An RVC Google Colab that ran on a TPU (Tensor Processing Unit) would be far quicker and more efficient than the GPU dedicated notebooks we currently have. Would there be any way or possible venture into writing a notebook that could utilise a Google TPU for faster training?
#Google Colabs TPU Implementation
spring iron
wicked havenBOT
whole apex
An RVC Google Colab that ran on a TPU (Tensor Processing Unit) would be far quic...
seems unlikely for the seemingly inactive RVC devs to bother on implementing TPU support
fickle pilot
That's possible in theory and may be even plug'n'play since pytorch has pytorch xla extension for cloud TPU devices. But may require more effort in general
But afaik, tpus work best with tensorflow, so it might be not worth it
wet ridge
That's possible in theory and may be even plug'n'play since pytorch has pytorch ...
Applio inference testing with the same model and a 1:25 min audio:
- T4 GPU: 14.77 seconds
- TPU v2: 40.31 seconds
- CPU: 425.75 seconds
wet ridge
But afaik, tpus work best with tensorflow, so it might be not worth it
yea, Tensor Processing Units as the name says, are better for TensorFlow, so knowing that RVC is built on pytorch it might not be the best
fickle pilot
Applio inference testing with the same model and a 1:25 min audio:
- T4 GPU: 14....
Nice. Have you tried training? I think TPU would benefit more from batched workload, inference runs on just one item
wet ridge
Nice. Have you tried training? I think TPU would benefit more from batched workl...
Tried but it froze /usr/lib/python3.10/multiprocessing/popen_fork.py:66: RuntimeWarning: os.fork() was called. os.fork() is incompatible with multithreaded code, and JAX is multithreaded, so this will likely lead to a deadlock.' when doing pitch extraction on my colab 
prolly have to fix smt as i did this fork https://github.com/Nick088Official/Applio-TPU just to mess around with TPUs
spring iron
Applio inference testing with the same model and a 1:25 min audio:
- T4 GPU: 14....
These are the current benchmarks as the code is GPU optimised, if the RVC devs optimised and adapted their code for TPU's then the processing speeds would be much faster. Especially training sessions.
spring iron
yea, Tensor Processing Units as the name says, are better for TensorFlow, so kn...
For inference the TPU might not outpreform to a worth while extent, but training on batches of data like TPU's are made to handle will be very worth while
slender oliveBOT
Ayo? @spring iron level 3 !!! 
wet ridge
For inference the TPU might not outpreform to a worth while extent, but training...
I don't think so, unless someone rewrites the whole code to Jax instead of pytorch
wet ridge
These are the current benchmarks as the code is GPU optimised, if the RVC devs o...
It can be done on pytorch but it's not worth it because pytorch doesn't have as good TPU support such as Jax
For example in stable diffusion they had to make a different Jax pipeline to make it worth it to use tpu
wet ridge
i searched up on internet and found this paper
@fickle pilot as we talked about TPUs here before, u might find what that paper says interesting