[litgpt benchmark] enable force_recompute_fp8_weight_in_bwd
when torchao.float8
is used with FSDP2
#5331
Job | Run time |
---|---|
10m 34s | |
30m 15s | |
10m 58s | |
30m 12s | |
9m 58s | |
31m 29s | |
7m 19s | |
25m 19s | |
25m 1s | |
24m 50s | |
24m 35s | |
28m 41s | |
1s | |
4h 19m 12s |