deepseek-v4.1POST
双 DGX Spark 跑 V4.1 Flash 的 OOM 与 rank 挂死排查
开发者定位两个问题：并发长提示会让 rank 0 单独分词并 OOM（每 100 万 token 请求约占 1.4GB），且一个 rank 挂掉后另一个会永久卡住；其分支可自动检测并重启双 rank，约 1 秒发现、1 分钟内恢复。
wojo
@rwojo10 天前收录原推发布于 2026.10.22
wojo @rwojo原推发布于 2026.10.22
Mine was doing similar things until I tracked down two causes: several long prompts at once can OOM rank 0 (rank 0 alone tokenizes, and held ~1.4 GB per 1M-token request until the reply finished; fix in MiaAI-Lab/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks#40), and when one rank dies the other just hangs forever. My branch (#30, or wojo/DeepSeek-v4.1-Flash-1017-2x-DGX-Sparks main) detects a broken pair and restarts both ranks automatically, about 1 s to notice a dead rank and under a minute to recover. What do your logs show when it crashes? A memguard kill, a CUDA/NCCL error, or does it just hang?
x.com/rwojo/status/2113104928155525605采集于 2026.10.22 11:59