PyTorch Installation Failing via Pixi on macOS ARM

[kennyjwilli]

I'm trying to install PyTorch using Pixi on macOS ARM M1 following the official guide, but encountering a symbol lookup error.

Failed Attempt: Installing via Pixi

When trying to verify the PyTorch installation:

pixi run --environment cpu python -c "import torch; print(torch.cuda.is_available())"

I get the following error:

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/Users/kenny/work/pytorch-test/.pixi/envs/cpu/lib/python3.11/site-packages/torch/__init__.py", line 367, in <module>
    from torch._C import *  # noqa: F403
    ^^^^^^^^^^^^^^^^^^^^^^
ImportError: dlopen(/Users/kenny/work/pytorch-test/.pixi/envs/cpu/lib/python3.11/site-packages/torch/_C.cpython-311-darwin.so, 0x0002): Symbol not found: __ZN4absl12lts_2024072212log_internal10LogMessagelsIiLi0EEERS2_RKT_
  Referenced from: <7DD6F527-7A4A-3649-87B6-D68B25F8B594> /Users/kenny/work/pytorch-test/.pixi/envs/cpu/lib/libprotobuf.28.2.0.dylib
  Expected in:     <D623F952-8116-35EC-859D-F7F8D5DD7699> /Users/kenny/work/pytorch-test/.pixi/envs/cpu/lib/libabsl_log_internal_message.2407.0.0.dylib

Working Alternative: Installing via Conda

The same installation works successfully using Conda:

conda create -n dev
conda install pytorch -c pytorch
python3 -c "import torch; print(torch.cuda.is_available())"
False

Environment Details

Here's my current configuration:

System Info (pixi info)

System
------------
       Pixi version: 0.39.4
           Platform: osx-arm64
   Virtual packages: __unix=0=0
                   : __osx=14.6.1=0
                   : __archspec=1=m1
          Cache dir: /Users/kenny/Library/Caches/rattler/cache
       Auth storage: /Users/kenny/.rattler/credentials.json
   Config locations: No config files found

Global
------------
            Bin dir: /Users/kenny/.pixi/bin
    Environment dir: /Users/kenny/.pixi/envs
       Manifest dir: /Users/kenny/.pixi/manifests/pixi-global.toml

Project
------------
               Name: pytorch-conda-forge
      Manifest file: /Users/kenny/work/pytorch-test/pixi.toml
       Last updated: 31-12-2024 15:13:40

Environments
------------
        Environment: default
           Features: default
           Channels: https://prefix.dev/conda-forge
   Dependency count: 0
   Target platforms: osx-arm64

        Environment: cpu
           Features: cpu, default
           Channels: https://prefix.dev/conda-forge
   Dependency count: 1
       Dependencies: pytorch-cpu
   Target platforms: osx-arm64

Project Configuration (pixi.toml)

[project]
channels = ["https://prefix.dev/conda-forge"]
name = "pytorch-conda-forge"
platforms = ["osx-arm64"]

[feature.cpu.dependencies]
pytorch-cpu = "*"

[environments]
cpu = ["cpu"]

Note: I only modified the platforms field to "osx-arm64" and removed the gpu bits since it'd cause a failure to resolve from the example in the docs.

Could someone help identify why the Pixi installation is failing while the Conda installation works?

[kennyjwilli]

Of course I found a solution shortly after posting! Following the "Installing from PyPi" section of the documentation worked successfully on my MacOS ARM system:

pixi run python -c "import torch; print(torch.cuda.is_available())"

Output:

/Users/kenny/work/pytorch-test/.pixi/envs/default/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:295: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /Users/runner/work/pytorch/pytorch/pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)
  cpu = _conversion_method_template(device=torch.device("cpu"))
False

This raises a question: Is it expected that the Conda-forge installation method doesn't work on MacOS ARM while the PyPi installation does? Some clarification in the documentation about platform-specific considerations would be helpful.

Note: While this works, there appears to be a missing NumPy dependency that should probably be added to the configuration.

[ruben-arts]

Hey I just tested on my M4 mac and it seems to work like expected:
pixi.toml:

[project]
channels = ["https://prefix.dev/conda-forge"]
name = "pytorch"
platforms = ["osx-arm64"]

[dependencies]
pytorch = "*"

> pixi run python -c "import torch; print(torch.__version__); print(torch.cuda.is_available())"
2.5.1.post8
False