torch.OutOfMemoryError: CUDA out of memory.

[Verdandy]

Dear cryofm team,

I encountered the issue shown in the figure when using the cfm denoise command. I am working with 4×16GB GPUs with 4 multiprocess, but the program still reports a CUDA out-of-memory error.

To troubleshoot this, I have already tried setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to reduce memory fragmentation as recommended by the output log but failed
.
Could you please provide a feasible solution or official low-memory configuration recommendations for running cfm denoise on 16GB GPUs?

Best,
Verdandy

[eugenejyuan]

Hi! Thanks for bringing this up!

The GPU memory usage is mainly controlled by --batch-size. Our default setting is --batch-size 4.
For a 16GB GPU, a safer choice is usually --batch-size 2 (or smaller if needed).

We’ve also updated the troubleshooting page with a reference table of GPU memory usage vs. batch size here:
common-issues: cuda-out-of-memory-oom

Hope this helps!

Best,
cryofm team