP / AgentX 1

No master config, launcher, dependency pin, retained recipe, or historical changelog is changed. This is a deletion-only cleanup against 8ec35e0e97c9f0846a132a4eb891b51e87d5e8b9, not a benchmark migration or model retirement.

Audit boundaries

  • Audited all 765 YAMLs in the single-node and multi-node srt-slurm recipe trees, including four shared EPLB assets.
  • None of the deleted recipe-relative paths occurs in any other tracked text file outside historical perf-changelog.yaml. Checked active and deprecated master configs, launchers, workflows, scripts, tests, docs, and other recipes. Single-node selection uses explicit srt-recipe; multi-node selection uses CONFIG_FILE.
  • Retained all 413 recipes referenced by active configs, all 328 referenced only by deprecated configs, and the four shared load-balancer YAMLs used through recipe bind mounts.
  • Compared complete parsed YAML against the pinned NVIDIA/srt-slurm checkout (8dace5f9596907a5075bf056251563b2e9563e7d) and NVIDIA/srt-slurm-recipes (2018b1dda0a2792edcb81fe7bae5293d25fd4ab2). Found no exact parsed-content duplicates to remove. The pinned runtime does not ship a historical recipes/ tree, and InferenceX stages its local recipes explicitly; an upstream recipe with similar settings is not a safe drop-in replacement.
  • This PR does not overlap the recipe-selector/tooling work in srt-slurm: resolve CONFIG_FILE recipe selectors in launchers and the matrix #3390 or consolidation work in srt-slurm: consolidate sibling recipes into named override variants #3391, and does not delete configs/deprecated as proposed in [Klaud Cold] Remove pointers to the deleted single-node bash folders; delete configs/deprecated and amd_utils/deprecated #3463.

Validation

  • git diff --check: passed.

  • All 413 unique active master-config recipe paths resolve. Deletion adds zero missing paths to active or deprecated configs. The deprecated NVIDIA config already has 398 missing legacy paths; those are unchanged.

  • Ran the real generate_config_matrix implementation for each of 199 active config keys before and after deletion in both default and all-eval modes. Outputs are byte-identical, including errors:

    • 197 successful configs: 1,713 default rows and 1,652 all-eval rows.
    • Two pre-existing failures: qwen3.5-fp4-b300-sglang-agentic-power-ab and qwen3.5-fp8-b300-sglang-agentic-power-ab reference cluster:b300-nv, whose available-cpu-dram-mib metadata is absent. An unfiltered full-sweep fails for this same reason before any edits.
  • Local matrix and SRT unit tests: 447 passed, 1 failed:

    python -m pytest utils/matrix_logic/ utils/test_srt_single_node.py \
      utils/test_synthetic_acceptance.py utils/test_srt_cluster_config.py -q --tb=short

    The failure is test_pool_launcher_stages_artifacts_and_propagates_failure[h100-dgxc-slurm-none]: its environment lacks RUNNER_NAME. Reproduced the same failure in a clean worktree of the base commit. No test/source changes are included to conceal it.

  • No GPU sweep or runtime-performance claim. No performance changelog entry: only unreferenced files are removed; no executed recipe or generated matrix changes.

AI model disclosure

GPT 6 Astra, via Perplexity Computer: repository audit, deletion selection, implementation, validation, and PR preparation. No delegated agents or additional models.

中文

概述

删除 20 个未被引用的 srt-slurm 配方 YAML,共减少 2,632 行:

  • Kimi-K2.5 / TRT-LLM / B200 FP4 / 8k1k:11 个。
  • MiniMax-M3 / vLLM / GB200 FP4 / AgentX:6 个。
  • Qwen3.5 / SGLang / GB200 FP4 / AgentX:1 个。
  • Kimi-K3 和 MiniMax-M3 / vLLM / MI355X FP4 MTP / AgentX:各 1 个。

本 PR 基于 8ec35e0e97c9f0846a132a4eb891b51e87d5e8b9,仅删除未引用文件,不修改主配置、启动器、依赖版本、保留配方或历史性能变更日志,不迁移基准测试,也不退役模型。

审计范围

检查了单节点和多节点目录中的全部 765 个 YAML,包括 4 个共享 EPLB 文件。除历史 perf-changelog.yaml 外,其他受版本控制的文本文件均未引用这些待删除配方的相对路径;检查覆盖活动及已弃用配置、启动器、工作流、脚本、测试、文档和其他配方。单节点通过显式 srt-recipe 选择配方,多节点使用 CONFIG_FILE。

保留活动配置引用的 413 个配方、仅已弃用配置引用的 328 个配方,以及被其他配方挂载使用的 4 个负载均衡 YAML。与 NVIDIA/srt-slurm 固定版本 8dace5f9596907a5075bf056251563b2e9563e7d 和 NVIDIA/srt-slurm-recipes 的 2018b1dda0a2792edcb81fe7bae5293d25fd4ab2 比较后,未发现解析内容完全相同的配方。当前固定运行时不包含历史 recipes/ 目录,不能因为上游存在相似配方就删除仍被引用的本地文件。

本 PR 不涉及 #3390 的选择器工具或 #3391 的配方合并,也不执行 #3463 提出的 configs/deprecated 删除。

验证

  • git diff --check 通过;活动配置引用的 413 个唯一配方路径全部存在。活动及已弃用配置均未新增缺失引用;已弃用 NVIDIA 配置原有的 398 个缺失历史路径保持不变。
  • 删除前后,使用真实 generate_config_matrix 对 199 个活动配置逐一生成默认及全部评估模式的矩阵,输出及错误逐字节一致。197 个配置成功,分别生成 1,713 行和 1,652 行。
  • qwen3.5-fp4-b300-sglang-agentic-power-ab 和 qwen3.5-fp8-b300-sglang-agentic-power-ab 在修改前就因 cluster:b300-nv 缺少 available-cpu-dram-mib 而失败;完整 sweep 生成也受此影响。
  • 本地矩阵与 SRT 测试:447 通过、1 失败。失败项为 test_pool_launcher_stages_artifacts_and_propagates_failure[h100-dgxc-slurm-none],其测试环境缺少 RUNNER_NAME;已在基础提交的干净 worktree 复现。未修改测试或源码来掩盖问题。
  • 未执行 GPU sweep,不声称验证运行时性能。由于只删除未引用文件,且执行配方与生成矩阵均不变,不新增性能变更日志条目。

AI 模型披露

GPT 6 Astra,通过 Perplexity Computer 完成仓库审计、删除范围选择、实现、验证和 PR 准备。未委派其他代理,也未使用其他模型。

Remove 20 unused recipe files without changing active or deprecated config references. Generated matrices are unchanged.\n\n删除 20 个未引用的 srt-slurm 配方文件,不修改活动或已弃用配置的引用,生成矩阵保持不变。
@github-actions

Copy link
Copy Markdown
Contributor

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 claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

Looks good, straightforward deletion-only cleanup. Reviewed: confirmed all 20 deleted recipe basenames/paths have zero remaining references in nvidia-master.yaml, amd-master.yaml, their deprecated counterparts, and perf-changelog.yaml; checked the kimik3/minimaxm3 single-node MI355X deletions specifically since amd-master.yaml still references sibling atom/mi355x-fp4-mtp/agentic.yaml paths (a different, unaffected variant); no code, launcher, or dependency files touched.

Extended reasoning...

Change is a pure deletion of 20 unreferenced srt-slurm recipe YAMLs (2632 lines removed, 0 added), no code/config/changelog files touched. Verified by grepping the deleted files' basenames and directory paths against nvidia-master.yaml, amd-master.yaml, both deprecated master configs, and perf-changelog.yaml, finding zero remaining references. No security-sensitive surface is touched; the only ambiguity (single-node MI355X recipes) resolved cleanly since active configs point at a differently-named atom/ variant, not the deleted vllm/ files.

…cipes

# Conflicts:
#	benchmarks/single_node/srt-slurm-recipes/kimik3/vllm/mi355x-fp4-mtp/agentic.yaml
@functionstackx

Copy link
Copy Markdown
Collaborator

Kept 9 of the YAMLs this PR deleted (restored from main), per maintainer request, and merged main in to clear the conflict with #3463:

  • benchmarks/multi_node/srt-slurm-recipes/minimaxm3/vllm/gb200-fp4/agentx/ (6 files)
  • benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/gb200-fp4/agentx/agg-tp4-mtp-hicache.yaml
  • benchmarks/single_node/srt-slurm-recipes/kimik3/vllm/mi355x-fp4-mtp/agentic.yaml and .../minimaxm3/vllm/mi355x-fp4-mtp/agentic.yaml

The PR now removes only the 11 kimik2.5/trtllm/b200-fp4/8k1k disagg recipes, so the title's "20" count is out of date.

@functionstackx
functionstackx merged commit 6429c33 into main Sep 26, 2026
1 check passed
@functionstackx
functionstackx deleted the cleanup/unused-srt-recipes branch September 26, 2026 06:57
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

2 participants