Feature/kv cache quantization

[cubist38]

feat: add KV cache quantization support for LM and multimodal models

Closes #276

Summary

• Add --kv-bits, --kv-group-size, and --quantized-kv-start CLI options and YAML config fields to enable KV cache quantization (including TurboQuant) for lm and multimodal model types
• Thread the parameters through config, handlers, and model wrappers to the underlying generate_step in both mlx-lm and mlx-vlm
• Validate and warn when KV cache quantization is specified for unsupported model types (e.g. embeddings, whisper)

Usage

CLI

mlx-openai-server launch --model-path <model> --kv-bits 4 --kv-group-size 64 --quantized-kv-start 0

YAML config

models:
  - model_path: mlx-community/Qwen3-VL-2B-Thinking-8bit
    model_type: multimodal
    kv_bits: 4
    kv_group_size: 64
    quantized_kv_start: 0

Changed files

• app/cli.py — New --kv-bits, --kv-group-size, --quantized-kv-start CLI options
• app/config.py — New fields on MLXServerConfig and ModelEntryConfig with validation
• app/server.py — Pass KV params to LM and VLM handler constructors
• app/handler/mlx_lm.py — Accept, store, and include KV params in model_params
• app/handler/mlx_vlm.py — Accept, store, and include KV params in model_params
• app/models/mlx_lm.py — Forward KV params to stream_generate
• app/models/mlx_vlm.py — Forward KV params to stream_generate