Skip to content

fix: experttensorparallel dynamo tracing crash #4973

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
devtyagi3909 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix: experttensorparallel dynamo tracing crash #4973

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4442abf
fix ExpertTensorParallel Dynamo crash
devtyagi3909 Oct 5, 2026
6686c7b
fix: restore _apply type hints and keyword signature compatibility
devtyagi3909 Oct 5, 2026
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
18 changes: 6 additions & 12 deletions test/prototype/moe_training/reference_parallel_styles.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,12 +165,12 @@ def _token_dispatch(self, mod, inputs, device_mesh):
routed_input, num_tokens_per_expert = inputs

routed_input = DTensor.from_local(
routed_input, device_mesh["tp"], (Replicate(),)
routed_input, self.tp_mesh, (Replicate(),)
).to_local(grad_placements=(Partial(),))

inputs = (routed_input, num_tokens_per_expert)

return super()._token_dispatch(mod, inputs, device_mesh["ep"])
return super()._token_dispatch(mod, inputs, self.ep_mesh)

def _partition_fn(self, name: str, mod: nn.Module, device_mesh: DeviceMesh) -> None:
from torch.distributed.tensor import distribute_tensor
Expand All @@ -189,15 +189,9 @@ def _partition_fn(self, name: str, mod: nn.Module, device_mesh: DeviceMesh) -> N
)

def _token_combine(self, mod, routed_output, device_mesh):
return super()._token_combine(mod, routed_output, device_mesh["ep"])
return super()._token_combine(mod, routed_output, self.ep_mesh)

def _apply(self, module: nn.Module, device_mesh: DeviceMesh) -> nn.Module:
from torch.distributed.tensor import distribute_module

return distribute_module(
module,
device_mesh,
partition_fn=self._partition_fn,
input_fn=self._token_dispatch,
output_fn=self._token_combine,
)
self.tp_mesh = device_mesh["tp"]
self.ep_mesh = device_mesh["ep"]
return super()._apply(module, device_mesh)