Skip to content

Fix: fix hang in FSDP UBR with symmetric registration #3565

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

Open
youngeunkwon0405 wants to merge 1 commit into
base: core_r0.16.0
Choose a base branch
from
+19 −15
Open

Fix: fix hang in FSDP UBR with symmetric registration #3565

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
34 changes: 19 additions & 15 deletions megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1613,24 +1613,28 @@ def __init__(
)
# Select the communicator groups to register FSDP buffers.
self.ubr_groups = [self.dist_index.get_fsdp_group(is_expert_parallel=False)]
if self.dist_index.get_fsdp_group(is_expert_parallel=True) is not None:
# Expert-DP group when using EP
self.ubr_groups.append(self.dist_index.get_fsdp_group(is_expert_parallel=True))
if self.dist_index.get_outer_fsdp_group() is not None:
# Outer/Inter-FSDP group when using hybrid FSDP
self.ubr_groups.append(self.dist_index.get_outer_fsdp_group())
if (
self.dist_index.get_fsdp_group(
is_expert_parallel=False, independent_all_gather=True
)
is not None
):
# All-gather group used when overlapping all-gather and gradient reduction.
self.ubr_groups.append(
# Currernlty we are not supporting symmetric registration for other groups.
# For now, we only allow to register buffer to other groups when symmetric
# registration is disabled. We will support it in the future.
if self.ddp_config.disable_symmetric_registration:
if self.dist_index.get_fsdp_group(is_expert_parallel=True) is not None:
# Expert-DP group when using EP
self.ubr_groups.append(self.dist_index.get_fsdp_group(is_expert_parallel=True))
if self.dist_index.get_outer_fsdp_group() is not None:
# Outer/Inter-FSDP group when using hybrid FSDP
self.ubr_groups.append(self.dist_index.get_outer_fsdp_group())
if (
self.dist_index.get_fsdp_group(
is_expert_parallel=False, independent_all_gather=True
)
)
is not None
):
# All-gather group used when overlapping all-gather and gradient reduction.
self.ubr_groups.append(
self.dist_index.get_fsdp_group(
is_expert_parallel=False, independent_all_gather=True
)
)

log_single_rank(
logger,
Expand Down
Loading