Skip to content

Fix loss scaling and backward call of ZenFlow #7793

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

Closed
tohtana wants to merge 2 commits into from
Closed

Fix loss scaling and backward call of ZenFlow #7793

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
6f21abd
Fix FP16 loss scaling in backward() for non-ZenFlow paths
tohtana Jan 18, 2026
f76a770
Add backward() method to Stage 3 optimizer for ZenFlow support
tohtana Jan 18, 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
4 changes: 2 additions & 2 deletions deepspeed/runtime/engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2505,9 +2505,9 @@ def backward(self, loss, retain_graph=False, scale_wrt_gas=True):

# TODO: handle these scaling with direct calls to loss.backward()
if isinstance(self.optimizer, ZeROOptimizer):
gas_scaled_loss = self.optimizer.scale_if_loss(gas_scaled_loss)
loss = self.optimizer.scale_if_loss(loss)
elif self.torch_autocast_z0_gradscaler:
gas_scaled_loss = self.torch_autocast_z0_gradscaler.scale(gas_scaled_loss)
loss = self.torch_autocast_z0_gradscaler.scale(loss)

with compiled_autograd(self._is_compiled_autograd_enabled, self._compile_kwargs):
# ZenFlow requires exclusive control over the backward pass to manage its
Expand Down
19 changes: 19 additions & 0 deletions deepspeed/runtime/zero/stage3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2376,6 +2376,25 @@ def backward_epilogue(self):
if self.swap_optimizer:
self.optimizer_swapper.post_backward()

def backward(self, loss, retain_graph=False):
Copy link
Collaborator

@Antlera Antlera Jan 19, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It looks like this interface is now designed for ZenFlow-like methods, which makes the integration easier and cleaner.
I'm not sure if DeepSpeed × PyTorch prefers keeping this fully PyTorch-aligned instead of adding framework-specific logic. Shall we add a TODO here?

Copy link
Collaborator

@Antlera Antlera Jan 19, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Otherwise, this code LGTM. Thanks!

Copy link
Collaborator Author

@tohtana tohtana Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's a good point. I wanted to quickly resolve this so we can pass the full CI test suite, but it makes the core part of the optimizer more Zenflow-specific. We probably shouldn’t cut corners for this. Let me close this for now, and then consider a more general approach.

"""
Backward pass for Stage 3 optimizer.

When ZenFlow is enabled, this method handles the specialized backward
pass with loss scaling and ZenFlow's prologue/epilogue hooks.
"""
self.backward_prologue()

self.enter_backward()
if self.custom_loss_scaler:
scaled_loss = self.external_loss_scale * loss
scaled_loss.backward(retain_graph=retain_graph)
else:
self.loss_scaler.backward(loss.float(), retain_graph=retain_graph)

self.backward_epilogue()
self.exit_backward()

def get_fp32_grad_partitions(self) -> Dict[int, Dict[int, Tensor]]:
"""get fp32 gradient partition dictionary
accessed as grad_dict[parameter_group_index][parameter_index]
Expand Down