Enable autograd graph to propagate after multi-device syncing (for loss functions in ddp
)
#4774
Job | Run time |
---|---|
13s | |
48m 9s | |
44m 57s | |
46m 19s | |
1h 9m 45s | |
1h 12m 22s | |
1h 12m 0s | |
1h 9m 22s | |
1h 12m 21s | |
1h 14m 11s | |
49m 8s | |
1h 8m 34s | |
1h 8m 37s | |
1h 4m 52s | |
1h 7m 32s | |
1h 21m 33s | |
38m 40s | |
42m 33s | |
35m 52s | |
44m 54s | |
47m 7s | |
1s | |
19h 39m 2s |