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.

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

[BUG] Padding of infeatures/outfeatures and packing #100

Closed
Qubitium opened this issue Jun 28, 2024 · 0 comments · Fixed by #98
Closed

[BUG] Padding of infeatures/outfeatures and packing #100

Qubitium opened this issue Jun 28, 2024 · 0 comments · Fixed by #98
Assignees
Labels
bug Something isn't working

Comments

@Qubitium
Copy link
Contributor

Qubitium commented Jun 28, 2024

Exllama v1/v2 + Marlin have (or will have) auto padding of infeatures and outfeatures so many models can run despite not exactly matching kernel restrictions.

However, if the padding is done dynamically at runtime, and we pack the padded tensors, the result weights may not be compatible with vllm/sglang as the shape of the tensors are off due to padding.

This is a theory at the moment and needs investigation. If confirmed, we need to trim the tensors during or post packing to original infeatures/outfeatures shape/size/

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug Something isn't working
Projects
None yet
Development

Successfully merging a pull request may close this issue.

2 participants