Skip to content

【FlexCheckpoint】fix early clearing of communication buffer in resharder #78575

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
swgu98 merged 1 commit into from
Apr 7, 2026
+18 −3
Merged

【FlexCheckpoint】fix early clearing of communication buffer in resharder #78575

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
2 changes: 1 addition & 1 deletion python/paddle/distributed/flex_checkpoint/dcp/load_state_dict.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,7 +69,7 @@

# When using the communication mode described below, newly created tensors will not be allocated GPU memory.
# The allocation of GPU memory for these tensors will occur only when meaningful values are written to them.
_UNINIT_TENSOR_MODES = ["send_recv", "grouped_send_recv", "parallel_broadcast"]
_UNINIT_TENSOR_MODES = ["send_recv", "grouped_send_recv"]

_metadata_manager = MetadataManager()

Expand Down
19 changes: 17 additions & 2 deletions python/paddle/distributed/flex_checkpoint/dcp/resharder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -402,6 +402,11 @@ def assign_sharded_weight(src, dst):
starts, ends = [], []
dst_starts, dst_ends = [], []

dest_tensor = dst.local_tensor
if not dest_tensor._is_initialized():
buffer = paddle.zeros_like(dest_tensor)
buffer._share_buffer_to(dest_tensor)

for i in range(ndim):
src_begin = src.global_offset[i]
src_end = src_begin + src.local_shape[i]
Expand Down Expand Up @@ -929,10 +934,12 @@ def broadcast_cross_p_group_and_assign(self, tensor_list, task_batches):
read_item.tensor_name
]:
if not target_sharded_weight.local_tensor._is_initialized():
buffer = paddle.zeros_like(
buffer_t = paddle.zeros_like(
target_sharded_weight.local_tensor
)
buffer_t._share_buffer_to(
target_sharded_weight.local_tensor
)
buffer._share_buffer_to(target_sharded_weight.local_tensor)

src_tensor = received_sharded_weight.local_tensor
tgt_place = target_sharded_weight.local_tensor.place
Expand Down Expand Up @@ -985,6 +992,14 @@ def broadcast_cross_global_group_and_assign(self):
for target_sharded_weight in self.grouped_target_state_dict[
read_item.tensor_name
]:
if not target_sharded_weight.local_tensor._is_initialized():
buffer_t = paddle.zeros_like(
target_sharded_weight.local_tensor
)
buffer_t._share_buffer_to(
target_sharded_weight.local_tensor
)

assign_sharded_weight(
src=received_sharded_weight,
dst=target_sharded_weight,
Expand Down
Loading