Reduce memory by using all_gather_into_tensor
#3277
| Job | Run time |
|---|---|
| 3m 36s | |
| 5m 17s | |
| 2m 40s | |
| 1m 12s | |
| 1m 18s | |
| 1m 6s | |
| 1m 8s | |
| 3m 58s | |
| 2m 44s | |
| 7m 29s | |
| 3m 38s | |
| 5m 31s | |
| 2m 9s | |
| 1m 55s | |
| 1m 24s | |
| 1m 48s | |
| 1m 5s | |
| 3m 41s | |
| 3m 28s | |
| 12m 52s | |
| 1h 7m 59s |