Skip to content

fix: when setting fsdp size unuse megatron for gather in npu #185

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
tastelikefeet merged 2 commits into from
Apr 24, 2026
Merged

fix: when setting fsdp size unuse megatron for gather in npu #185

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
dc7438b
fix: when setting fsdp size unuse megatron for gather in npu
0hujun Apr 23, 2026
badf1a3
fix: when setting fsdp size unuse megatron for gather in npu
0hujun Apr 23, 2026
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
7 changes: 4 additions & 3 deletions src/twinkle/utils/framework.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,16 +43,17 @@ def gather_object(object: Any, device_mesh: DeviceMesh, process_group=None):
output_objects = [object]
group_size = 1
if dist.is_available() and dist.is_initialized():
if Platform.device_prefix() == 'npu':
if Platform.device_prefix() == 'npu' and not device_mesh.has_dim('fsdp'):
# On NPU, letting Python object collectives use the default HCCL
# group previously hung in 8-card metric collection at
# ``dist.all_gather_object(...)``. Reuse Megatron's dedicated Gloo
# DP group instead. When CP is enabled we must pick the DP+CP
# variant, otherwise the rank span for metric aggregation is wrong.
if importlib.util.find_spec('megatron.core') is not None:
from megatron.core import parallel_state as mpu
process_group = mpu.get_data_parallel_group_gloo(
with_context_parallel=getattr(device_mesh, 'cp_world_size', 1) > 1)
if mpu.model_parallel_is_initialized():
process_group = mpu.get_data_parallel_group_gloo(
with_context_parallel=getattr(device_mesh, 'cp_world_size', 1) > 1)
group_size = dist.get_world_size(group=process_group)
if group_size > 1:
output_objects = [None for _ in range(group_size)]
Expand Down
Loading