CUDA OOM in model v1.1; Not able to run v1.0 about openlrm HOT 2 CLOSED

Hi,

Thank you for your interest!

I have lowered the frame_size here to 2 by default, which controls the number of video frames rendered per batch. If you find that the VRAM usage is still high, you can further reduce it.

Moreover, could you please share your environment settings? Here are some things to check:

• Make sure that xFormers memory-efficient attention is enabled and compatible with your CUDA version. You should see UserWarning: xFormers is available if it is working correctly.
• The recommended version of PyTorch is >=2.1.2. It is crucial to have PyTorch >=2.0 to enable native torch FlashAttention.

As a reference, the VRAM usage for base-sized models should be below 20GB for inference on an A100 with frame_size=2.

For v1.0 models, if you are using d4caebb commit, could you plz share with us more information? From this JSONDecodeError, we couldn't locate the exact line to throw this error. Thanks!

from openlrm.

Hi, thanks for the reply. I think it's because i changed the code to take all images in a folder as input which cause the OOM issue. I will further test and raise questions if i meet any trouble. Thanks!

from openlrm.