Skip to content

Fix tensor parallel with non-floating dtypes #37790

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
merged 1 commit into from
Apr 25, 2025
Merged

Fix tensor parallel with non-floating dtypes #37790

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
17 changes: 6 additions & 11 deletions src/transformers/integrations/tensor_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -307,8 +307,7 @@ def partition_tensor(self, param, empty_param, param_type, param_casting_dtype,
parameter = parameter.contiguous()
if self.use_dtensor:
parameter = DTensor.from_local(parameter, device_mesh, shard, run_check=False)
requires_grad = True if parameter.is_floating_point() else False
return nn.Parameter(parameter, requires_grad=requires_grad)
return nn.Parameter(parameter, requires_grad=parameter.is_floating_point())

@staticmethod
def _prepare_output_fn(output_layouts, use_local_output, mod, outputs, device_mesh):
Expand All @@ -330,8 +329,7 @@ def partition_tensor(self, param, empty_param, param_type, param_casting_dtype,
parameter = parameter.contiguous()
if self.use_dtensor:
parameter = DTensor.from_local(parameter, device_mesh, [Shard(-2)], run_check=False)
requires_grad = True if parameter.is_floating_point() else False
return nn.Parameter(parameter, requires_grad=requires_grad)
return nn.Parameter(parameter, requires_grad=parameter.is_floating_point())


class RowwiseParallel(TensorParallelLayer):
Expand Down Expand Up @@ -383,8 +381,7 @@ def partition_tensor(self, param, empty_param, param_type, param_casting_dtype,
parameter = parameter.contiguous()
if self.use_dtensor:
parameter = DTensor.from_local(parameter, device_mesh, shard, run_check=False)
requires_grad = True if parameter.is_floating_point() else False
return nn.Parameter(parameter, requires_grad=requires_grad)
return nn.Parameter(parameter, requires_grad=parameter.is_floating_point())

@staticmethod
def _prepare_input_fn(input_layouts, desired_input_layouts, mod, inputs, device_mesh):
Expand Down Expand Up @@ -446,8 +443,7 @@ def partition_tensor(self, param, empty_param, param_type, param_casting_dtype,
parameter = parameter.contiguous()
if self.use_dtensor:
parameter = DTensor.from_local(parameter, device_mesh, [Shard(-1)], run_check=False)
requires_grad = True if parameter.is_floating_point() else False
return nn.Parameter(parameter, requires_grad=requires_grad)
return nn.Parameter(parameter, requires_grad=parameter.is_floating_point())


class SequenceParallel(TensorParallelLayer):
Expand Down Expand Up @@ -531,8 +527,7 @@ def partition_tensor(self, param, empty_param, param_type, param_casting_dtype,
parameter = parameter.contiguous()
if self.use_dtensor:
parameter = DTensor.from_local(parameter, device_mesh, [Replicate()], run_check=False)
requires_grad = True if parameter.is_floating_point() else False
return nn.Parameter(parameter, requires_grad=requires_grad)
return nn.Parameter(parameter, requires_grad=parameter.is_floating_point())


SUPPORTED_TP_STYLES = {
Expand Down Expand Up @@ -671,7 +666,7 @@ def shard_and_distribute_module(
# SUPER IMPORTANT we have to use setattr
# otherwise loading is crazy slow
if not isinstance(param, torch.nn.Parameter):
param = torch.nn.Parameter(param)
param = torch.nn.Parameter(param, requires_grad=param.is_floating_point())
setattr(module_to_tp, param_type, param)
# module_to_tp.load_state_dict({param_type: param}, strict=False, assign=True)
return param