ciflow/torchtitan/198536: [DTensor] Migrate unbind sharding propagation to single-dim strategy
- PyTorch 近 90 天出现 1457 次
- PyTorch 近 90 天第 1438 次发版
- 上一次:同一天稍早 · ciflow/torchtitan/193844: [UPDATE] Update
发生了什么
aten.unbind.int used a legacy op strategy that enumerated the input's existing sharding candidates and raised RuntimeError whenever the unbind dim was sharded, instead of redistributing. It also recorded an all-zero redistribute cost matrix, understating redistribution cost. Replace it with a single-dim strategy mirroring split: shard rules for every dim except the unbind dim (output placeholder shifted down by one…
摘要按规则整理自下方来源原文