Skip to content

cp: reduce peak memory for qwen3 moe recipes (1054) into r0.2.0 #1112

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鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ananthsub merged 1 commit into from
Oct 28, 2025
Merged

cp: reduce peak memory for qwen3 moe recipes (1054) into r0.2.0 #1112

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 12 additions & 11 deletions src/megatron/bridge/models/model_provider.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -677,17 +677,18 @@ def _ddp_wrap(
else:
DP = DistributedDataParallel

model = [
DP(
config=get_model_config(model_chunk),
ddp_config=ddp_config,
module=model_chunk,
# Turn off bucketing for model_chunk 2 onwards, since communication for these
# model chunks is overlapped with compute anyway.
disable_bucketing=(model_chunk_idx > 0) or overlap_param_gather_with_optimizer_step,
)
for (model_chunk_idx, model_chunk) in enumerate(model)
]
with torch.cuda.stream(torch.cuda.Stream()):
model = [
DP(
config=get_model_config(model_chunk),
ddp_config=ddp_config,
module=model_chunk,
# Turn off bucketing for model_chunk 2 onwards, since communication for these
# model chunks is overlapped with compute anyway.
disable_bucketing=(model_chunk_idx > 0) or overlap_param_gather_with_optimizer_step,
)
for (model_chunk_idx, model_chunk) in enumerate(model)
]

# Broadcast params from data parallel src rank to other data parallel ranks.
if data_parallel_random_init:
Expand Down
2 changes: 0 additions & 2 deletions src/megatron/bridge/recipes/qwen/qwen3_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,8 +225,6 @@ def _qwen3_moe_common(

if precision_config is None:
precision_config = bf16_mixed()
if isinstance(precision_config, MixedPrecisionConfig):
precision_config.grad_reduce_in_fp32 = False

# MoE-specific pipeline split configurations
if account_for_embedding_in_pipeline_split:
Expand Down