feat: Add MLX Manifold Pipeline

[justinh-rahb]

Introducing the MLX Manifold Pipeline (v2.0)

This PR introduces a significant enhancement to the MLX pipeline with the MLX Manifold Pipeline, located in a new file: mlx_manifold_pipeline.py. This new pipeline offers major improvements:

Manifold Support:

• The pipeline automatically detects and lists all available HF models with a defined filter keyword.
• This removes the need for manual model configuration and makes model selection easier.

Dynamic Model Loading:

• The pipeline dynamically loads and uses the selected MLX model when run.
• This eliminates static model definitions and allows for more flexible model experimentation and deployment.

Version 2.0:

• This update is a major version bump (v2.0) due to the substantial changes introduced.
• The existing MLX pipeline remains untouched to ensure backward compatibility.

Migration:

• Users can switch to mlx_manifold_pipeline.py to use the new Manifold Pipeline, but should not run both on the same MLX backend server.

Testing:

• The MLX Manifold Pipeline needs more thorough testing before merging, help wanted!

Using the Valves:

• MLX Model Filter: Filter available models to show only ones matching the filter keyword.
• MLX Default Model: Set a default model to be automatically loaded.
  • If not found locally, it will be downloaded (similar to Ollama's pull feature).
• MLX Stop: Define stop words that will trigger the inference to end.
• MLX Chat Template: Specify a custom chat template to use.
• MLX Use Default Chat Template: True/False (default False)
• HuggingFace Token: Provide your token for accessing private models.

Note: Setting an environment variable for your HuggingFace cache can point to an alternate model location, bypassing the need for the MLX Model Pattern filter.

---

This enhancement makes the MLX pipeline more usable and flexible, allowing users to better utilize their MLX models.

[tjbck]

Can't wait to try this 🤩 Thanks!

[justinh-rahb]

Forgot the teasers 😬

Screen.Recording.2024-06-30.at.12.23.31.AM.mp4

[justinh-rahb]

TODO for MLX Manifold Pipeline PR:

• Remove MLX_SUBPROCESS=False option: This is essential for the dynamic model loading. DONE
• Cleanup/Rename Valves: Improve clarity by cleaning up and renaming some Valves. DONE
• Consider a helper script: Explore creating a script that simplifies installation and running of the MLX Pipelines server on macOS.
• Investigate mlx_lm direct usage (v2.0 ONLY): Since MLX_SUBPROCESS=False is being removed in v2.0, directly using the mlx_lm library instead of mlx_lm.serve might be possible and more efficient. This needs further investigation.
• Document containerization limitations: Running Pipelines on bare metal is required due to the removal of the subprocess option.

[tjbck]

Let's goooooo 🚀