Skip to content

Fix partition id in the fp32->fp16 param copying step for z2+cpu-offload #2059

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
jeffra merged 7 commits into from
Jul 7, 2022
Merged

Fix partition id in the fp32->fp16 param copying step for z2+cpu-offload #2059

Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
c101885
fix partition id for cpu offload
siddharth9820 Jun 27, 2022
feb7270
formatting changes
siddharth9820 Jun 27, 2022
dec4386
Merge branch 'master' into siddharth/moe-z2+offload-bug-fix
tjruwase Jul 5, 2022
a037f10
corrected bug in torch cpu adam code path
siddharth9820 Jul 6, 2022
6dfbe1b
Merge branch 'siddharth/moe-z2+offload-bug-fix' of github.com:microso…
siddharth9820 Jul 6, 2022
78eef29
Merge branch 'master' into siddharth/moe-z2+offload-bug-fix
siddharth9820 Jul 6, 2022
945f0aa
Merge branch 'master' into siddharth/moe-z2+offload-bug-fix
siddharth9820 Jul 7, 2022
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
16 changes: 11 additions & 5 deletions deepspeed/runtime/zero/stage_1_and_2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1750,14 +1750,20 @@ def step(self, closure=None):
if self.deepspeed_adam_offload:
from deepspeed.ops.adam import DeepSpeedCPUAdam
if type(self.optimizer) == DeepSpeedCPUAdam and self.dtype == torch.half:
bit16_param_groups = [[
bit16_partitions[partition_id]
] for bit16_partitions in self.parallel_partitioned_bit16_groups]
bit16_param_groups = [
[
bit16_partitions[dist.get_rank(
group=self.real_dp_process_group[group_id])]
] for group_id,
bit16_partitions in enumerate(self.parallel_partitioned_bit16_groups)
]
self.optimizer.step(fp16_param_groups=bit16_param_groups)
else:
self.optimizer.step()
# after step(), single_partition_of_fp32_groups has the local optimizer's own partition of updated params
for bit16_partitions, fp32_partition in zip(self.parallel_partitioned_bit16_groups, self.single_partition_of_fp32_groups):
for group_id, (bit16_partitions, fp32_partition) in enumerate(zip(self.parallel_partitioned_bit16_groups, self.single_partition_of_fp32_groups)):
partition_id = dist.get_rank(
group=self.real_dp_process_group[group_id])

bit16_partitions[partition_id].data.copy_(fp32_partition.data)
else:
self.optimizer.step()
Expand Down