Skip to content

Workaround for torch.compile limit when using fused backpass #1523

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
dxqb merged 2 commits into from
Jun 14, 2026
Merged

Workaround for torch.compile limit when using fused backpass #1523

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
76fba6d
fix
Koratahiu Jun 14, 2026
0f9eefc
Mark init_compile() workaround call sites with issue link
dxqb Jun 14, 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
1 change: 1 addition & 0 deletions modules/trainer/GenericTrainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -587,6 +587,7 @@ def __optimizer_step(tensor: Tensor, param_group=param_group, i=i):
tensor.grad = None

def __grad_hook(tensor: Tensor, param_group=param_group, i=i):
init_compile() # workaround for https://github.com/pytorch/pytorch/issues/186537
if self.__is_update_step(self.model.train_progress):
if fused_reduce:
multi.reduce_grads_mean(
Expand Down
4 changes: 1 addition & 3 deletions modules/util/checkpointing_util.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,9 +112,7 @@ def __init__(self, orig_module: nn.Module, orig_forward, train_device: torch.dev
self.layer_index = layer_index

def __checkpointing_forward(self, dummy: torch.Tensor, call_id: int, *args):
# during the backward pass, this runs on an autograd worker thread, which needs its own
# dynamo config initialization (see init_compile)
init_compile()
init_compile() # workaround for https://github.com/pytorch/pytorch/issues/186537
if self.layer_index == 0 and not torch.is_grad_enabled():
self.conductor.start_forward(True)

Expand Down