Homebrew ollama 0.15.5 on Apple Silicon prints MLX: Failed to load symbol: mlx_metal_device_info with mlx-c 0.5.0 (go generate/build tags mismatch)

[ohzls]

brew config AND brew doctor output OR brew gist-logs <formula> link

HOMEBREW_VERSION: 5.0.14-19-gea0351c
ORIGIN: https://github.com/Homebrew/brew
HEAD: ea0351ca9dda55624657ee78812f399c6c0e8113
Last commit: 17 hours ago
Branch: main
Core tap JSON: 10 Feb 04:58 UTC
Core cask tap JSON: 10 Feb 04:58 UTC
HOMEBREW_PREFIX: /opt/homebrew
HOMEBREW_CASK_OPTS: []
HOMEBREW_DOWNLOAD_CONCURRENCY: 28
HOMEBREW_FORBID_PACKAGES_FROM_PATHS: set
HOMEBREW_MAKE_JOBS: 14
HOMEBREW_NO_COLOR: set
HOMEBREW_SORBET_RUNTIME: set
Homebrew Ruby: 3.4.8 => /opt/homebrew/Library/Homebrew/vendor/portable-ruby/3.4.8/bin/ruby
Clang: 17.0.0 build 1700
Git: 2.52.0 => /opt/homebrew/bin/git
Curl: 8.7.1 => /usr/bin/curl
macOS: 26.2-arm64
CLT: 26.2.0.0.1.1764812424
Xcode: 26.2

(AND)

seongwonseo@seongwonui-Macmini ~ % brew doctor
Your system is ready to brew.
seongwonseo@seongwonui-Macmini ~ %

Verification

• My brew doctor output says Your system is ready to brew. and am still able to reproduce my issue.
• I ran brew update and am still able to reproduce my issue.
• I have resolved all warnings from brew doctor and that did not fix my problem.
• I searched for recent similar issues at https://github.com/Homebrew/homebrew-core/issues?q=is%3Aissue and found no duplicates.
• My issue is not about a failure to build a formula from source.

What were you trying to do (and why)?

I installed Ollama via Homebrew on Apple Silicon and tried to run it (ollama serve / ollama -h) for local LLM inference.

What happened (include all command output)?

After installing ollama (0.15.5) and mlx-c (0.5.0), running Ollama prints an MLX dynamic loader warning:

$ /opt/homebrew/Cellar/ollama/0.15.5/bin/ollama -h
MLX: Failed to load symbol: mlx_metal_device_info
...

(or)

$ OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama serve
MLX: Failed to load symbol: mlx_metal_device_info
...
Listening on 127.0.0.1:11434 (version 0.15.5)

Also, mlx-c 0.5.0 does not export that symbol:
$ nm -gU /opt/homebrew/Cellar/mlx-c/0.5.0/lib/libmlxc.dylib | rg "mlx_metal_device_info"
(no output)

What did you expect to happen?

Ollama should start without MLX missing-symbol warnings. The packaged build should have MLX wrappers consistent with the installed mlx-c version.

Step-by-step reproduction instructions (by running brew commands)

brew update
brew install mlx-c ollama

# Repro (either one):
/opt/homebrew/bin/ollama -h
# or
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 /opt/homebrew/bin/ollama serve

[Elusv]

also happens on brew reinstall ollama as upgrade from prior version. mlx-c already at 0.5.0

[p-linnane]

This is indeed an issue with mlx-c being upgraded to an incompatible version. Unfortunately, the ollama formula doesn't have any sort of test related to mlx support. If you can propose a test that doesn't require downloading a huge amount of data, we could prevent this in the future.

[ghuh]

Just running ollama ps or ollama list without any models loaded causes "MLX: Failed to load symbol: mlx_metal_device_info" to get written to stderr.

So, I suspect just running one of those (after starting ollama) and checking if anything is written to stderr would be a sufficient check. No need to download any models or other large data.

Something like:

brew install ollama
brew services start ollama
ollama list 2>/tmp/ollama_stderr_check
[[ -s /tmp/ollama_stderr_check ]] && echo "failed"

If you wanted to check for MLX specifically, you can check if the error message starts with ^MLX:

[p-linnane]

I'm not getting any type of error when running ollama list or ollama ps after a fresh install with no models.

[ghuh]

Oh interesting, I have models downloaded, but not running. That sucks, so apparently a model needs to be download to trigger the error.

Unless anyone else has any ideas?

The smallest model I could find is tinyllama weighing in at 638MB which is probably more than you want to download for a simple test.

Anyone else know of a smaller model?

[p-linnane]

@jmorganca @dhiltgen sorry for the ping, but could one of you weigh in on the above? We're looking for a minimal way to test MLX without downloading an entire model. Just ensuring it doesn't emit errors would be enough.