Skip to content

fix bug: fsdp2 cannnot run with npu, because the hardcode with cuda #3790

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
wants to merge 2 commits into
base: v1.7.0-release
Choose a base branch
from
Open

fix bug: fsdp2 cannnot run with npu, because the hardcode with cuda #3790

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d5fbf81
fix bug: fsdp2 cannnot run with npu, because the hardcode with cuda
Sep 22, 2025
e41b6d7
fix transformers UT error: FAILED tests/trainer/test_trainer_utils.py…
Sep 25, 2025
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
4 changes: 2 additions & 2 deletions src/accelerate/utils/fsdp_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -498,8 +498,8 @@ def _cast_and_contiguous(tensor, to_contiguous, dtype):

if accelerator.is_main_process:
for (param_name, full_param), sharded_param in zip(full_sd.items(), meta_sharded_sd.values()):
full_param = full_param.detach().cuda()
mesh = sharded_param.device_mesh
full_param = full_param.detach().to(mesh.device_type)
dist.broadcast(full_param, src=0, group=mesh.get_group())
sharded_tensor = distribute_tensor(full_param, mesh, sharded_param.placements)
to_contiguous, casting_dtype = _infer_parameter_dtype(
Expand All @@ -512,8 +512,8 @@ def _cast_and_contiguous(tensor, to_contiguous, dtype):
# We need this else to have a matching `broadcast` for all of the ranks, else we deadlock
else:
for param_name, sharded_param in meta_sharded_sd.items():
full_tensor = torch.empty(sharded_param.size(), device="cuda", dtype=sharded_param.dtype)
mesh = sharded_param.device_mesh
full_tensor = torch.empty(sharded_param.size(), device=mesh.device_type, dtype=sharded_param.dtype)
dist.broadcast(full_tensor, src=0, group=mesh.get_group())
sharded_tensor = distribute_tensor(full_tensor, mesh, sharded_param.placements)
to_contiguous, casting_dtype = _infer_parameter_dtype(
Expand Down
10 changes: 9 additions & 1 deletion src/accelerate/utils/memory.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,7 +153,15 @@ def find_executable_batch_size(

def reduce_batch_size_fn():
nonlocal batch_size
batch_size = batch_size // 2
new_batch_size = int(batch_size * 0.9)
if new_batch_size == 0:
new_batch_size = 1
if new_batch_size == batch_size:
if batch_size > 1:
new_batch_size = batch_size - 1
else:
raise RuntimeError("Batch size reduced to 1 and still out of memory.")
batch_size = new_batch_size
return batch_size

def decorator(*args, **kwargs):
Expand Down