use XLA patched linear in FSDP (fix #3811 and #3718) and expose options on param sharding dim and pinning memory #3830
ronghanghu
force pushed
from
2ddc6a04
to
c905ba8f
4 years ago
ronghanghu
marked this pull request as ready for review 4 years ago
hjm-aws
requested changes
on 2022-08-05
ronghanghu
force pushed
from
2e12e877
to
c18e4865
4 years ago
ronghanghu
force pushed
from
f5dd6e46
to
c15b5453
4 years ago
hjm-aws
approved these changes
on 2022-08-15
hjm-aws
approved these changes
on 2022-08-15
use XLA patched linear in FSDP; expose options for padding in all_gat…
74335297
directly patch `nn.Linear`'s forward method instead of `torch.nn.func…
38de3e24
introduce `shard_param_on_dim_0` to handle potential compiler issues
690947d3
ronghanghu
force pushed
from
c15b5453
to
690947d3
4 years ago
ronghanghu
changed the title use XLA patched linear in FSDP (fix #3811 and #3718) and expose options on padding in all-gather and pinning memory use XLA patched linear in FSDP (fix #3811 and #3718) and expose options on param sharding dim and pinning memory 4 years ago
check `full_param.dim() == 1` instead of catching exceptions in `_con…
34519228
JackCaoG
approved these changes
on 2022-08-17
JackCaoG
merged
3f47abab
into master 4 years ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub