14

I am running PyTorch in Python 3.8.5 on a M1 Macbook Pro, and I get these error messages:

Intel MKL FATAL ERROR: This system does not meet the minimum requirements for use of the Intel(R) Math Kernel Library.
The processor must support the Intel(R) Supplemental Streaming SIMD Extensions 3 (Intel(R) SSSE3) instructions.
The processor must support the Intel(R) Streaming SIMD Extensions 4.2 (Intel(R) SSE4.2) instructions.
The processor must support the Intel(R) Advanced Vector Extensions (Intel(R) AVX) instructions.

What do they mean, and what can I do to resolve them?

mathwizurd
  • 1,387
  • 1
  • 13
  • 15
  • 1
    I was getting the same error messages the last few days (PyTorch had worked fine a week ago). For me what has worked so far has been: make a new conda environment. conda install nomkl. pip install torch. Doesn't feel robust, but wanted to pass it on. The idea to use pip instead of conda was from https://github.com/pytorch/pytorch/issues/71022 These are just the random steps that worked (so far) for me. I hope someone else gives a more correct answer. – CJD Feb 19 '22 at 21:59
  • I corrupted my conda env with the same error lately. I installed pytorch with some instructions for M1 I found online. Now anything I import raises this error. It is thrown when I do any import in python: "Intel MKL FATAL ERROR: This system does not meet the minimum requirements for use of the Intel(R) Math Kernel Library." – Intelligent-Infrastructure Mar 01 '22 at 09:22

4 Answers4

12

I ran into this error when trying to install an intel env via conda on an M1 mac. With a native env (osx-arm64) it worked, but I needed an intel env (osx-64) for some other reason.

I managed to resolve this error by downgrading mkl to version 2021 instead of version 2022 which was installed as a transitive dependency.

I simply added this to my environment.yml and re-installed it:

- mkl < 2022

Note that if your conda is M1 native (i.e. has a platform of osx-arm64), and you want to install an intel environment, you need to do it like so:

CONDA_SUBDIR=osx-64 conda env update -f environment.yml

The env also contains pytorch, and this approach allowed me to install everything via conda, instead of relying on a hybrid of conda/pip like the other solutions.

avivr
  • 1,393
  • 15
  • 28
  • 1
    This was exactly my issue: using Torch in an `osx-64` environment on an Apple M2 chip. This worked for me. One small aside: If you don't want to add this permanently to your `environment.yml` (because you want it to be cross-platform), then you can just use the one-time command: `conda install -c conda-forge "mkl<2022"`. See [here](https://stackoverflow.com/questions/71515117/how-to-set-up-a-conda-osx-64-environment-on-arm-mac) about creating Intel envs vs. AMD envs. – Neil Traft Aug 10 '23 at 23:03
7

As @CJD suggested installing PyTorch using pip apparently solves the problem, even on a Mac with M1 or newer versions M1 Pro and M1 Max [Tested]. Using pip will install PyTorch without MKL.

Steps:

  1. Uninstall PyTorch using conda (if it is installed): conda remove pytorch
  2. Install it using pip: pip install torch .

Check the following link for further details: https://github.com/pytorch/pytorch/issues/71022

Predator1432
  • 99
  • 3
2

I was getting the same error messages for any package I tried to import (ie, numpy, pandas, etc). I am running OSX Monterey on a late 2008 macbook 5,1 through opencore. The key bit for me was just installing nomkl as per CJD's comment above. It didn't matter whether it was a new or base environment. I don't really understand what the problem was or the possible limitations of the fix, but it is working now.

mhenn
  • 43
  • 5
  • ähm - its working because python is running as INTEL not ARM - check with Activity Monitor. I would have hoped pytorch is running native on arm64 architecture though.. – x y Feb 28 '22 at 12:15
0

Got the same error when trying Huggingface models. M1 Mac with conda environment.

Apparently "problem" is that conda installs packages for M1 Mac (who would have thought, huh) but not every library works with arm processor. This seems to be the case with Torch library.

How I fixed this.

  1. Created new env
  2. With pip install torch, torchvision
  3. conda install the rest of dependencies.

The problem might not be in Torch package itself but rather the way conda picks other dependencies. It looks for best ones for particular environment. On the other hand pip doesn't bother much to find most suitable ones particularly for arm processor and installs what works.

Ernest
  • 186
  • 10