Repository navigation
[Klaud Cold] Update MiniMax-M3.1-Flash FP4 GB200 SGLang AgentX DSpark image to nightly-cu134-20261026-5ef9901 / [Klaud Cold] 将 MiniMax-M3.1-Flash FP4 GB200 SGLang AgentX DSpark 镜像更新至 nightly-cu134-20261026-5ef9901 - #4559
Klaud-Cold wants to merge 4 commits into
Conversation
…34-20261026-5ef9901 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献！
|
|
Initial attempt · Running · Run 40426139872 / attempt 1 · 2026-10-27 00:26 UTC
Note: All rows: pending. Next: Wait for the smoke run (c1, c32 HiCache, c72 TP4/EP4 and the c80 gsm8k eval), then append the perf-changelog entry and dispatch the final sweep if everything passes. 中文初次尝试 · 运行中 · Run 40426139872 / attempt 1 · 2026-10-27 00:26 UTC |
|
parked · Session 1 is waiting on 40426139872. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 1 个会话正在等待 40426139872。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
Bot
added
the
klaud-parked
|
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。 |
github-actions
Bot
added
klaud-in-progress
|
Initial attempt · Running · Run 40426139872 / attempt 7 · 2026-10-27 11:08 UTC
Note: AgentX c1 TP4 EP1 5acf52: 13 request errors.
Next: Run attempts 1-5 were preempted by the scheduler and in attempt 6 an external scancel killed the healthy c32 HiCache job 46 minutes into profiling (7/3,512 request errors), so only c32 is rerun as attempt 7 without using a repair; then append the perf-changelog entry and dispatch the final sweep. 中文初次尝试 · 运行中 · Run 40426139872 / attempt 7 · 2026-10-27 11:08 UTC |
|
parked · Session 2 is waiting on 40426139872. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 2 个会话正在等待 40426139872。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
Bot
added
klaud-parked
|
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
Bot
added
klaud-in-progress
|
Initial attempt · Passed · Run 40426139872 / attempt 9 · 2026-10-27 17:27 UTC
Note: AgentX c1 TP4 EP1 5acf52: 13 request errors.
Next: Run attempt 8 was preempted by the scheduler and recovery's attempt 9 completed c32 successfully, so all smoke benchmarks and the c80 gsm8k eval passed; append the perf-changelog entry and dispatch the final sweep. 中文初次尝试 · 已通过 · Run 40426139872 / attempt 9 · 2026-10-27 17:27 UTC |
…8312e0-65da3ed6a0f4e188
…ark image bump Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 40550986062 / attempt 1 · 2026-10-27 20:21 UTC
Note: AgentX c1 TP4 EP1 5acf52: 14 request errors.
Next: All 12 points and the c80 gsm8k eval passed, but main then appended another perf-changelog entry (#4446), so merge main, keep this entry at the tail and dispatch the final sweep again without using a repair. 中文最终完整 sweep · 已通过 · Run 40550986062 / attempt 1 · 2026-10-27 20:21 UTC |
|
parked · Session 3 is waiting on 40550986062. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 3 个会话正在等待 40550986062。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
Bot
added
klaud-parked
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40577167180 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40577167180 |
|
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
Bot
added
klaud-in-progress
…8312e0-65da3ed6a0f4e188 # Conflicts: # inferencex-e2e/perf-changelog.yaml
Klaud-Cold
removed
the
full-sweep-fail-fast
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 40577167180 / attempt 1 · 2026-10-27 22:03 UTC
Note: AgentX c1 TP4 EP1 5acf52: 13 request errors.
Note: gsm8k/em_strict · c64: Δ N/A: no matched eval baseline. Next: Mark ready for maintainer review. 中文最终完整 sweep · 已通过 · Run 40577167180 / attempt 1 · 2026-10-27 22:03 UTC |
|
parked · Session 4 is waiting on 40577167180. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 4 个会话正在等待 40577167180。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
Bot
added
klaud-parked
|
validated · Repairs: unknown · Runs: 40426139872, 40550859768, 40550986062, 40577090774, 40577167180 中文validated · 修复次数：未知 · 运行：40426139872, 40550859768, 40550986062, 40577090774, 40577167180 |
github-actions
Bot
added
klaud-finished
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| # MiniMax-M3.1-Flash AgentX on GB200 with SGLang: upstream arm64 CUDA 13.4 | ||
| # nightly, which carries the MiniMax-M3.1-Flash support of sgl-project/sglang#45925. |
There was a problem hiding this comment.
🟡 nit (optional): The rewritten comment above this config adds image provenance, which AGENTS.md's comment rules prohibit. Lines 5749-5750 say the image is an "upstream arm64 CUDA 13.4 nightly" carrying the support from sgl-project/sglang#45925. That is a header comment describing the config and where its image came from. Under AGENTS.md "Comments (mandatory, every file)", Git and perf-changelog.yaml are the record for this. Fix: delete those two lines and keep only the DSpark/acceptance-source lines that explain the block.
Why this was flagged
The diff rewrites the comment at inferencex-e2e/configs/nvidia-master.yaml:5749-5750 for minimaxm31flash-fp4-gb200-sglang-agentic-dspark. The new text describes where the image came from (an upstream nightly carrying sgl-project/sglang#45925). AGENTS.md:56-57 prohibits header comments that describe a config, and history comments that say what code replaced or where it came from. It names Git and perf-changelog.yaml as the record. The base comment also carried provenance (cookbook link, verification date), but this PR rewrote those lines and kept the same kind of content instead of removing it. The new perf-changelog.yaml entry already records the image change. The cost is that the comment goes stale at the next nightly bump and adds to config noise. Nothing breaks at runtime.
Verification: The diff replaces the comment at inferencex-e2e/configs/nvidia-master.yaml:5749-5750 with a header and the image's "upstream arm64 CUDA 13.4 nightly" origin. AGENTS.md:54-57 says: "No header comments describing a file, topology, config, recipe, override, or section." Git and perf-changelog.yaml are the record. Nothing breaks at runtime.
Goal: Update SGLang image from
lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509bcb729c653528d448bff9850f8977fcad9d12ba56bd6dbe3d7985e87tolmsysorg/sglang:nightly-cu134-20261026-5ef9901@sha256:64460b432679d4555b6579832454d8b9a21cc8591b2d16b340b94abb59de28e6.Baseline: 2026-10-26 ·
lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509bcb729c653528d448bff9850f8977fcad9d12ba56bd6dbe3d7985e87Mean latency · Sources: API 1, API 2, API 3, API 4
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:dev-minimax-m31-flash@sha256:614b1509bcb729c653528d448bff9850f8977fcad9d12ba56bd6dbe3d7985e87更新为lmsysorg/sglang:nightly-cu134-20261026-5ef9901@sha256:64460b432679d4555b6579832454d8b9a21cc8591b2d16b340b94abb59de28e6。**基线：**2026-10-26 ·
lmsysorg/sglang:dev-minimax-m31-flash@sha256:614b1509bcb729c653528d448bff9850f8977fcad9d12ba56bd6dbe3d7985e87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。