Unable to use PyTorch from Mojo
I just wanted to try Pytorch and created a project using nightly build.
I then installed PyTorch using: magic add "pytorch".
This is the code I tried:
from python import Python
fn main() raises:
torch = Python.import_module("torch")
x = torch.tensor([1, 2, 3])
print(x)
It resulted in the following error:
Unhandled exception caught during execution: libmkl_intel_lp64.so.2: cannot open shared object file: No such file or directory
mojo: error: execution exited with a non-zero result: 1
I can see "libmkl_intel_lp64.so.2" in the .magic\envs\default\lib folder so I do not understand what the problem is.18 Replies
Did you pull in Intel Python? Pytorch usually links to OpenBLAS.
How do I do that?
What does your
pixi.toml/mojoproject.toml file look like?This the content of momproject.toml:
[project]
channels = ["conda-forge", "https://conda.modular.com/max-nightly"]
description = "Add a short description here"
name = "mojo_by_example-nightly"
platforms = ["linux-64"]
version = "0.1.0"
[tasks]
[dependencies]
max = ">=25.1.0.dev2024121305,<26"
numpy = ">=1.26.4,<2"
jax = ">=0.4.35,<0.5"
pydicom = ">=3.0.1,<4"
torchvision = ">=0.20.1,<0.21"
pytorch = ">=2.5.1,<3"And you are inside of the magic environment when compiling and running?
Yes, I made sure to double-check.
One second, let me run it locally.
Ok, so I have oneAPI installed and it picked it up as my system BLAS.
Let me try that in a container.
I can reproduce it, so it's a Mojo issue.
@Gennadiy It's probably easiest to add the
mkl package to get the dependency while we sort out what exactly pytorch did, since it seems like this is their fault.Knowledge Base | conda-forge | community-driven packaging for conda
Finding examples with Github Code Search
Thanks. I will read through that page.
As a note, MKL is quite a bit faster than OpenBLAS, which is what pytorch and numpy used before.
So this is an upgrade, just one that appears to have had some consequences for the ecosystem.
Is there any performance-related issue given that I am on an AMD system?
What generation?
It's an AMD Ryzen 9 4900HS.
MKL's alleged "crippling of AMD CPUs" was actually around servers that didn't support AVX-512, and MKL only has AVX-512 and scalar fallbacks, but Zen 4 and 5 support AVX-512 and see the same perf as Intel CPUs.
That one is probably going to be a bit worse than openblas, since iirc it has AVX2.
But, this is not going to be a gigantic difference for most workloads unless you are doing AI on the CPU.
Ah, that;'s good to know.
One last question, can I add the dependency for openblas using magic and how do I do it? I am still figuring my way around mojo.
Congrats @Gennadiy, you just advanced to level 2!
The basic syntax is in that docs page I sent.
How exactly
pixi, the underlying tool for magic, decided to translate that to toml I don't know, you may have to play around a bit.OK, thanks. I will have a play around now.