Refactor Server Configuration to Dataclass for Python 3.14 Compatibility and Programmatic Usage

[giskarda]

Context
Currently, the mlx-lm server relies on argparse.Namespace to pass configuration around. While this works for CLI usage, it presents some limitations:

Python 3.14 Compatibility: Newer versions of Python are stricter with dataclass validation, leading to ValueError or AttributeError when handling mutable defaults or missing attributes in certain environments.

Programmatic Access: It is currently difficult to launch the server as a library within other Python scripts without manually mocking argparse objects.

Robustness: Accessing aliases like .temp or .num_tokens directly on the config object can lead to runtime crashes if the object isn't perfectly initialized.

Proposal
I have developed a refactor that introduces a structured MLXServerConfig dataclass. This change:

Implements default_factory: Properly handles mutable types (lists/dicts) to ensure compatibility with modern Python versions and test runners.

Adds run_server(config): A new entry point that allows launching the server by passing a config object, making mlx-lm easier to integrate into existing pipelines.

Benefits

Full compatibility with Python 3.14.

Improved testability (cleaner mocking of server settings).

Better developer experience for those using MLX programmatically.

I have already implemented these changes along with a new test suite in tests/test_server.py. I would love to hear the maintainers' thoughts on this direction before opening a formal Pull Request.

Looking forward to your feedback!