[Power] Require GLM-5.2 and add DSV4 H200 telemetry / 强制 GLM-5.2 并新增 DSV4 H200 功耗采集

[edwingao28]

Summary

• Enable H200 dcgm-power telemetry for the GLM-5.2 FP8 2P2D AgentX recipe and fail closed with required: true.
• Enable the DSV4 FP8 aggregate H200 recipe in optional mode with required: false; aggregate data publishes whole-deployment energy metrics and intentionally omits prefill/decode role-local axes.
• Pin the power producer to immutable fork commit e5c837f06a362dc888dfea2ee588e9f19c298270 and preserve its SHA plus the exporter image digest in the audit bundle.
• Forward the manual require-power input through both workflow layers so a strict dispatch cannot silently run with REQUIRE_POWER=0.
• Keep Kimi-K3 deferred because it uses a separate vLLM/external-fork launcher path.

This draft is stacked on #2683 and depends on its H200 launcher, custom-window, and central aggregation infrastructure. Until #2683 merges, this PR intentionally shows both commits against main.

Roadmap: #2681

Validation

• CI-equivalent power/process-result suite: 241 passed.
• Current-pin H200 strict hardware run: benchmark job 96292249318; dashboard overlay.
  • Exact producer SHA e5c837f06a362dc888dfea2ee588e9f19c298270; exporter SHA-256 dd64a6be....
  • 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.
  • Dashboard values include 273 W/chip average and 37,103 J/successful query; artifacts also contain 246 W/chip prefill, 300 W/chip decode, and 67.62 J/output token.
  • Disaggregated J/token values use whole-deployment energy/token accounting; role watts and role-local energy fields remain separately labeled.

GitHub marks the workflow container as failed even though every visible benchmark, collection, and success-rate job succeeded and no failed job is present. The artifact audit above is the authoritative power verdict.

• Earlier DSV4 aggregate hardware run against producer a1b8c7af...: benchmark job 96258235915; dashboard overlay.
  • required: false, publication_valid: true, 8/8 GPUs, schema v2 and power_valid: 1.
  • Dashboard values include average 405 W/chip, 13.6 J/output token, 12,696 J/query, and 57.9% TDP.
  • Prefill/decode axes correctly show no data for the aggregate topology.

The current strict GLM run closes the finalized producer exact-SHA gate. DSV4 remains optional and retains its earlier recipe-level H200 evidence.

中文说明

• 为 GLM-5.2 FP8 H200 2P2D AgentX recipe 启用 dcgm-power，并设为 required: true，功耗采集异常时按失败关闭。
• 为 DSV4 FP8 H200 聚合式 recipe 启用可选功耗采集，保持 required: false；发布全部署口径的能耗指标，并按设计不生成预填充/解码 role-local 轴。
• 将功耗 producer 固定到不可变 fork commit e5c837f06a362dc888dfea2ee588e9f19c298270，并在审计产物中保留 producer SHA 与 exporter 镜像摘要。
• 在两层工作流中完整透传手动 require-power 输入，避免 strict dispatch 实际以 REQUIRE_POWER=0 静默运行。
• Kimi-K3 继续暂缓；它使用独立的 vLLM/外部 fork 启动路径，需要单独评估。

本 draft PR stacked 在 #2683 之上，依赖其中的 H200 启动器、custom window 与集中式聚合基础设施。在 #2683 合入前，本 PR 相对 main 会暂时显示两个提交。

验证结果

• CI 同款功耗与结果处理测试：241 个全部通过。
• 当前 pin 的 H200 strict 真机运行：benchmark job 96292249318；dashboard overlay。
  • producer SHA 精确为 e5c837f06a362dc888dfea2ee588e9f19c298270；exporter SHA-256 为 dd64a6be...。
  • 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。
  • Dashboard 展示平均 273 W/chip 与 37,103 J/successful query；产物还包含预填充 246 W/chip、解码 300 W/chip、67.62 J/output token。
  • 分离式 J/token 使用全部署能耗/Token 口径；role watts 与 role-local 能耗字段保持独立标注。

GitHub 将 workflow 容器级结论标为 failed，但所有可见的 benchmark、结果收集与 success-rate job 均成功，也不存在 failed job。以上产物审计结果是权威功耗判定。

• 较早的 DSV4 aggregate 硬件运行使用 producer a1b8c7af...：benchmark job 96258235915；dashboard overlay。
  • required: false、publication_valid: true、8/8 GPU、schema v2、power_valid: 1。
  • Dashboard 展示：平均 405 W/chip、13.6 J/output token、12,696 J/query、57.9% TDP。
  • 聚合式拓扑的预填充/解码轴正确显示无数据。

当前 GLM strict 运行关闭了最终 producer 的 exact-SHA gate。DSV4 保持 optional，并沿用此前的 recipe 级 H200 硬件证据。