[Klaud Cold] Retire GLM-5.1 and its B200 TileRT config / [Klaud Cold] 退役 GLM-5.1 及其 B200 TileRT 配置

[functionstackx]

Description

Retires GLM-5.1 completely. glm5.1-fp8-b200-tilert was the last GLM-5.1 config and the last Single-turn 1k1k config, so this also ends 1k1k coverage.

Removed

• glm5.1-fp8-b200-tilert from nvidia-master.yaml (1k1k + 8k1k, B200 Nscale, vLLM prefill + TileRT 0.1.5 decode)
• benchmarks/multi_node/srt-slurm-recipes/glm5.1/ (both recipes) and srt-slurm-recipes/configs/tilert-b200-setup.sh
• B200 Nscale launcher rows used only by this config:
  • policy.NATIVE_SRT_LANES GLM-5.1 TileRT match
  • power.py GLM-5.1 fixed-sequence DCGM rule
  • models.OVERRIDES["b200-nscale"] (GLM-5.1-FP8@shared)
  • TILERT_ENV (its only row was b200-nscale's UCX settings); runtime_env now just layers the cluster env and the given settings
  • the /tilert_weights lane mount
• runners.yaml b200-nscale: GLM-5.1-FP8 / GLM-5.1-FP8@shared entries and the now-unused shared-models / tilert-weights volumes
• glm5.1 gsm8k eval threshold
• Docs: the "TileRT fixed-sequence recipes" section (EN/ZH), the "use 1k1k only for GLM-5.1" notes in configuration-procedures / testing (EN/ZH), the stale tilert-weights volume name in CONFIGS.md, and the GLM-5.1 exception in the claude.yml agent prompts and .claude/commands/add-model-hardware.md
• MODELS.md / MODELS_zh.md: the 1k1k scenario row and the GLM-5 / GLM-5.1 support-matrix row now record the retirement, plus a dated retirement note

Kept

• Generic TileRT support (driver, synthetic acceptance, tilert framework handling), still used by glm5.3-fp8-mi355x-tilert-agentic. The TileRT driver test moved from the GLM-5.1 B200 lane to the GLM-5.3 MI355X lane instead of being deleted.
• srt_fixed_sequence.sh, shared with the DSR1 B200 recipes
• Historical text: README news items, the dated 2026-09-21 parity audit, and perf-changelog.yaml
• operatorx/testlists/*.json GLM-5.1 kernel shapes (OperatorX microbenchmarks, not InferenceX configs)

No perf-changelog entry, matching earlier retirement PRs (#3618, #3595, #3467). No benchmark is affected, so no sweep label.

Validation

• infx.matrix.generate full-sweep over both master configs: 1715 rows, 0 glm5.1, 0 ISL-1024
• pytest infx/tests (serial, srt-slurm submodule initialized): all pass apart from results/test_collect_eval_results.py, which fails to import because a results extra isn't installed locally. It fails the same way on main.

AI model disclosure

• Model/version: claude-opus-5-5[1m] (Claude Opus 5.5, 1M context), via Claude Code
• Role: found every GLM-5.1 reference, made the deletions and test/doc updates, ran local validation, and drafted this PR

Type of Change

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

Checklist

• I have completed the AI model disclosure and kept it current
• 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 inferencex-e2e/perf-changelog.yaml and have not edited historical entries (N/A: retirement only)

中文

改动说明

彻底退役 GLM-5.1。glm5.1-fp8-b200-tilert 是最后一个 GLM-5.1 配置，也是最后一个单轮 1k1k 配置，因此本 PR 同时结束了 1k1k 覆盖。

删除

• nvidia-master.yaml 中的 glm5.1-fp8-b200-tilert（1k1k + 8k1k，B200 Nscale，vLLM prefill + TileRT 0.1.5 decode）
• benchmarks/multi_node/srt-slurm-recipes/glm5.1/（两个配方）及 srt-slurm-recipes/configs/tilert-b200-setup.sh
• 仅供该配置使用的 B200 Nscale 启动器条目：native lane 匹配、DCGM 功耗规则、GLM-5.1-FP8@shared checkpoint 覆盖、TILERT_ENV（唯一一行是 b200-nscale 的 UCX 设置；runtime_env 现在只叠加集群 env 与传入设置）、/tilert_weights 挂载
• runners.yaml b200-nscale 中的 GLM-5.1 checkpoint 条目及不再使用的 shared-models / tilert-weights 卷
• glm5.1 的 gsm8k eval 阈值
• 文档：「TileRT 固定序列长度配方」章节（中英）、「1k1k 仅用于 GLM-5.1」的说明、CONFIGS.md 中过时的 tilert-weights 卷名，以及 claude.yml 智能体提示和 .claude/commands/add-model-hardware.md 中的 GLM-5.1 例外
• MODELS.md / MODELS_zh.md：1k1k 场景行与 GLM-5 / GLM-5.1 支持矩阵行改为记录退役，并新增带日期的退役说明

保留

• 通用 TileRT 支持，仍由 glm5.3-fp8-mi355x-tilert-agentic 使用；TileRT driver 测试从 GLM-5.1 B200 lane 移至 GLM-5.3 MI355X lane，而非删除
• 与 DSR1 B200 配方共用的 srt_fixed_sequence.sh
• 历史文本：README 新闻条目、2026-09-21 一致性核查、perf-changelog.yaml
• operatorx/testlists/*.json 中的 GLM-5.1 kernel 形状（OperatorX 微基准，不属于 InferenceX 配置）

与此前的退役 PR（#3618、#3595、#3467）一致，不新增 perf-changelog 条目；不影响任何基准测试，因此不加 sweep 标签。

验证

• 对两个主配置运行 infx.matrix.generate full-sweep：1715 行，glm5.1 0 行，ISL-1024 0 行
• pytest infx/tests（串行，已初始化 srt-slurm 子模块）：除 results/test_collect_eval_results.py 外全部通过；该文件因本地未安装 results 可选依赖而无法 import，在 main 上同样失败

AI 模型使用说明

• 模型/版本：claude-opus-5-5[1m]（Claude Opus 5.5，1M 上下文），经 Claude Code 使用
• 工作内容：查找全部 GLM-5.1 引用，完成删除及测试/文档更新，运行本地验证，起草本 PR

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Looks good, straightforward and thorough retirement change — correctly scoped and consistent across code, config, tests, and docs.

What was reviewed:

• Launch policy/driver removals (policy.py NATIVE_SRT_LANES/TILERT_ENV, lanes.py mount, models.py override, power.py rule) — confirmed each was only reachable via the deleted glm5.1 b200-nscale TileRT config, and generic TileRT support (used by glm5.3-fp8-mi355x-tilert-agentic) is untouched.
• runtime_env simplification after dropping TILERT_ENV — verified other TileRT callers (drivers/srt/run.py, config.py, __init__.py) don't depend on the removed cluster-specific UCX env.
• Config/runner removals (nvidia-master.yaml, runners.yaml, thresholds.yaml) and deleted recipe files — grepped the repo for dangling references (shared-models, tilert-weights, tilert-b200-setup, active glm5.1 configs) and found none outside historical perf-changelog.yaml entries, which the PR intentionally leaves untouched.
• Test updates in test_launch_policy.py, test_srt_driver.py, test_srt_policy.py — confirmed the TileRT driver test was moved (not deleted) to the GLM-5.3 MI355X lane, consistent with the PR description.

Extended reasoning...

The PR removes the last GLM-5.1 config and all code/config/docs paths exclusively serving it, across launch policy (policy.py, lanes.py, models.py, power.py), config files (nvidia-master.yaml, runners.yaml, thresholds.yaml), deleted recipe files, and docs/tests. No security-sensitive surface (auth, crypto, injection) is touched — this is CI/benchmark orchestration config and test code. I traced each removed code branch to confirm it was only reachable by the deleted glm5.1 config and verified via grep that no dangling references to removed volumes/scripts/configs remain outside intentionally-preserved historical changelog entries; tests were moved rather than deleted to cover the retained generic TileRT path (glm5.3-fp8-mi355x-tilert-agentic).

This review covers commit 0099e53, which is no longer the latest commit on this pull request; later commits are not covered by it.