v2 rework

[adibarra]

Description

Start the full-YAML rework with SRT-Slurm as the execution component. This draft is the shared integration PR for the work tracked in #2382.

The initial change points utils/srt-slurm at SemiAnalysisAI/srt-slurm and documents remote synchronization for existing checkouts. It retains the exact runtime pin 984180e5b8755aef85e9995048b5a16cb5336bce (upstream v2.2.1), so this first commit changes the dependency source without changing benchmark execution.

Work remaining

• Sync the fork with upstream and select a validated integration revision. At creation, fork main is two commits behind upstream 5df10653f70b8076b64c8cd261389d27cf4dbeb0; changing the remote alone does not perform that sync.
• Make recipe selection a first-class master-config field instead of additional-settings.CONFIG_FILE.
• Reuse native SRT parameterization/overrides to share recipe settings across concurrency points.
• Migrate active recipes, including aggregated serving, while keeping cluster configuration separate from benchmark settings. Deprecated configs are excluded; their archival is already merged in Consolidate deprecated configs and remove retired runtime support #3349.
• Reconcile the existing H100 pilot ([wip ai slop] Implement Phase 1 native H100 pilot / 实现阶段 1 原生 H100 试点 #3299 and [wip ai slop] Add prepared H100 single-job execution / 新增 H100 预备式单任务执行 srt-slurm#21), validate the migrated runtime and result contracts, and return to the upstream dependency when the required functionality is available there.

Validation

• git diff --check passed.
• Initialized utils/srt-slurm from the SemiAnalysisAI fork and verified checkout of the unchanged pinned commit.
• GPU qualification remains part of the migration work. No runtime upgrade or recipe change is included in this starter commit, so no performance-changelog entry is needed yet.

Related Issue

#2382

Type of Change

• Bug fix
• New feature
• Configuration change
• Documentation update
• Other (please describe)

Checklist

• I have tested my changes locally
• I have updated documentation if necessary
• For every change that can affect benchmark performance and every recipe addition or modification, I have appended a new entry to the physical end of perf-changelog.yaml and have not edited historical entries. Not applicable to the starter commit; required for subsequent runtime/recipe changes.
• Before merging via reuse, an authorized maintainer has commented /use <run_id> (or the legacy /reuse-sweep-run) after a successful full sweep with evals. Pending; this is a draft.

中文

说明

以 SRT-Slurm 作为执行组件，启动全 YAML 重构。本草稿作为 #2382 所跟踪工作的共享集成 PR。

初始变更将 utils/srt-slurm 的来源改为 SemiAnalysisAI/srt-slurm，并更新现有检出目录同步远程地址的说明。运行时仍固定在上方列出的上游 v2.2.1 提交，因此本次仅改变依赖来源，不改变基准执行行为。

后续工作

• 将分支仓库与上游同步，并选定完成验证的集成版本。创建 PR 时，分支仓库的 main 比上方列出的上游提交落后两个提交；修改远程地址本身不会完成同步。
• 将配方选择提升为主配置的一等字段，替代 additional-settings.CONFIG_FILE。
• 使用原生 SRT 参数化和覆盖能力，让不同并发点共享配方设置。
• 迁移仍在使用的配方，包括聚合式部署，并保持集群配置与基准设置分离。不迁移已弃用配置；其归档已通过 Consolidate deprecated configs and remove retired runtime support #3349 合并。
• 与现有 H100 试点（[wip ai slop] Implement Phase 1 native H100 pilot / 实现阶段 1 原生 H100 试点 #3299 和 [wip ai slop] Add prepared H100 single-job execution / 新增 H100 预备式单任务执行 srt-slurm#21）协调，验证迁移后的运行时与结果契约；待上游具备所需能力后切回上游依赖。

验证与状态

git diff --check 通过；已从 SemiAnalysisAI 分支仓库初始化子模块，并确认检出原有固定提交。GPU 验收留待迁移阶段完成。本次初始提交不升级运行时或修改配方，因此暂不需要性能变更日志；后续涉及运行时或配方的变更必须追加日志。关联 issue 为 #2382；变更类型为配置与文档更新。本 PR 保持草稿，完整 sweep、eval 和复用合并条件尚未满足。