For 台灣人工智慧實驗室、半導體模型研發團隊及分散式訓練架構師

TWCC 台灣智慧雲 GPU 叢集 Xid 79 PCIe 掉卡與 Slurm 任務卡死自動診斷

精確區分 Slurm 節點被排程管理器取消、單卡散熱降頻導致的全叢集拖後腿，以及 PCIe 匯流排掉卡 (Xid 79) 硬體故障。

Replay one resolved incident Try 30 days free

Free 30-day Team evaluation. No card or automatic subscription. Work email verification required.

Live diagnosis, no signup

重現 TWCC 台灣智慧雲 Slurm 訓練任務故障日誌

Upload your log file. Review the selected evidence and masking, then run diagnosis. Keep the original log for any follow-up questions.

Before uploading, review cloud data handling and local options.

training_logs.txt

No log to hand? Try one:

Upload or drop a plain-text log

Upload your log file. Large files are processed in your browser. Only selected evidence is submitted when you run diagnosis.

3 free diagnoses/day

Try Denpex for 30 days. FREE.

No credit card. No automatic subscription. See whether Denpex helps with an incident you already understand.

Start your free 30-day Team evaluation

Up to 50 diagnoses a day for 30 days. Verify your work email to activate an eligible company evaluation, then replay a resolved incident and compare the evidence and next steps with your team's confirmed cause.

What a useful evaluation should establish

在數十個甚至數百個 GPU 的分散式集體逾時中，秒速鎖定唯一觸發崩潰的實體卡與主機節點，加快工單處理與復原流程。

Use Denpex alongside your existing telemetry tools to investigate submitted incident evidence. With monitoring configured on a supported plan, detected failures can trigger alerts and an automatic diagnosis from the captured evidence.

Inside Denpex

See the evidence, not just the claim

Follow the causal chain, inspect the quoted evidence and prepare the next handoff. These actual product captures use founder-account examples and the VS Code extension, with the source identified under each image. Open a capture to inspect it at full size.

Denpex founder-account example: a causal graph connects an Xid 48 ECC event to a rank exit and downstream collective failures, with quoted evidence for each stage. — Live founder-account capture from the bundled NCCL/ECC example. Follow the hardware event, rank exit and downstream symptoms with a supporting log quote at each stage.Inspect original image

Denpex hardware handoff summary asks the operator to capture diagnostic evidence and verify RMA eligibility with the vendor. — Actual hardware handoff summary from the same founder-account example. Capture diagnostic evidence before an approved reset, then confirm RMA eligibility with the vendor.Inspect original image

Denpex historical example says a collective timeout does not establish the initiating rank, cause or failure domain and requests distinguishing evidence before a remedy. — Founder-account example restored from history. The original timeout remains unattributed, and the UI keeps the missing-evidence assessment visible.Inspect original image

Denpex VS Code extension displaying a bundled OOM-before-NCCL example, matched log evidence, verification guidance and an investigation handoff. — Actual VS Code extension capture of the bundled OOM-before-NCCL example. Review matched log evidence, verification guidance and the investigation handoff without leaving the editor.Inspect original image

Failure Signature Triage

TWCC (台灣智慧雲 / Taiwania) - Observed Runtime Evidence

TWCC 台灣智慧雲官方技術文檔

Incident pattern summary

在台灣杉超級電腦上執行多節點 FSDP 訓練任務時，因節點 4 的單一 GPU 掉卡 (Xid 79)，導致其餘所有 GPU 在 all_reduce 屏障集體卡死。Denpex 於 14 毫秒內隔離出真正的故障節點，避免誤怪健康節點。

Terminal traceback excerpt (TWCC (台灣智慧雲 / Taiwania))

NVRM: GPU at PCI 0000:43:00.0: GPU has fallen off the bus.
NVRM: A GPU internal error has occurred. (Xid: 79)
slurmd[node04]: error: *** JOB 948211 ON node04 CANCELLED AT 2026-10-28T04:12:00 DUE TO TIME LIMIT ***
[E ProcessGroupNCCL.cpp:915] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=1284, OpType=ALLREDUCE, Timeout(ms)=600000)