Low GPU usage

[canxerian]

GPU usages appears low (~15%) during training using simple_trainer.py.

Repro steps: Install from source, e.g:

pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
pip install git+https://github.com/nerfstudio-project/gsplat.git

# ..and install requirements.txt

Then, on a COLMAP dataset:

python ./gsplat/examples/simple_trainer.py default --disable-viewer --save-ply --max_steps 50000 --data_dir ./splat_project --data_factor 1 --result_dir ./splat_project/splat_result --ply_steps 50000 --save_steps 50000 --eval_steps

GPU usage via nvidia-smi:

System:

RTX 5090 CUDA 12.8 PyTorch 2.7.1

[houchenfeng]

I have encountered the same problem. I use 2DGS along with its corresponding densification method, and train it using the image-fitting.py script. This issue is quite confusing. Are there any operations being performed on the CPU?