Tobias van der Werff

I agree with @jveitchmichaelis that being able to disable box filtering or being able to filter masks along with the bounding boxes would be a nice feature to have.

I think I found a decent solution. The `interpolate_patch_14to16.py` script can be modified in the following way: - The p14 checkpoints contain the weights under the `module` key, not `model`....

I think this is still an issue? I'm running version 1.9.4-3 but still getting this error message.

I'd love to work on this @HDCharles

One question about this @HDCharles. The SpinQuant repo has a dependency on the [CUDA fast Hadamard transform](https://github.com/Dao-AILab/fast-hadamard-transform) package for doing the actual Hadamard transform. Would it be acceptable to include...

@jerryzh168 I'd be happy to work together on this. For now, I'll be implementing the rotation matrices one by one, and document the results in the PR. Let me know...

Good to know @HDCharles, I'll keep the tensor subclasses in mind. I was wondering, will the choice to integrate this into torchao depend on the performance delta it produces? Currently,...

Update: I'm currently somewhat stuck on this PR. The R2 and R4 matrices are both implemented, and show small perplexity improvements for in4wo-64 quantization (not much though, see table above)....

I have unblocked myself somewhat regarding the R1 rotation matrices: the authors provide [downloads](https://drive.google.com/drive/folders/1nV9juzE6_OHr10y6Ke5KCyOiGqDr0srX) for the optimized R1/R2 weights. I could try these out to see what kind of performance...

I'll do a final reformat and add some more results in the next few days @HDCharles