Packages not found error about miniforge

[hmaarrfk]

I guess we are closing because:

1. We know this is a very popular package.
2. Packaging this for OSX-arm64 is going to be challenging.
3. Repackaging everything on conda-forge for OSX-arm64 is already on the roadmap.

i hope you have fun with your new toy @slerman12 :D

[rgommers]

There aren't any released pytorch packages for osx-arm64 yet, so this is not surprising. You can grab a nightly wheel or build from source I think, see pytorch/pytorch#48145

[slerman12]

Thanks, I'll try that. Would installing Gym/atari work?

[hmaarrfk]

i'm not so sure. You will have to experiment with functionaly.

Honestly, hold on to your old workstation for a while. The transition won't be instant, and working with bleeding edge hardware accelerated software is going to take a while.

https://anaconda.org/pytorch/pytorch/files

Pytorch doesn't list any build for osx other than osx-64 so your mileage may varry.

[slerman12]

I was able to install pytorch from source. Works, no problem. Is there a way I can now use it with Gym? Can I install gym without the apple silicon compatibility?

[hmaarrfk]

again, i'm not sure.

[slerman12]

Is there something wrong with the NumPy distribution? Running np.empty(2) returns an array of non-NaNs.

[rgommers]

Is there something wrong with the NumPy distribution? Running np.empty(2) returns an array of non-NaNs.

That's normal. empty doesn't give you nans, it gives you whatever random data is already present in the block of memory it grabs.

[slerman12]

Ah yes, I thought that was the cause of the inconsistency in performance of my code but it must be something else then. It’s very odd. My model performs well on my old MacBook, but the exact same code gets much worse results on my new M1 MacBook with Pytorch and gym installed from source, and everything else installed with miniforge’s conda.