Skip to content

[Bugfix] Fix k_proj's bias for whisper self attention #12342

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 5 commits into from
Jan 23, 2025
Merged

[Bugfix] Fix k_proj's bias for whisper self attention #12342

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
d49dd8e
clean up
Isotr0py Jan 22, 2025
f77cf40
revert whisper
Isotr0py Jan 23, 2025
b4f1c89
fix whisper self attn linear
Isotr0py Jan 23, 2025
7ca5443
another solution
Isotr0py Jan 23, 2025
bee02df
clean up
Isotr0py Jan 23, 2025
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
21 changes: 18 additions & 3 deletions vllm/model_executor/models/whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -729,7 +729,22 @@ def sample(
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
loader = AutoWeightsLoader(self, skip_prefixes=["proj_out."])
loaded_weights = [(name, loaded_weight)
for name, loaded_weight in weights]
mapper = WeightsMapper({".fc1.": ".mlp.fc1.", ".fc2.": ".mlp.fc2."})
return loader.load_weights(loaded_weights, mapper=mapper)
# add fake zeros bias for k_proj to state_dict
weights = _create_fake_bias_for_k_proj(weights)
return loader.load_weights(weights, mapper=mapper)


def _create_fake_bias_for_k_proj(
weights: Iterable[Tuple[str, torch.Tensor]]
) -> Iterable[Tuple[str, torch.Tensor]]:
"""
Create full zeros bias for k_proj weight in self-attention layers.
So that the bias for k_proj in qkv_proj can be initialized with zeros.
"""
for name, weight in weights:
if ".self_attn.k_proj.weight" in name:
bias = torch.zeros(weight.size(0))
bias_name = name.replace("weight", "bias")
yield from [(name, weight), (bias_name, bias)]
yield name, weight
Comment on lines +738 to +750
Copy link
Member Author

@Isotr0py Isotr0py Jan 23, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

After reconsideration, I decided to add full zeros bias for k_proj to loaded_weights, so that we don't need to modify the linear method implementation and the bias of qkv_proj can also be initialized for k_proj.

Loading