ZeRO3, improved parameter all-gather operation #1188
remove norm(), avoid memcpy after allgather
1e73e758
zarzen
marked this pull request as draft 4 years ago
zarzen
force pushed
to
1e73e758
4 years ago
WIP: wrapped ncclAllgather as customized op in DS
67b3db3e
WIP: integrated into partition_parameters
70e681f0
Fix format
81b4fc4a
zarzen
marked this pull request as ready for review 4 years ago
Merge branch 'master' into impr_allgather_params
8a14e434
cleaned dead code, modified unit test
32c8fa72
Merge branch 'master' into impr_allgather_params
c4728f50
Merge branch 'master' into impr_allgather_params
e075fd4d
removed customized c++ extension
52085080
Merge remote-tracking branch 'origin/master' into impr_allgather_params
ffd3d3b0
change torch.ones to torch empty
1ed96ce2
Merge branch 'master' into impr_allgather_params
220f2e0e
Merge branch 'master' into impr_allgather_params
8f655940
typo
0e6d8e0c
Merge branch 'master' into impr_allgather_params
691749fe
Merge branch 'master' into impr_allgather_params
88e750e6
Merge branch 'master' into impr_allgather_params
497ee7d7
Merge branch 'master' into impr_allgather_params
bd8839c5
Merge branch 'master' into impr_allgather_params
25829106
Merge branch 'master' into impr_allgather_params
4ca0d391
Merge branch 'master' into impr_allgather_params
56de9ad9
Merge branch 'master' into impr_allgather_params
056cf101
Merge branch 'master' into impr_allgather_params
aac09cd2
jeffra
commented
on 2021-10-07
Merge branch 'master' into impr_allgather_params
6201b295
warn if not cuda tensor for allgather
50a9215d
Merge branch 'master' into impr_allgather_params
c554a589
Merge branch 'master' into impr_allgather_params
b7e131d4
fix formatting
813cb227
Merge branch 'master' into impr_allgather_params
588d3d0e
Merge branch 'master' into impr_allgather_params
eb0a540f
fix: move ds_tensor to cuda device
c092b789
Merge branch 'master' into impr_allgather_params
e73809dc
Merge branch 'master' into impr_allgather_params
d1d3c28f
Merge branch 'master' into impr_allgather_params
62cb1049
Merge branch 'master' into impr_allgather_params
ab64b17a
remove try clause on the path for fetching params
7a801721
Merge branch 'microsoft:master' into impr_allgather_params
f01dad85
tjruwase
approved these changes
on 2021-10-30
Merge branch 'master' into impr_allgather_params
524e6096
Merge branch 'master' into impr_allgather_params
d7fff585
tjruwase
enabled auto-merge (squash) 4 years ago
tjruwase
merged
c0eeb69d
into master 4 years ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub