Problem in signature of backward

[itemizh-sys]

Hello! I tried to run your code with Torch 2.4.0 + cu121 (the original 3dgs runs on this combination), but I get an error "TypeError:
_RasterizeGaussians.backward() takes 4 positional arguments but 5 were given".

I tried to change the function signature and add a dummy variable, but then another error occurs "RuntimeError:
function _RasterizeGaussiansBackward returned a gradient different than None at position 9, but the corresponding forward input was not a Variable".
Here, I also tried to change the backward output and replaced the 9th position with None, and the code ran, but the model stopped learning (the loss only gets worse).

Do you know how to correctly fix the issue with the signatures?

[zmliao]

This repo does not support training.

You can try it with https://github.com/zmliao/TCGS_Dev/tree/master

[itemizh-sys]

Hello, thank you for your response!
I have a question, as far as I understand, the training issues arise because computations on tensor cores occur at reduced precision fp16.
Have you tried to fix this in any way?
For example, scaling or computing in TF32?
What problems did you encounter and why didn't it work?