Skip to content

Samyamr/largest partitioned params calculation fix #1150

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tjruwase merged 7 commits into from
Jun 16, 2021
+5 −2
Merged

Samyamr/largest partitioned params calculation fix #1150

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
b16c781
largest_partitioned_params calculation fix
samyam Jun 8, 2021
c1ec6e2
Update stage3.py
samyam Jun 8, 2021
b52b333
Update stage3.py
samyam Jun 8, 2021
b7e9e9b
formatting fix
samyam Jun 9, 2021
307ccdf
changing sub-group size default to 1e9
samyam Jun 9, 2021
75bdbbc
Merge branch 'master' into samyamr/largest-partitioned-params-calcula…
tjruwase Jun 11, 2021
faaa87d
Merge branch 'master' into samyamr/largest-partitioned-params-calcula…
tjruwase Jun 15, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion deepspeed/runtime/zero/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,7 +91,7 @@
ZERO_OPTIMIZATION_OFFLOAD_OPTIMIZER_DEFAULT = None

ZERO_OPTIMIZATION_SUB_GROUP_SIZE = 'sub_group_size'
ZERO_OPTIMIZATION_SUB_GROUP_SIZE_DEFAULT = 1000000000000
ZERO_OPTIMIZATION_SUB_GROUP_SIZE_DEFAULT = 1000000000

#maximum number of parameters per GPU before releasing them
ZERO_OPTIMIZATION_MAX_LIVE_PARAMETERS = 'stage3_max_live_parameters'
Expand Down
5 changes: 4 additions & 1 deletion deepspeed/runtime/zero/stage3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -852,7 +852,10 @@ def __init__(self,
count = count + 1

#Largest partitioned param
largest_partitioned_param_numel = max(self.fp16_partitioned_groups_flat_numel)
largest_partitioned_param_numel = max([
max([tensor.numel() for tensor in fp16_partitioned_group])
for fp16_partitioned_group in self.fp16_partitioned_groups
])
print_rank_0(
f'Largest partitioned param numel = {largest_partitioned_param_numel}',
force=True)
Expand Down