Skip to content
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

Fix allreduce for BF16 and ZeRO0 #5170

Merged
merged 5 commits into from
Feb 21, 2024
Merged

Fix allreduce for BF16 and ZeRO0 #5170

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
794e992
fix gradient clipping
tohtana Feb 19, 2024
9871421
Merge branch 'master' into tohtana/fix_fp32_clipping
loadams Feb 20, 2024
84ced5b
perform allreduce on FP32 when BF16 optimizer is enabled
tohtana Feb 21, 2024
fde1df0
ZZMerge branch 'master' into tohtana/fix_bf16_z0_reduce
tohtana Feb 21, 2024
23df2cf
Merge branch 'master' into tohtana/fix_bf16_z0_reduce
tohtana Feb 21, 2024
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
9 changes: 5 additions & 4 deletions deepspeed/runtime/engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1911,9 +1911,6 @@ def print_forward_breakdown(self, fwd_time):

@instrument_w_nvtx
def allreduce_gradients(self, bucket_size=MEMORY_OPT_ALLREDUCE_SIZE):
assert not (self.bfloat16_enabled() and self.pipeline_parallelism), \
f'allreduce_gradients() is not valid when bfloat+pipeline_parallelism is enabled'

# Pass (PP) gas boundary flag to optimizer (required for zero)
self.optimizer.is_gradient_accumulation_boundary = self.is_gradient_accumulation_boundary()
# ZeRO stage >= 2 communicates during non gradient accumulation boundaries as well
Expand All @@ -1926,7 +1923,11 @@ def allreduce_gradients(self, bucket_size=MEMORY_OPT_ALLREDUCE_SIZE):
self.optimizer, 'reduce_gradients'):
self.optimizer.reduce_gradients(pipeline_parallel=self.pipeline_parallelism)
else:
self.buffered_allreduce_fallback(elements_per_buffer=bucket_size)
grads = None
if hasattr(self.optimizer, "get_grads_for_reduction"):
# This is currently for BF16 optimizer
grads = self.optimizer.get_grads_for_reduction()
self.buffered_allreduce_fallback(grads=grads, elements_per_buffer=bucket_size)

@instrument_w_nvtx
def backward(self, loss, allreduce_gradients=True, release_loss=False, retain_graph=False, scale_wrt_gas=True):
Expand Down
Loading