feat: batch text-only MLX VLM requests

[qinxuye]

Summary

• enable continuous batching for text-only MLX VLM requests via the underlying language model
• keep image requests on the mlx-vlm path
• isolate MLX batch generator state per model instance and add a VLM text-only parallel regression test

Tests

• python -m py_compile xinference/model/llm/mlx/core.py xinference/model/llm/mlx/tests/test_mlx.py
• python -m black --check xinference/model/llm/mlx/core.py xinference/model/llm/mlx/tests/test_mlx.py
• PYTHONPATH=... python -X faulthandler -m pytest xinference/model/llm/mlx/tests/test_mlx.py::test_mlx_vision_text_only_parallel_inference -q -s
• PYTHONPATH=... python -X faulthandler -m pytest xinference/model/llm/mlx/tests/test_mlx.py::test_load_mlx_vision -q -s

[gemini-code-assist]

Code Review

This pull request enables continuous batching for text-only requests in MLX Vision models by introducing the _MLXLogitsModelAdapter and refactoring MLXBatchModel to use instance-level state instead of class-level variables. It also includes a new test case to verify parallel inference for text-only prompts. Feedback was provided to simplify the logic in _is_text_only_prompt for better readability and maintainability.

[qinxuye]

/gemini review

[gemini-code-assist]

Code Review

This pull request enables continuous batching for text-only requests in MLX vision models by introducing a logits adapter and routing logic. It also refactors MLXBatchModel to use instance-level storage for improved state isolation. A review comment suggests making the text-only prompt detection more robust by checking for additional modalities like video and audio.

[qinxuye]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors MLXBatchModel to use instance-level state instead of class-level variables and enables continuous batching for text-only prompts within MLXVisionModel. It introduces a _MLXLogitsModelAdapter to bridge mlx-vlm models with the mlx-lm API and includes new tests for prompt detection and parallel inference. Feedback was provided to move the _lock attribute in MLXBatchModel from the class level to the instance level to prevent unnecessary performance contention when multiple models are running in parallel.