CUDA memory requirement about mp-senet HOT 5 CLOSED

I used two 3090ti GPUs to train the MP-SENet model, each with a memory size of 24 GB.

from mp-senet.

@yxlu-0102
Is it the same for inference?

from mp-senet.

One GPU with a memory size of 24 GB is enough for the inference process.

from mp-senet.

@yxlu-0102
I just tried inferencing on my 3090ti (24gb vram), and I am still getting OOM error. Is there a way to reduce the memory usage?

from mp-senet.

Due to the variable length of speech, the memory consumption for generating each speech is not fixed, and it doesn't depend on whether the GPU is a 3090 or not. It primarily depends on the amount of GPU memory you have.

During training, you can reduce GPU memory usage by adjusting the batch size and segment length, but this may have an impact on the training results.

from mp-senet.