Fixes for training models with bf16 + freshly initialized optimizer via load_module_only
#6545
Job | Run time |
---|---|
1m 44s | |
2m 26s | |
2m 7s | |
2m 0s | |
1m 59s | |
10m 16s |
load_module_only
#6545
Job | Run time |
---|---|
1m 44s | |
2m 26s | |
2m 7s | |
2m 0s | |
1m 59s | |
10m 16s |