feat: add KV cache quantization args to server

[deceptech-packet-ninja]

Summary

• Adds --kv-bits, --kv-group-size, and --quantized-kv-start CLI arguments to mlx_lm.server
• Wires these through to stream_generate → generate_step, which already supports KV cache quantization
• Disables batching when kv_bits is set since BatchQuantizedKVCache does not exist yet

Closes #1043.

Motivation

KV cache quantization has been available via mlx_lm.generate CLI (--kv-bits) since v0.22, but the server has no way to enable it. Users running long-context inference through the OpenAI-compatible API cannot reduce KV cache memory usage.

Changes

mlx_lm/server.py (26 lines added):

• Three new CLI arguments in main(): --kv-bits (int, default None), --kv-group-size (int, default 64), --quantized-kv-start (int, default 0)
• In _serve_single: constructs kv_kwargs dict and passes to stream_generate via **kv_kwargs
• In _is_batchable: returns False when kv_bits is set (batched quantized cache NYI)

Test plan

• python -m mlx_lm.server --model mlx-community/Qwen2.5-0.5B-Instruct-4bit --kv-bits 4 starts without error
• Sending a request generates correct output
• Without --kv-bits, behavior is unchanged
• With --kv-bits, batching is correctly disabled (sequential path used)

🤖 Generated with Claude Code

[Thump604]

Clean server plumbing for #1074. The --kv-bits flag is needed for anyone running long-context inference through the API. I would use this immediately for my 122B production server.

One minor note: disabling batching when kv_bits is set makes sense for now, but BatchedEngine + quantized KV cache is a natural follow-up since continuous batching is where long-context memory pressure hits hardest.