How can I install Pytorch?

[enesgur]

Hello,

I’m using Pyston for the AI scripts, but I've got some errors during the install Pytorch 1.8.0 package.

The error is “wheel is not supported on this platform”. I upgraded the wheel and pip but I failed.

I am using Debian 10 amd64 architecture. Pyston is 2.3.1.

[undingen]

Unfortunately there are no prebuild packages for pyston available yet which makes installing much harder.
You will have to follow their guide on installing it from source for now: https://github.com/pytorch/pytorch#from-source

[enesgur]

I see. Thanks for your reply.

[makslevental]

@undingen i'm comfortable compiling pytorch from source so i'd like to experiment with this. problem is i have no idea what this actually entails, i.e., do i just compile using your llvm dep? i'm imagining it must be something else along the lines of -fc++-abi=...?

[kmod]

It depends on how pytorch is configured -- but most projects automatically configure themselves to build against the Python interpreter that you use to run the build process, so it hopefully is as simple as saying pyston setup.py install

[makslevental]

@kmod thanks for the quick response. will report back on results :)

[undingen]

Also we do have a precompiled pytorch if you can use conda.
More infos about how to use conda with pyston: https://github.com/pyston/pyston/wiki/Using-conda

[makslevental]

@undingen interesting. so do you see perf improvements? the last blog post i read (not sure from when) said that you operate pytorch in fallback mode and therefore there's no perf improvement

[undingen]

We have not tested how the performance compares but it should be at least as fast as using PyTorch with CPython - there is no fallback/compatibility overhead. But we can only speedup Python code so if all of the pytorch code is running inside C/GPU etc Pyston will not speed it up.
Would be really interested in hearing what performance you are seeing if you end up experimenting with pyston.