[Power] Add optional B200 and B300 multinode telemetry / 添加可选 B200 和 B300 多节点功耗采集

[edwingao28]

What changed

• Port the H200-validated power launcher wiring to launch_b200-dgxc.sh, launch_b200-nscale-slurm.sh, and launch_b300-nv.sh.
• Enable dcgm-power with required: false on 37 fixed-sequence recipes: 18 SGLang and 19 vLLM, including 15 dedicated-infrastructure topologies.
• Route only power-enabled recipes through immutable producer e5c837f06a362dc888dfea2ee588e9f19c298270; preserve every existing non-power clone revision and keep Kimi-K3 and AgentX recipes unchanged.
• Provision the DCGM exporter and record producer/exporter provenance in the power artifact.
• Use fail-closed vLLM port ranges and pre-submit topology validation from the fork prerequisite.
• Add dynamic contract and CI path coverage for all selected recipes.

Stack

• This PR is stacked on InferenceX #2683 at 8073ca3fe3982146ca93ad6687b3c3aac8703f5c.
• Producer prerequisite: edwingao28/srt-slurm #2 at e5c837f06a362dc888dfea2ee588e9f19c298270.
• Review the Phase 3 commit after the PR-A commits in this draft; rebase onto main after PR-A merges.

Validation

• CI-equivalent power and result-processing suite: 244 passed.
• Official performance changelog validator passed against the PR-A base.
• All 37 recipes load and pass complete topology/port preflight; the largest topology has 32 processes.
• Three launchers pass bash -n; git diff --check passed.

Hardware status

No B200, B300, GB200, or GB300 job was dispatched. The current validation environment is limited to H100/H200, so all 37 recipe opt-ins remain required: false. Before rollout, an authorized environment must verify node port availability, exporter reachability, numa-bind vLLM paths, schema-v2 artifacts, role watts, whole-deployment joules, and dashboard points.

中文说明

变更内容

• 将已经过 H200 验证的功耗 launcher 接线移植到 launch_b200-dgxc.sh、launch_b200-nscale-slurm.sh 和 launch_b300-nv.sh。
• 为 37 个固定序列 recipe 以 required: false 启用 dcgm-power：其中 18 个 SGLang、19 个 vLLM，并包含 15 个独立基础设施节点拓扑。
• 仅让启用功耗的 recipe 使用不可变 producer e5c837f06a362dc888dfea2ee588e9f19c298270；保留全部现有非功耗 clone revision，且不修改 Kimi-K3 和 AgentX recipe。
• 准备 DCGM exporter，并在功耗产物中记录 producer/exporter provenance。
• 使用 fork 前置提供的 fail-closed vLLM 端口区间与提交前拓扑校验。
• 为全部选中 recipe 增加动态 contract 与 CI path 覆盖。

依赖关系

• 本 PR stacked 在 InferenceX #2683 的 8073ca3fe3982146ca93ad6687b3c3aac8703f5c 上。
• producer 前置为 edwingao28/srt-slurm #2 的 e5c837f06a362dc888dfea2ee588e9f19c298270。
• 本 Draft 中请在 PR-A commits 之后 review Phase 3 commit；PR-A 合入后再 rebase 到 main。

验证

• CI 同款功耗与结果处理测试：244 个通过。
• 官方 performance changelog validator 基于 PR-A base 验证通过。
• 37 个 recipe 全部可以加载并通过完整 topology/port preflight；最大拓扑包含 32 个 process。
• 三条 launcher 均通过 bash -n；git diff --check 通过。

硬件状态

本 PR 未派发任何 B200、B300、GB200 或 GB300 作业。当前可用验证硬件仅限 H100/H200，因此 37 个 recipe 全部保持 required: false。正式 rollout 前，需要在获准环境中验证节点端口占用、exporter 可达性、numa-bind vLLM 路径、schema v2 产物、角色功耗、全部署焦耳口径及 dashboard 出点。