Skip to content

[Klaud Cold] Update GLM-5.2 FP8 MI355X SGLang AgentX MTP image to v0.5.22 / [Klaud Cold] 将 GLM-5.2 FP8 MI355X SGLang AgentX MTP 镜像更新至 v0.5.22 - #4518

Closed
Klaud-Cold wants to merge 65 commits into
mainfrom
klaud/auto-5c762659ac355952-8c92c0a196b4eef1
Closed

Klaud-Cold wants to merge 65 commits into
mainfrom
klaud/auto-5c762659ac355952-8c92c0a196b4eef1

Conversation

@Klaud-Cold

@Klaud-Cold Klaud-Cold commented Oct 25, 2026 •

Copy link
Copy Markdown
Collaborator

Goal: Update SGLang image from lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260917 to lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11.
Baseline: 2026-09-19 · lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260917
AgentX · TP8/EP1 · Mean latency · Sources: API 1, API 2, API 3, API 4

Concurrency Total tok/s/GPU ↑ Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 2,799.42 8.3 1,194.29 5.95
2 2,784.99 10.78 874.37 6.43
4 4,012.14 14.57 841.78 6.82
8 7,042.11 26.78 746.79 8.73
10 8,172.44 33.19 988.41 10.19
12 7,710.19 32.25 1,221.74 12.08
16 9,244.63 36.48 2,326.62 23.02
Eval Score ↑ Samples
gsm8k/em_strict · c16 96.66% 1,319

Stop: Close this PR at any time to stop this candidate now; Klaud cancels its runs and deletes its branch.

中文

**目标:**将 SGLang 镜像从 lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260917 更新至 lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11。
**基线:**2026-09-19 · lmsysorg/sglang-rocm:v0.5.19-rocm720-mi35x-20260917
AgentX · TP8/EP1 · 平均延迟 · 来源: API 1, API 2, API 3, API 4;数值及异常说明见上表。
**停止:**随时关闭此 PR 即可立即停止该候选:Klaud 会取消其运行并删除其分支。

…0.5.22

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

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 run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
  • 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:**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目,并且恰好带有一个主标签时,才会运行扫描:full-sweep-fail-fast(强烈推荐;canary 加逐矩阵 fail-fast)、full-sweep-enabled(有 canary;矩阵任务在失败后继续运行)或 non-canary-full-sweep-enabled(无 canary,也无 fail-fast)。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
  • **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。

@Klaud-Cold

Klaud-Cold commented Oct 25, 2026 •

Copy link
Copy Markdown
Collaborator Author

Initial attempt · Passed · Run 40230541980 / attempt 1 · 2026-10-25 22:31 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · 88588aaf353a · AgentX · TP8/EP1 · Mean latency
Change: Set the master image to the official SGLang v0.5.22 ROCm 7.2 MI35X release, pinned by digest; source diff from main a9fb1c3238 (old image) to tag v0.5.22 keeps every recipe flag and env (tilelang DSA backends, glm47/glm45 parsers, EAGLE MTP, HiCache write_through_selective/page_first_direct, SGLANG_OPT_USE_TOPK_V2) with unchanged semantics, while AITER moves 4ad99832 to e7d2453f and v0.5.22 newly enables AITER allreduce fusion on decode for GlmMoeDsaForCausalLM on ROCm by default.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 14.61 (N/A) 936.21 (N/A) 5.12 (N/A)
8 42.79 (N/A) 626.02 (N/A) 8.11 (N/A)

Note: All rows: Δ N/A: no matched baseline.

Eval Score ↑ Samples
gsm8k/em_strict · c16 96.51% (-0.15 pp) 1,319 each

Next: Smoke on ref 60df019 passed (c1, c8 HiCache, zero request errors, gsm8k 96.51% at c16); append the perf-changelog entry and dispatch the full-sweep-fail-fast final sweep.

中文

初次尝试 · 已通过 · Run 40230541980 / attempt 1 · 2026-10-25 22:31 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · 88588aaf353a · AgentX · TP8/EP1 · 平均延迟
**变更:**将主镜像设为官方 SGLang v0.5.22 ROCm 7.2 MI35X 发布版并按 digest 固定;从旧镜像的 main a9fb1c3238 到 tag v0.5.22 的源码对比显示,所有 recipe 参数和环境变量(tilelang DSA 后端、glm47/glm45 解析器、EAGLE MTP、HiCache write_through_selective/page_first_direct、SGLANG_OPT_USE_TOPK_V2)均保留且语义不变,AITER 从 4ad99832 升级到 e7d2453f,且 v0.5.22 在 ROCm 上默认为 GlmMoeDsaForCausalLM 的 decode 启用 AITER allreduce fusion。实测数值及异常说明见上表。
**下一步:**ref 60df019 的 smoke 运行已通过(c1、c8 HiCache,零请求错误,c16 gsm8k 96.51%);接下来追加 perf-changelog 条目并启动 full-sweep-fail-fast 最终扫描。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 1 is waiting on 40230541980. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 1 个会话正在等待 40230541980。恢复流程会重跑被抢占的运行,并在运行结束后启动下一个会话。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

resumed · The session ended before finishing; session 2 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束;第 2 个会话将在其规划运行的新运行中继续。

… v0.5.22

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Klaud-Cold Klaud-Cold added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Oct 25, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Queued · Run 40240738227 / attempt 1 · 2026-10-25 22:34 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · f5c1fdf25ca4 · AgentX · TP8/EP1 · Mean latency
Change: Append the perf-changelog entry for the SGLang v0.5.22 image and run the full-sweep-fail-fast final sweep over every AgentX point.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 N/A N/A N/A
2 N/A N/A N/A
4 N/A N/A N/A
8 N/A N/A N/A
10 N/A N/A N/A
12 N/A N/A N/A
16 N/A N/A N/A

Note: All rows: pending.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.

Next: Wait for the queued final sweep (c1-c16 and the c16 gsm8k eval) to complete, then verify artifacts against the frozen baseline.

中文

最终完整 sweep · 排队中 · Run 40240738227 / attempt 1 · 2026-10-25 22:34 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · f5c1fdf25ca4 · AgentX · TP8/EP1 · 平均延迟
**变更:**为 SGLang v0.5.22 镜像追加 perf-changelog 条目,并对全部 AgentX 点运行 full-sweep-fail-fast 最终扫描。实测数值及异常说明见上表。
**下一步:**等待已排队的最终扫描(c1-c16 及 c16 gsm8k 评测)完成,然后对照冻结的基线核验产物。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 2 is waiting on 40240738227. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 2 个会话正在等待 40240738227。恢复流程会重跑被抢占的运行,并在运行结束后启动下一个会话。

@github-actions

github-actions Bot commented Oct 26, 2026 •

Copy link
Copy Markdown
Contributor

@github-actions github-actions Bot added klaud-parked Klaud parked until its queued or running benchmark runs end klaud-in-progress Klaud is working on this candidate or will resume it and removed klaud-parked Klaud parked until its queued or running benchmark runs end klaud-in-progress Klaud is working on this candidate or will resume it labels Oct 26, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

resumed · The session ended before finishing; session 3 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束;第 3 个会话将在其规划运行的新运行中继续。

@github-actions github-actions Bot added klaud-in-progress Klaud is working on this candidate or will resume it and removed klaud-parked Klaud parked until its queued or running benchmark runs end labels Oct 27, 2026
…355952-8c92c0a196b4eef1

# Conflicts:
#	inferencex-e2e/perf-changelog.yaml
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Failed · Run 40240738227 / attempt 4 · 2026-10-27 06:43 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · f5c1fdf25ca4 · AgentX · TP8/EP1 · Mean latency
Change: Final sweep on f5c1fdf after three preempted attempts; c10 failed in the mi355x-amds host_setup hook because node mia1-p01-g15 still held 23% VRAM from a prior job after the 15 min drain wait, before the SGLang server started, and fail-fast cancelled the other points.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 N/A N/A N/A
2 N/A N/A N/A
4 N/A N/A N/A
8 N/A N/A N/A
10 N/A N/A N/A
12 N/A N/A N/A
16 N/A N/A N/A

Note: c1, c2, c4, c8, c12, c16: cancelled.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.
Note: c10: failed.

Eval Score ↑ Samples
gsm8k/em_strict · c16 96.97% (+0.30 pp) 1,319 each

Next: This is a node-state infrastructure failure, not an image failure (gsm8k c16 still passed at 96.97%); main also appended perf-changelog entries, so merge main and rerun the final sweep on the new head without using a repair.

中文

最终完整 sweep · 失败 · Run 40240738227 / attempt 4 · 2026-10-27 06:43 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · f5c1fdf25ca4 · AgentX · TP8/EP1 · 平均延迟
**变更:**f5c1fdf2 上的最终扫描在三次被抢占后运行;c10 在 SGLang 服务启动前于 mi355x-amds host_setup 钩子中失败,原因是节点 mia1-p01-g15 在 15 分钟排空等待后仍保留前一作业 23% 的显存,fail-fast 随后取消了其他点。实测数值及异常说明见上表。
**下一步:**这是节点状态导致的基础设施失败,并非镜像问题(c16 gsm8k 仍以 96.97% 通过);main 也追加了 perf-changelog 条目,因此合并 main 并在新 head 上重新运行最终扫描,不消耗修复次数。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Queued · Run 40462652710 / attempt 1 · 2026-10-27 06:43 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · dabf564932ac · AgentX · TP8/EP1 · Mean latency
Change: Merge main (new perf-changelog entries) keeping this entry at the physical tail, and rerun the full-sweep-fail-fast final sweep over every AgentX point.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 N/A N/A N/A
2 N/A N/A N/A
4 N/A N/A N/A
8 N/A N/A N/A
10 N/A N/A N/A
12 N/A N/A N/A
16 N/A N/A N/A

Note: All rows: pending.
Note: All rows: request errors unavailable.
Note: All rows: Δ N/A: point unavailable.

Next: Wait for the queued final sweep (c1-c16 and the c16 gsm8k eval) to complete, then verify artifacts against the frozen baseline.

中文

最终完整 sweep · 排队中 · Run 40462652710 / attempt 1 · 2026-10-27 06:43 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · dabf564932ac · AgentX · TP8/EP1 · 平均延迟
**变更:**合并 main(新增 perf-changelog 条目)并保持本条目位于文件末尾,然后对全部 AgentX 点重新运行 full-sweep-fail-fast 最终扫描。实测数值及异常说明见上表。
**下一步:**等待已排队的最终扫描(c1-c16 及 c16 gsm8k 评测)完成,然后对照冻结的基线核验产物。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 3 is waiting on 40462652710. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 3 个会话正在等待 40462652710。恢复流程会重跑被抢占的运行,并在运行结束后启动下一个会话。

@github-actions github-actions Bot added klaud-parked Klaud parked until its queued or running benchmark runs end and removed klaud-in-progress Klaud is working on this candidate or will resume it labels Oct 27, 2026
@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

resumed · The session ended before finishing; session 4 continues it in a new run of its planner.

中文

resumed · 会话未完成即已结束;第 4 个会话将在其规划运行的新运行中继续。

@github-actions github-actions Bot added klaud-in-progress Klaud is working on this candidate or will resume it and removed klaud-parked Klaud parked until its queued or running benchmark runs end labels Oct 27, 2026
@Klaud-Cold

Klaud-Cold commented Oct 27, 2026 •

Copy link
Copy Markdown
Collaborator Author

Final full sweep · Passed · Run 40462652710 / attempt 3 · 2026-10-27 17:50 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · dabf564932ac · AgentX · TP8/EP1 · Mean latency
Change: Validate the complete updated-image family at this head.

Concurrency Output tok/s/GPU ↑ TTFT ms ↓ TPOT ms ↓
1 14.68 (+76.9%) 860.52 (-27.9%) 5.07 (-14.8%)
2 18.25 (+69.4%) 523.32 (-40.1%) 5.56 (-13.5%)
4 23.44 (+60.8%) 522.92 (-37.9%) 5.99 (-12.2%)
8 42.78 (+59.7%) 498.02 (-33.3%) 8.27 (-5.3%)
10 52.38 (+57.8%) 620.1 (-37.3%) 9.67 (-5.1%)
12 50.48 (+56.5%) 650.51 (-46.8%) 10.19 (-15.6%)
16 64.24 (+76.1%) 1,165.09 (-49.9%) 16.2 (-29.6%)

Note: c12: 2 request errors.
Note: c16: 1 request error.

Eval Score ↑ Samples
gsm8k/em_strict · c16 97.19% (+0.53 pp) 1,319 each

Next: Mark ready for maintainer review.

中文

最终完整 sweep · 已通过 · Run 40462652710 / attempt 3 · 2026-10-27 17:50 UTC
lmsysorg/sglang:v0.5.22-rocm720-mi35x@sha256:c4b61360b251d62bb17da63f175d571ab22479eb46a144548d79f3ed3998ea11 · dabf564932ac · AgentX · TP8/EP1 · 平均延迟
**变更:**验证此提交更新镜像后的完整配置族。实测数值及异常说明见上表。
**下一步:**标记为就绪,等待维护者审查。

@Klaud-Cold

Copy link
Copy Markdown
Collaborator Author

parked · Session 4 is waiting on 40462652710. Recovery reruns preempted runs and starts the next session once they end.

中文

parked · 第 4 个会话正在等待 40462652710。恢复流程会重跑被抢占的运行,并在运行结束后启动下一个会话。

@adibarra

Copy link
Copy Markdown
Collaborator

not minimal changes only

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) klaud-finished Klaud verified the full sweep; ready for review

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

2 participants