PRoPE dimension bug

[rmbrualla]

Found with AI -- seems like a real bug.

In PRopeFlashAttention.forward, the image_width and image_height parameters passed to _prepare_apply_fns are swapped. The function receives H (height) and W (width) from the standard PyTorch image format, but passes image_width=H and image_height=W. This reverses the dimensions, causing incorrect camera intrinsic normalization when intrinsics K are provided. The intrinsic matrix elements fx, cx would be divided by height instead of width, and fy, cy by width instead of height.

[yyfz]

Hi, rmbrualla!

Thanks a lot for pointing this out. I checked, and it is indeed a bug. When I initially experimented with camera injection, I fed the intrinsics into the prope and found the results were poor—it turns out this was the reason why.

Although the final Pi3X doesn't input intrinsics at the prope stage, fixing this bug and retraining should definitely improve the model's ability to follow intrinsics.