has anyone been able to successfully run NeuralForecast on Apple Silicon leveraging GPU acceleration? i cannot get

tensorflow-metal

and

tensorflow

to play nice, even using aarm64 Docker image.

I have found this issue a few weeks ago with no solution for now. It just seems that pytorch is not finished implementing all the methods necessary for running properly with MPS. https://developer.apple.com/metal/pytorch/ https://github.com/Nixtla/neuralforecast/issues/709 The only thing that I I found that worked was to fall back to CPU using this:

import os

os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1'

at the beginning of your script.

I used the same in a Jupyter notebook:

%env PYTORCH_ENABLE_MPS_FALLBACK=1

It looks like some functions still run on GPU and only fallback to CPU when they are not available with MPS

MPS?

"Metal Performance Shaders" is a collection of highly optimized compute and graphics shaders designed to integrate into apps using Apple's Metal API.

any chance you can share your installation/setup? im having a lot of trouble with this. thanks

Sure, happy to help! I installed the following packages into a conda environment (first set from pip, second directly conda). Then I start a notebook in this environment and add the command I sent in my previous comment at the top of the notebook.

[pip.deps]
juliapkg = ""
ipywidgets = ""
ipykernel = ""
neuralforecast = ""
mlforecast = ""
juliacall = ""

[deps]
scikit-learn = ""
pytorch = ""
pandas = ""
matplotlib = ""
statsmodels = ""
seaborn = ""
numpy = ""

You can probably skip the julia-related packages.

and then you just set

gpus=1

in your NF model definition and it automatically picks up the GPU with

PYTORCH_ENABLE_MPS_FALLBACK=1

?

I did not specify the

gpus

at all.

and why is conda crucial here? what does the environment get you? (other than just virtualization)

conda is not crucial; the environment setup guarantees that all dependencies relevant to this project are in a "sandbox" and there are not conflicts among other packages that may be installed in the global system/environment.

ok good to know, thanks so much Dimitris