trunk/05b558d4e79ea7a049dbc351a9067f0bf2a63faf: [c10d] Bind local rank before eager NCCL2 initialization (#193237)
- PyTorch 近 90 天出现 38 次
- PyTorch 近 90 天第 38 次发版
- 上一次：同一天稍早 · trunk/26b9ddd7f8a46a15067a7cdc789623f3cdee2cb1
发生了什么
NCCL2 initializes eagerly, so selecting torch.cuda.current_device() before honoring LOCAL_RANK can bind forked workers to the same GPU. Prefer LOCAL_RANK and make DistributedTestBase bind the selected accelerator before init_process_group; the helper also accepts an explicit backend for compatibility tests. Test Plan: CUDA_VISIBLE_DEVICES=2,3 VIRTUAL_ENV=/home/tristanr/pytorch/.venv PATH=/home/tristanr/pytorch/.venv…
摘要按规则整理自下方来源原文