BladeDISC
BladeDISC copied to clipboard
support collective operators
To optimize distributed training graph (DP, FSDP), DISC needs to support collective ops as a preliminary preparation
- [ ] support collective ops compilation and execution (all_reduce, all_gather, reduce_scatter) @Yancey1989
- #1275
- #1288
- [ ] support async collective ops execution @eedalong
- #1287