Skip to content

Conversation

@sahas3
Copy link
Member

@sahas3 sahas3 commented Nov 25, 2025

@sahas3 sahas3 marked this pull request as ready for review November 25, 2025 15:26
@sahas3 sahas3 requested a review from sjarus November 25, 2025 15:26
@sahas3
Copy link
Member Author

sahas3 commented Nov 25, 2025

FYI, while some quantized e2e tests like Conv2dQInt8Module_basic are now passing, the conv operator is actually dispatching to the fp32 version and not i8 version unlike linalg-on-tensors path. This is because the FuseQuantizedOps pass is not run in the tosa pipeline now.

FuseQuantizedOps pass fuses in(i8)->dequant(fp32)->conv(fp32)->quant(i8)->out(i8) into in(i8)->conv(i8)->out(i8). Enabling that pass however results in numeric mis-match for Conv2dQInt8Module_basic since the legalization of aten.conv to tosa doesn't handle the quantized semantics correctly. I will look into that once this PR lands.

Copy link
Collaborator

@sjarus sjarus left a comment

Choose a reason for hiding this comment

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

Thanks once again @sahas3 !

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants