Skip to content

[Bugfix] Fix torch dynamo fixes caused by replace_parameters #8748

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
Sep 24, 2024
Merged

[Bugfix] Fix torch dynamo fixes caused by replace_parameters #8748

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
12 changes: 8 additions & 4 deletions vllm/model_executor/layers/quantization/utils/layer_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,13 +21,17 @@ def replace_parameter(mod: torch.nn.Module, name: str,
new: Union[torch.Tensor, torch.nn.Parameter]) -> None:

old = getattr(mod, name)
if old.dtype == new.dtype and \
if type(old) is type(new) and old.dtype == new.dtype and \
old.untyped_storage().nbytes() == new.untyped_storage().nbytes():
# If we can just update in-place to avoid re-registering
# can be faster if the underlying storage is the same
update_tensor_inplace(old, new)
else:
# Fallback re-register parameter
# Fallback re-register parameter, convert to Parameter if necessary
# this not only ensures we don't register a tensor as a parameter, but
# also ensures that all parameter subclasses get re-registered as
# parameters for `torch.compile` compatibility
if not isinstance(new, torch.nn.Parameter):
new = torch.nn.Parameter(new)
mod.register_parameter(name, torch.nn.Parameter(new))
new = torch.nn.Parameter(new, requires_grad=False)
mod.register_parameter(name,
torch.nn.Parameter(new, requires_grad=False))
Loading