Skip to content
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

Jump to bottom

[Bugfix] Fix dummy weight for fp8 #4916

Merged
merged 4 commits into from
May 20, 2024
Merged

[Bugfix] Fix dummy weight for fp8 #4916

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
993f39c
Fix dummy weight for fp8
May 19, 2024
ed12d1f
Add comment, uniform is not supported in FP8
May 20, 2024
90ad378
remove torch from comment
May 20, 2024
0a3fff8
.to creates a copy and doesn't assign it param.data, changed it so the
May 20, 2024
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
9 changes: 8 additions & 1 deletion vllm/model_executor/model_loader/weight_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -369,4 +369,11 @@ def initialize_dummy_weights(
"""
for param in model.state_dict().values():
if torch.is_floating_point(param):
param.data.uniform_(low, high)
if torch.finfo(param.data.dtype).bits < 16:
Copy link
Collaborator

Choose a reason for hiding this comment

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

Can you add a comment saying we do this because uniform doesn't support <8-bit data type?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Sure, Added a comment :) Thanks

# uniform_ doesn't support < 16-bit datatypes (FP8)
dtype = param.data.dtype
tmp_param = param.data.to(torch.float16)
tmp_param = tmp_param.uniform_(low, high).to(dtype)
param.data.copy_(tmp_param)
else:
param.uniform_(low, high)
Loading