sam3.1: CUDA out of memory

[Xpangz]

I'm using the RTX 4090, which only has 24GB of memory. When running my own video using examples/sam3.1_video_predictor_example.ipynb, an Out Of Memory (OOM) error occurred in propagate_in_video. I previously tried the tracking of sam3 and did not exceed the memory limit. Why is that? Can this method utilize the video memory of other gpus?

[karlygach481-rgb]

Дисциплина коюңуз өзүңүзгө

[Xpangz]

While the code was running the propagate_in_video function at 25% | 16/63 [00:04<00:13, 3.49it/s], a CUDA out-of-memory error occurred.

[fschvart]

You need to reduce the size of the video. I'm usually able to pull easily videos of 15 1024x1024 frames with a RTX4090.
You can downsample the resolution or just split the video into several shorter clips

[Xpangz]

You need to reduce the size of the video. I'm usually able to pull easily videos of 15 1024x1024 frames with a RTX4090. You can downsample the resolution or just split the video into several shorter clips

Thank you very much. The reason for the OOM issue is indeed that the memory is limited, which has restricted the frame rate.