Keep FFN output layer in float32 for T5 models #2310
Job | Run time |
---|---|
41m 24s | |
5m 55s | |
39m 9s | |
34m 56s | |
1h 3m 31s | |
1h 49m 12s | |
9m 47s | |
6m 50s | |
6m 12s | |
10m 43s | |
9m 14s | |
10s | |
43s | |
13m 7s | |
7m 37s | |
3m 35s | |
0s | |
6h 2m 5s |
Job | Run time |
---|---|
41m 24s | |
5m 55s | |
39m 9s | |
34m 56s | |
1h 3m 31s | |
1h 49m 12s | |
9m 47s | |
6m 50s | |
6m 12s | |
10m 43s | |
9m 14s | |
10s | |
43s | |
13m 7s | |
7m 37s | |
3m 35s | |
0s | |
6h 2m 5s |