VLLM Metal Engine Error (`ValueError: Model type qwen3_5 not supported` when running `Qwen3.5-4B-MLX-4bit` via `docker model run`)

[emi-dm]

Description:
When attempting to run the Qwen3.5-4B-MLX-4bit model from the MLX community using the docker model run command, the model fails to preload. The underlying vllm-metal runner crashes with a ValueError indicating that the qwen3_5 model type is not supported by mlx_lm.

Steps to Reproduce:

1. Run the following command in the terminal:

   docker model run hf.co/mlx-community/Qwen3.5-4B-MLX-4bit

2. Enter a prompt (e.g., > Hola).
3. Observe the crash and the traceback.

Expected Behavior:
The model should load successfully and generate a response to the prompt.

Actual Behavior / Error Logs:
The background model preload fails with a status 500. Here is the traceback:

background model preload failed: preload failed: status=500 body=unable to load runner: error waiting for runner to be ready: vllm-metal terminated unexpectedly: vllm-metal failed:    model, config = load_model(model_path, lazy, model_config=model_config)
                                                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/emi/.docker/model-runner/vllm-metal/lib/python3.12/site-packages/mlx_lm/utils.py", line 201, in load_model
    model_class, model_args_class = get_model_classes(config=config)
                                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/emi/.docker/model-runner/vllm-metal/lib/python3.12/site-packages/mlx_lm/utils.py", line 75, in _get_classes
    raise ValueError(msg)
ValueError: Model type qwen3_5 not supported.
packages/vllm_metal/server.py", line 288, in create_engine
    runner.load_model()
  File "/Users/emi/.docker/model-runner/vllm-metal/lib/python3.12/site-packages/vllm_metal/model_runner.py", line 51, in load_model
    self.model, self.tokenizer = mlx_load(
                                 ^^^^^^^^^
  File "/Users/emi/.docker/model-runner/vllm-metal/lib/python3.12/site-packages/mlx_lm/utils.py", line 322, in load

Environment details:

• OS: macOS (Apple Silicon) M3
• Command line tool: Docker Desktop local AI features (docker model CLI)

Additional Context:
It appears the version of mlx_lm bundled with the Docker model-runner (vllm-metal) might be outdated and missing support for the qwen3_5 architecture, which was added in more recent versions of the MLX framework.

[doringeman]

Hey @emi-dm,

On it! Are you willing to run Model Runner from this repo directly or do you want a full new Docker Desktop build?
Once merged, the changes from #769 will become available sometime next week in Docker Desktop, in its new version.

[emi-dm]

Hi @doringeman ! Thanks for your very quick response! I prefer the Docker Desktop build to run it, for me is quicker!

[doringeman]

Hey @emi-dm, wanna give this build a go? 😀
https://desktop-stage.docker.com/mac/main/arm64/222295/Docker.dmg

This will be part of next week Docker Desktop update, too.