tortoise-tts
tortoise-tts copied to clipboard
neonbjb
High processor load and no GPU load in WSL2 Ubuntu
I installed the app in ubuntu in wsl2. CUDA is available, everything is working(audio is generated), and tests say that CUDA is available
import torch print(torch.cuda.is_available()) // true
torch.cuda.get_device_name(0) // NVIDIA GeForce RTX 3080
But when I launch audio generation, my processor core(one of the 4) is at max load, and GPU is at zero load.
What might be the problem?
If you reinstall torch inside of your environment a link will be made between tortoise-tts and torch. You're not using your card, just the CPU.
I'm having the same issue on Windows 10 x64. Works fine but no GPU usage and takes a while on CPU.
Verified the same things that OP did. Not sure what the solution is, don't quite understand plincter10's comment. Re-innstalling torch didn't seem to change anything.
Mangaroo, you create an environment first, with python or conda and you put all your software inside the environment. You may have torch installed in your computer as a whole and it checks out fine but as soon as you create an env. it doesn't see the torch. Start fresh and delete everything and start over. Creating the environment is key. You can download the zip file and it will do everything for you. It just won't update, you have to download a new zip file to update. You can also try this fork, it works OK in a Gradio gui. https://git.ecker.tech/mrq/ai-voice-cloning/wiki You can run it alongside your other install. Installed quick.
Thank you for the reply, I actually did set it all up using conda. Would it even be outputting results if it didn't see the torch?
Running the commands in the OP through Anaconda Prompt shows the same results (cuda is available and the GPU) but it's only really using my CPU.
I'm having the same issue. Only the CPU seems to be getting utilized. I have tested narviii's commands and torch.cuda.is_available() returns True. The device name returns "NVIDIA GeForce RTX 2060" as it should.
I have installed tortoise-tts without Conda, only using pip. Is there any way to force it to use the GPU? How do I set up the environment if I have to do that? Any way to force it to use the GPU without starting everything over?
My operating system is Windows 10.
I have the same issue with gtx4090. I tried using pytorch with 11.7, 11.8, nightly. Nothing works.
Same issue with gtx1080Ti on Windows 10. I have Cuda 12 drivers installed. I'll try downgrading to Cuda 11.8 and check if it works.
Mangaroo, you create an environment first, with python or conda and you put all your software inside the environment. You may have torch installed in your computer as a whole and it checks out fine but as soon as you create an env. it doesn't see the torch. Start fresh and delete everything and start over. Creating the environment is key. You can download the zip file and it will do everything for you. It just won't update, you have to download a new zip file to update. You can also try this fork, it works OK in a Gradio gui. https://git.ecker.tech/mrq/ai-voice-cloning/wiki You can run it alongside your other install. Installed quick.
You are a ultimate life-saver, friend. That Gradio gui repo you have linked works smoother than butter. 🫶🏻
Btw, I found that Tortoise works very well just in Windows, no need for WSL2, actually
@narviii And did you do anything in particular for it to work with the GPU on Windows? I'm facing the same problem; CUDA is available, it finds the GPU with the correct name but, my processor is at max load, and GPU is at zero load, like yours.
Edit: Okay, nevermind just happened to be looking at GPU whenever it was at 0-1%, but it was at 88ºC so it definitely had some peaks of doing things.
Same here torch and cuda are installed correctly but tortoise is not using it
