CUDA out of memory

[Ericwen2001]

Thank you for open-sourcing this cool project. I noticed that the memory usage during the training process with your default settings is close to 24GB after the warming, which causes an out-of-memory error on 3090/4090 GPU. Could you empirically suggest any strategies to reduce the memory usage for 24GB GPUs without significantly impacting performance?

[lcc815]

same issue. Have you finally solved this problem?

[lingbo-yu]

@Ericwen2001 same issue. can you provide some advice? thankyou

[Michael-Evans-Savitar]

same issue, it seems that the project doesn't support 3090 GPU with 24G memory. What should be done, any suggestions for running on a single GPU?

[Treeplanter1]

Same issue, I have 2 4090GPU, but I found that only one was used during training. And I got "CUDA out of memory" just after the warm up process.

[lingbo-yu]

https://github.com/nnanhuang/S3Gaussian/blob/6c96925981a7a02328f382691d38a20bfe8c05a2/scene/dataset_readers.py#L609C1-L619C26
refer the the code below,
maybe you guys can change the image size loaded or reduce the cameras used

[Fdarco]

Same issue, are there any methods to train the model with multi-GPUs?

[HHollowAU]

同样的问题，有什么方法可以训练具有多 GPU 的模型吗？

how to use multy GPU,have slove? l use "export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128",can run ,but every slow