Potential Gradient Error when Reloading Frozen Weights in qmodule.py _load_from_state_dict

[cjfghk5697]

There is a potential issue in the _load_from_state_dict method where reloading frozen weights into a frozen module might cause a gradient-related error. The FIXME comment in the code points out this problem.

if type(self.weight.data) is not type(deserialized_weight):
    # Reloading frozen weights into unfrozen module: move to the correct device and force assignment
    self.weight = torch.nn.Parameter(deserialized_weight.to(self.weight.device))
else:
    # FIXME: here we should copy frozen weights into frozen module, but this leads to grad error
    self.weight = torch.nn.Parameter(deserialized_weight.to(self.weight.device))

Proposed Solution:

Explicitly set the requires_grad state of the reloaded parameters to False to prevent errors.
Alternatively, use the .data attribute to directly copy the values without altering the requires_grad state.

self.weight = torch.nn.Parameter(deserialized_weight.to(self.weight.device))
self.weight.requires_grad = False

or

self.weight.data.copy_(deserialized_weight.to(self.weight.device).data)

This fix would prevent potential gradient errors when reloading frozen weights into a frozen module, ensuring the operation is safe and consistent.

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[github-actions]

This issue was closed because it has been stalled for 5 days with no activity.

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[github-actions]

This issue was closed because it has been stalled for 5 days with no activity.