It seems that CUDA is not enabled

[cmh1027]

The program is strangely slow and it seems that the computation is done on CPU, not cuda (As you can see in the image, CPU utilization is 2108%) It's quite weird because I checked that parameters in GaussianModel are loaded on cuda. Modified diff-gaussian-rasterization package is highly suspicious.

[jamesjjk]

100% modified diff-gaussian-rasterization is missing CUDA optimisations / overwritten this is not running on CUDA other then loading.

[cmh1027]

@jamesjjk Maybe I have to compare the code with the original one before authors reply the issue

[Alexander0Yang]

Thanks for your interest. It's impractical to run CUDA kernel on CPU. I suspect the I/O bottleneck is the culprit for your slowly training. If you have sufficient memory, you can try to disable the dataloader.

[Zerg-Overmind]

same here... training on dynerf dataset is extremely slow (~1.12it/s) also, loss backward done per instance within a batch was also confusing and slow...

[feixue94]

same problem. 1.3it/s on RTX 3090

[qrcat]

It seem to KNN's problem. If you disable rigid loss, it will be faster.