[Power] Wire H200 multinode AgentX power / 接入 H200 多节点 AgentX 功耗链路

[edwingao28]

What changed

• Pin the GLM-5.2 H200 AgentX power lane to the immutable fork producer e5c837f06a362dc888dfea2ee588e9f19c298270 from srt-slurm #2.
• Add recipe-scoped dcgm-power detection, lane guards, DCGM exporter provisioning, container injection, and producer/exporter provenance.
• Emit protected formal measurement windows for every AgentX concurrency and join the central power package into each aggregate.
• Preserve schema-v2 whole-deployment joules and prefill/decode role watts while failing closed on invalid or incomplete telemetry.
• Keep non-power H200 clone routes and single-node AgentX behavior unchanged.

Roadmap: #2681

Validation

• CI-equivalent power and result-processing suite: 236 passed.
• H200 launcher bash -n and git diff --check passed.
• Fork prerequisite focused suite: 268 passed; Ruff check and format check passed.
• The producer pin is asserted by both launcher rev-parse validation and the H200 contract test.
• H200 strict hardware revalidation on this PR head and producer e5c837f06a362dc888dfea2ee588e9f19c298270: run 32324249431, benchmark job 96292249318.
  • required: true, publication_valid: true, 32/32 GPUs, 16 prefill + 16 decode, no validation failures.
  • Aggregate schema v2 and power_valid: 1; 292 successful requests, 15 warmup-only drops, zero error drops.
  • 273 W/chip average, 246 W/chip prefill, 300 W/chip decode, 67.62 J/output token, and 37,103 J/successful query.
  • Dashboard overlay renders the H200 measured-power and J/query rows.

Dependencies

• Fork prerequisite: edwingao28/srt-slurm #2
• Recipe rollout: InferenceX #2684

中文说明

变更内容

• 将 GLM-5.2 H200 AgentX 功耗链路固定到 srt-slurm #2 的不可变 producer e5c837f06a362dc888dfea2ee588e9f19c298270。
• 增加按 recipe 生效的 dcgm-power 检测、lane 守卫、DCGM exporter 准备、容器映射注入及 producer/exporter provenance 留档。
• 为每个 AgentX 并发点写入受保护的正式测量窗口，并把集中式功耗包汇入对应 aggregate。
• 保持 schema v2 的全部署焦耳口径和预填充/解码角色功耗；遥测无效或不完整时 fail closed。
• 保持非功耗 H200 clone 路径与单节点 AgentX 行为不变。

验证

• CI 同款功耗与结果处理测试：236 个通过。
• H200 launcher bash -n 与 git diff --check 通过。
• fork 前置相关测试：268 个通过；Ruff check 与 format check 通过。
• launcher 的 rev-parse 断言及 H200 contract test 会同时校验 producer pin。
• 已在本 PR head 与 producer e5c837f06a362dc888dfea2ee588e9f19c298270 上完成 H200 strict 真机复验：run 32324249431，benchmark job 96292249318。
  • required: true、publication_valid: true、32/32 GPU（16 张预填充 + 16 张解码），validation failures 为空。
  • 聚合结果为 schema v2、power_valid: 1；292 个成功请求、15 个仅 warmup drop、error drop 为 0。
  • 平均 273 W/chip、预填充 246 W/chip、解码 300 W/chip、67.62 J/output token、37,103 J/successful query。
  • Dashboard overlay 已渲染 H200 measured-power 与 J/query 行。

依赖关系

• fork 前置：edwingao28/srt-slurm #2
• recipe rollout：InferenceX #2684