[NV][AgentX] Add glm5.1-fp8-b200-tilert-agentic (TileRT PD disagg) / [NV][AgentX] 新增 glm5.1-fp8-b200-tilert-agentic（TileRT PD 分离）

[CrimsonDump]

Description

Add the GLM-5.1-FP8 B200 AgentX (agentic-coding) submission on TileRT PD disaggregation as one
new master-config key and one new multi-node recipe, and pin TileRT to 0.1.5.post3.

• glm5.1-fp8-b200-tilert-agentic — vLLM prefill + TileRT decode, 1P1D with TP8 on each side
  and nixl KV transfer, driven by the new recipe
  benchmarks/multi_node/agentic/glm5.1_fp8_b200_tilert-disagg.sh. The fixed-sequence-length
  recipe for the same topology already exists; this is its agentic sibling and reuses the same
  tilert_utils/ entry points.
• A separate config key rather than a second scenario on glm5.1-fp8-b200-tilert, because
  agentic master configs must declare a cluster: runner. configs/runners.yaml gains the
  cluster:b200-dgxc-slurm label mapping to b200-dgxc-slurm_7/8/9.
• Concurrency is the single point [1]. A TileRT decode engine serves one sequence at a time,
  so the trajectory count stays at 1 and the router queues the concurrent turns that an agentic
  session fans out into. kv-offloading: none — all KV state lives on the GPU and TileRT
  exposes no DRAM offload tier.
• runners/launch_b200-dgxc.sh honors SCENARIO_SUBDIR, so agentic-coding configs pick the
  recipe under multi_node/agentic/ instead of the one at the multi_node/ root.
• benchmarks/multi_node/tilert_utils/run_node.sh gains the agentic branch: an AIPerf trace
  replay per concurrency under LOGS/agentic/conc_<N>/, the router queue timeout, and a second
  --served-model-name so vLLM accepts the HF repo id that the replay sends as --model.

Why TileRT 0.1.5.post3

post3 carries two changes this submission depends on.

• PD overhead reduction. The decode loop moved from the prefill-phase entry point to the
  decode-phase one, with chained AR steps and a single bulk device-to-host copy of the AR
  buffer instead of a per-step synchronize and per-token readback. This closes most of the
  remaining gap to the non-PD engine reference.
• Request queueing for AgentX. A TileRT decode engine serves one sequence at a time, and an
  agentic session fans out into concurrent turns (parent plus sub-conversations). The router
  previously answered 429 as soon as every decode node was reserved. --queue-timeout <s>
  makes the request wait for a free node instead, which is what an AgentX replay needs: the
  burst is wider than the pool but short-lived. The default 0 keeps the fail-fast behaviour,
  so fixed-sequence-length runs are unaffected.

benchmarks/multi_node/tilert_utils/setup_deps.sh pins tilert==0.1.5.post3, and the router
version metadata on both TileRT master-config keys is bumped to match.

Validation

• utils/validate_perf_changelog.py against origin/main passes, with zero deleted lines.
• generate_sweep_configs.py test-config --config-keys glm5.1-fp8-b200-tilert-agentic --scenario-type agentic-coding --conc 1 generates exactly one job.
• pytest utils/matrix_logic — 231 matrix and validation tests pass.
• YAML parsing and bash -n pass on every touched file.
• End-to-end on two 8×B200 nodes: a 3600 s replay at concurrency 1 completed 539 requests
  with zero errors and submission_valid: true.

Open questions

Three things we cannot settle on our side.

1. Runner label. Agentic master configs must use a cluster:<name> runner, and no existing
cluster label covers the multi-node B200 machines (cluster:b200-dgxc is the ten single-node
b200-dgxc_* runners). This PR adds the cluster:b200-dgxc-slurm mapping to
configs/runners.yaml, but the label also has to exist on the runners themselves. Could you
add it to b200-dgxc-slurm_7 / _8 / _9 — the same alias pattern as cluster:mi355x-amds vs
mi355x, no new machines, the same three nodes b200-multinode already covers? We would also
add the matching hardware: entry — are the values the same as cluster:b200-dgxc
(available-cpu-dram-mib: 3095781, gpus-per-node: 8)?

2. No engine-side Prometheus metrics. TileRT exposes /metrics on neither the router nor
the decode engine; only the vLLM prefill does. The aggregate therefore carries
server_metrics.present: false, and 9 of the 12 metrics_plots.png panels render empty (KV
utilization, queue depth, prefix-cache hit rate, throughput over time, preemptions, and the
KV-offload panels — empty for us in any case, since kv-offloading: none). Client-side metrics
are unaffected: TTFT, TPOT, ITL, E2EL, interactivity, throughput per GPU, token counts and the
theoretical cache-hit rate all come from the AIPerf records. Is that acceptable for ingest and
the dashboard?

3. Corpus coverage. GLM-5.1's native context is 202752 positions. --max-context-length
drops a trace whenever its peak context exceeds that, so 175 of the 393 traces in
cc-traces-weka-062126-256k remain eligible — about 14.7% of the corpus's requests by count.
Is that acceptable as-is?

中文说明

新增 GLM-5.1-FP8 B200 AgentX（agentic-coding）在 TileRT PD 分离上的提交，包含一个新的主配置键
与一个新的多节点配方，并将 TileRT 固定到 0.1.5.post3。

• glm5.1-fp8-b200-tilert-agentic —— vLLM prefill + TileRT decode，1P1D、两侧各 TP8，KV 走
  nixl 传输，由新增配方 benchmarks/multi_node/agentic/glm5.1_fp8_b200_tilert-disagg.sh
  驱动。相同拓扑的定长配方已存在，本配方是其 agentic 版本，复用同一套 tilert_utils/ 入口。
• 采用独立的配置键而非在 glm5.1-fp8-b200-tilert 上追加场景，因为 agentic 主配置必须声明
  cluster: runner。configs/runners.yaml 新增 cluster:b200-dgxc-slurm label，映射到
  b200-dgxc-slurm_7/8/9。
• 并发取单点 [1]。TileRT 的 decode 引擎一次只服务一条序列，因此 trajectory 数保持 1，
  agentic 会话扇出的并发轮次由 router 排队。kv-offloading: none —— KV 状态全部驻留 GPU，
  TileRT 不提供 DRAM offload 层。
• runners/launch_b200-dgxc.sh 识别 SCENARIO_SUBDIR，使 agentic-coding 配置选中
  multi_node/agentic/ 下的配方，而非 multi_node/ 根目录那份。
• benchmarks/multi_node/tilert_utils/run_node.sh 新增 agentic 分支：按并发在
  LOGS/agentic/conc_<N>/ 下运行 AIPerf trace 回放、设置 router 排队超时，并追加第二个
  --served-model-name，使 vLLM 接受回放以 --model 发来的 HF repo id。

为什么需要 TileRT 0.1.5.post3

post3 带来本次提交依赖的两项改动。

• PD 开销优化。 decode 循环从 prefill 阶段入口改为 decode 阶段入口，配合链式 AR 与对 AR
  缓冲区的单次整块 D2H 拷贝，取代原先每步一次同步、每 token 一次回读。这基本抹平了与非 PD
  引擎参照之间的剩余差距。
• 面向 AgentX 的请求排队。 TileRT 的 decode 引擎一次只服务一条序列，而 agentic 会话会扇出
  成并发的多轮（父对话加子对话）。此前 router 在所有 decode 节点被占满时立即返回 429。
  --queue-timeout <秒> 改为让请求等待空闲节点，这正是 AgentX 回放所需：突发宽度大于池子但
  持续很短。默认值 0 保持原有的 fail-fast 行为，因此定长场景不受影响。

benchmarks/multi_node/tilert_utils/setup_deps.sh 将依赖固定为 tilert==0.1.5.post3，两个
TileRT 主配置键上的 router 版本元数据同步更新。

本地验证

• utils/validate_perf_changelog.py 对 origin/main 校验通过，删除行为 0。
• generate_sweep_configs.py test-config --config-keys glm5.1-fp8-b200-tilert-agentic --scenario-type agentic-coding --conc 1 恰好生成 1 个 job。
• pytest utils/matrix_logic —— 231 项矩阵与校验测试全部通过。
• 所有改动文件的 YAML 解析与 bash -n 均通过。
• 双节点 8×B200 真机端到端：并发 1 下的 3600 秒回放完成 539 个请求、零错误，
  submission_valid: true。

待确认问题

以下三点我们这边无法自行决定。

1. Runner label。 agentic 主配置必须使用 cluster:<name> runner，而现有的 cluster label
没有一个覆盖多节点 B200 机器（cluster:b200-dgxc 指的是 b200-dgxc_* 那十台单机）。本 PR 已在
configs/runners.yaml 中新增 cluster:b200-dgxc-slurm 映射，但该 label 同时需要真实存在于
runner 上。能否为 b200-dgxc-slurm_7 / _8 / _9 添加此 label —— 与 cluster:mi355x-amds 对
mi355x 是同一种别名做法，不需要新机器，就是 b200-multinode 已经在用的这三台。我们也会补上
对应的 hardware: 条目 —— 取值是否与 cluster:b200-dgxc 相同（available-cpu-dram-mib: 3095781、gpus-per-node: 8）？

2. 引擎侧没有 Prometheus 指标。 TileRT 的 router 与 decode 引擎均不暴露 /metrics，只有
vLLM prefill 有。因此聚合结果会是 server_metrics.present: false，metrics_plots.png 的 12 个
面板中会有 9 个为空（KV 利用率、队列深度、prefix cache 命中率、吞吐时序、preemption，以及那几个
KV offload 面板 —— 我们 kv-offloading: none，本就应为空）。客户端指标不受影响：TTFT、TPOT、
ITL、E2EL、interactivity、每 GPU 吞吐、token 计数与理想缓存命中率均来自 AIPerf 自身记录。
这对 ingest 与 dashboard 可接受吗？

3. 语料覆盖率。 GLM-5.1 的原生上下文为 202752。--max-context-length 按整条 trace 的峰值
上下文丢弃，因此 cc-traces-weka-062126-256k 的 393 条中仅剩 175 条可用，按请求数计约覆盖语料
的 14.7%。这样可以接受吗？

Related Issue

N/A

Type of Change

• New feature
• Configuration change

Checklist

• I have tested my changes locally
• 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
• Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /reuse-sweep-run on this PR