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

Converting fp32 models to fp16 without cuda. #1783

Closed
pashu123 opened this issue Jan 9, 2023 · 2 comments
Closed

Converting fp32 models to fp16 without cuda. #1783

pashu123 opened this issue Jan 9, 2023 · 2 comments
Assignees
@pashu123
Labels
enhancement New feature or request

Comments

@pashu123
Copy link
Member

pashu123 commented Jan 9, 2023

Currently, for fp16 models we need torch's Cuda version since some of the fp16 ops are not supported on CPU, hence it's not able to trace.
I'd like to write a pass that converts fp32 weights to fp16 and forwards the types. Would you happen to have any suggestions? @silvasean @ramiro050
@pashu123 pashu123 added the enhancement New feature or request label Jan 9, 2023
@pashu123 pashu123 self-assigned this Jan 9, 2023
@silvasean
Copy link
Contributor

We don't want such a pass upstream. As discussed in the developer hour today, this is probably a feature request for PyTorch to support program capture for fp16 operators without CUDA.

Related: #1615

@pashu123
Copy link
Member Author

We don't want such a pass upstream. As discussed in the developer hour today, this is probably a feature request for PyTorch to support program capture for fp16 operators without CUDA.

Related: #1615

Thanks for the feedback.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@pashu123 pashu123

Labels
enhancement New feature or request
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@pashu123 @silvasean