llama-server Options Guide

* --spec-mtp-cr-depth N
  Sets the MTP Compact Rollback depth. Recommended value is 1 with --spec-draft-n-max 5.
  If omitted, full MTP depth is used by default.

* --spec-draft-adaptive
    Dynamically adjusts the MTP draft limit, up to --spec-draft-n-max, based on
    recent draft acceptance. Most useful for less predictable text at large context.
    Default is off.

* --pipeline-parallel auto|on|off
  Controls pipeline parallelism. Set to off to prioritize available context memory.
  Default is auto (standard llama.cpp behavior).

* --hip-fa-force-vec on|off
  Forces the supported quantized-KV HIP Flash Attention VEC path. Enables larger context 
  at the cost of slower prompt processing speed.
  Default is off (ROCm/HIP only).