[Bug] Error on Fine Tuning LFM2.5 230M

[jerinajay]

Note: Please do not remove the questions. Answer beside them.

1. Did you update? pip install --upgrade unsloth unsloth_zoo - in mac m5 not working
2. Colab or Kaggle or local / cloud - local - local mac air m5 32gb
3. Number GPUs used, use nvidia-smi - unsloth studio default settings

Collecting xformers>=0.0.27.post2 (from unsloth)
  Using cached xformers-0.0.35.tar.gz (4.3 MB)
  Installing build dependencies ... error
  error: subprocess-exited-with-error
  
  × installing build dependencies for xformers did not run successfully.
  │ exit code: 1
  ╰─> [5 lines of output]
      Collecting setuptools>=64
        Using cached setuptools-82.0.1-py3-none-any.whl.metadata (6.5 kB)
      ERROR: Ignored the following versions that require a different python version: 83.0.0 Requires-Python >=3.10
      ERROR: Could not find a version that satisfies the requirement torch>=2.10 (from versions: 1.9.0, 1.10.0, 1.10.1, 1.10.2, 1.11.0, 1.12.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0, 2.0.1, 2.1.0, 2.1.1, 2.1.2, 2.2.0, 2.2.1, 2.2.2, 2.3.0, 2.3.1, 2.4.0, 2.4.1, 2.5.0, 2.5.1, 2.6.0, 2.7.0, 2.7.1, 2.8.0)
      ERROR: No matching distribution found for torch>=2.10
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed to build 'xformers' when installing build dependencies for xformers
jerin@Jerins-MacBook-Air ~ % 

[merajstack]

The error is not actually about LFM2.5 230M. It's happening before training starts, while pip is trying to install xformers.

The key error is:

ERROR: Could not find a version that satisfies the requirement torch>=2.10

On an Apple Silicon (M5) Mac:

xformers is primarily designed for NVIDIA CUDA GPUs.
It is generally not supported on Apple's Metal (MPS) backend.
Many Unsloth features are primarily tested and optimized on Linux with NVIDIA GPUs. It's unclear whether this installation path is expected to work on Apple Silicon.

So pip install --upgrade unsloth unsloth_zoo fails because it attempts to install dependencies (xformers) that don't have compatible wheels for your platform.

This appears to be an Apple Silicon compatibility issue rather than an LFM2.5 issue.

Environment:

macOS on Apple M5 (32 GB RAM)
Python:
Local installation

Installation fails before training begins while installing xformers:

ERROR: Could not find a version that satisfies the requirement torch>=2.10
ERROR: Failed to build 'xformers'

Is Apple Silicon (MPS) currently supported for Unsloth Studio/local training, or is Linux + NVIDIA GPU required? If MPS is supported, is there a recommended installation method that skips xformers?

Also check

Run these commands and include their output:

python3 --version
pip --version
python3 -c "import platform; print(platform.platform())"
python3 -c "import torch; print(torch.version)"

One thing stands out:

torch>=2.10

At the time reflected by your error, PyPI only lists up to torch 2.8.0. If xformers is requesting torch>=2.10, it suggests one of these possibilities:

a packaging or dependency metadata bug,
an unreleased or incorrect xformers version being selected,
or an incompatibility with your Python version.

Including your Python version (python3 --version) will help determine which of those is most likely.

[alkinun]

unsloth==2026.7.4 on pypi still pulls xformers>=0.0.27.post2 by default on linux/win x86_64. pip resolves that to xformers 0.0.35, whose build metadata requires torch>=2.10, but PyTorch 2.10 is not available in the environment/index, so install fails before training starts.
the current repo pyproject.toml already moves torch/xformers out of default dependencies, so this likely needs a new pypi release built from the current metadata. also worth checking why the mac setup is resolving the linux/win x86_64 xformers marker, since native apple silicon shouldnt normally match it.
hope this helps.