RuntimeError during Sampling and thus error during training about jukebox HOT 4 OPEN

The second error comes from having too little data, you have 15 files and the function calculate_bandwidth by default iterates over about 600 sec of audio. Try passing a smaller duration to it.

from jukebox.

For the first error, how much CUDA memory do you have? The 5b_lyrics model needs 16GB for the default hps. Try lowering max_batch_size from 3 to say 1 in sample.py?

from jukebox.

Thanks for the answers. I have a GeForce 740M, a low-end GPU, with only 2GB of memory, therefore I guess that's what CUDA has, which is far from ideal.
That's also what the sampling command outputs when it fails to allocate more memory.
RuntimeError: CUDA out of memory. Tried to allocate 44.00 MiB (GPU 0; 1.96 GiB total capacity; 1.83 GiB already allocated; 2.75 MiB free; 1.87 GiB reserved in total by PyTorch)

I'm not too sure what the sampling part is compared to the training part. I think i'll have to read the paper more in depth.
However, I have now more than 800s of audio to work with, but I still get a similar error (now with other values in the AssertionError message) :
AssertionError: Midpoint 21112952 of item beyond total length 20853793
Is there anything I did wrong, or is it just an hardware limitation due to my GPU?

from jukebox.

the function calculate_bandwidth by default iterates over about 600 sec of audio. Try passing a smaller duration to it.

The above error is not due to GPU. Could you try lowering 600 seconds to a smaller value?

Training the small_prior with a batch size of 2, 4, and 8 requires 6.7 GB, 9.3 GB, and 15.8 GB of GPU memory

2GB is perhaps not enough. You may want to make small_prior smaller in width or depth. For VQ-VQAE training, you may want to decrease --sample_length

from jukebox.