RVC-Project
Training on Apple Silicon
Hey all,
I used my M1 MacBook Pro and trained a model with a dataset with .wav format that is 30 min long without slicing it, and I have some questions:
- Is it reasonable that it took more than 6 hours?
- Do slicing the dataset into < 10sec for every part, can make the training process faster?
- Is it that slow because RVC still doesn't support Apple Silicon, and the GPU is not being used?
- Are there any MPS updates coming to make training faster?
Hope some of you can help me with this 👯
Hi eliasnassar,
I think I can answer your questions.
- Yes. It takes long time to train even for 10 min long, I left it overnight.
- Not sure, but slicing the dataset should be necessary.
- No, you can absolutely train a model using Apple Silicon. And the GPU is being used but some function not supported with MPS (Metal Performance Shaders), so it will use the CPU as a fallback :(
- I hope so. I saw some of the RVC developer use Macs as well :D
Hi eliasnassar,
After I finished using m2 max, I encountered the following error when executing python ./infer-web.py
Traceback (most recent call last):
File "/Users/guhaizhou/work/AI/RVConversion-WebUI/infer-web.py", line 21, in
It's obvious that I saw faiss cpu=>1.7.4 in requirements. txt
I really hope to receive your help. Thank you！！
@eliasnassar I can’t even get my to train on Mac m2. I had to use a windows gpu pc.
@Naozumi520 What particular part of the training pipeline isn't compatible with MPS?
@Naozumi520 What particular part of the training pipeline isn't compatible with MPS?
Not sure, but you can find out with the line os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1" removed. This will stop using cpu as a fallback while training.
i have the same problem everything comes up with a big red ERROR when i boot up the live link.
