OOM Error

[anandravi24]

Hey,

Congratulations on getting selected for CVPR. I was trying to test the work using the custom images demo but seems there is some issue somewhere in the code. The memory requirement balloons and doesn't work even on default Gradio images.

I am using a machine with A100 80 GB

[hanyang-21]

Thank you for bringing this to our attention. In our tests on an A800 80GB GPU, the demo required approximately 65GB of VRAM. If you still encounter memory issues, you can enable the cpu_offload option to reduce VRAM usage. To do this, uncomment lines 256-258 in /noposplat/src/app.py. Please note that while this will significantly decrease VRAM consumption, it may also reduce inference speed.

[zwl995]

Running demo.sh but OOM occurred. My machine is 24G. What is the minimum GB required to run the demonstration? Or can I run it in any other way?

[crmlei]

Thank you for sharing your work！ When I run demo.sh on an A100 80G, I always get an OOM error. Even after enabling the three lines in app.py as suggested, I still encounter this error. And there's an interesting error message: “Tried to allocate 130515.12 GiB. GPU 0 has a total capacity of 79.15 GiB of which 74.21 GiB is free。” What could be the problem?