Skip to content

fix train deepseek V4 with fsdp2: AttributeError: 'Tensor' object has no attribute 'device_mesh' #4023

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
frozenleaves wants to merge 1 commit into
base: main
Choose a base branch
from
+19 −0
Open

fix train deepseek V4 with fsdp2: AttributeError: 'Tensor' object has no attribute 'device_mesh' #4023

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
19 changes: 19 additions & 0 deletions src/accelerate/utils/fsdp_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -518,6 +518,15 @@ def _cast_and_contiguous(tensor, to_contiguous, dtype):
f"Full state dict has {len(full_sd)} keys, sharded has {len(meta_sharded_sd)} keys."
)
full_param = full_sd[param_name]
# Persistent buffers are not sharded by fully_shard and remain plain Tensors.
# Broadcast them and keep as-is instead of trying to access DTensor attributes.
if not isinstance(sharded_param, DTensor):
full_param = full_param.detach().to(accelerator.device)
dist.broadcast(full_param, src=0, group=dist.group.WORLD)
if cpu_offload:
full_param = full_param.to("cpu")
sharded_sd[param_name] = full_param
continue
device_mesh = sharded_param.device_mesh
full_param = full_param.detach().to(device_mesh.device_type)
if isinstance(full_param, DTensor):
Expand All @@ -540,6 +549,15 @@ 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():
# Persistent buffers are not sharded by fully_shard and remain plain Tensors.
# Broadcast them and keep as-is instead of trying to access DTensor attributes.
if not isinstance(sharded_param, DTensor):
full_tensor = torch.empty(sharded_param.size(), device=accelerator.device, dtype=sharded_param.dtype)
dist.broadcast(full_tensor, src=0, group=dist.group.WORLD)
if cpu_offload:
full_tensor = full_tensor.to("cpu")
sharded_sd[param_name] = full_tensor
continue
device_mesh = sharded_param.device_mesh
full_tensor = torch.empty(sharded_param.size(), device=device_mesh.device_type, dtype=sharded_param.dtype)
dist.broadcast(full_tensor, src=0, group=dist.group.WORLD)
Expand Down Expand Up @@ -916,3 +934,4 @@ def get_parameters_from_modules(
for module in modules:
parameters.extend(list(module.parameters()))
return set(parameters)