Spaces:

r3gm

/

SoniTranslate_translate_audio_of_a_video_content

like 114

Running on Zero

same

Thank you very much for your work, I was still experiencing 1000 'CUDA run out of memory' or 'Float=16 not supported' without modifying anything. At first it worked, then an update broke everything. I reinstalled my entire computer with Pop_Os with version Pytorch 2.2.2_12.1 and Cuda 12.1 and Cudnn. Nvidia drivers 550 internally. Reading Whisperx we learn that the large model is for 8GB GPU and mine is 6GB. I lowered the batch to 8 and put Int8 on a medium or small model in gradio, and it works faster than with the installation of torch in CPU which is very slow. I updated the requirements leaving only the Whisperx repository and it works faster than with Colab with my Geforce 1060. Here is a copy of my conda env. https://disk.yandex.com/d/Y3UnUfgGx830kg I still have the problem of Omegaconf which is incompatible with Pyannote_3.1 . But it finally works.

Thanks for sharing your experience and setup. It's impressive how you optimized it despite GPU constraints. To deal with the compatibility issues with Pyannote, I made changes to the requirements in my own fork.

What does it mean "Specifify a media or SRT file in advanced settings"? I can't figure out how to fix it.

"504 Gateway Time-out"

:(