#!/bin/sh # FREESTREAM sovereign LLM server on the mac mini. # ./serve_local_llm.sh mlx -> MLX (Apple Silicon native, fastest on M-series) # ./serve_local_llm.cpp -> llama.cpp (brew install llama.cpp) # Both expose OpenAI-compatible APIs; point FREESTREAM at: # export FREESTREAM_LLM_BASE_URL=http://localhost:8080/v1 # export FREESTREAM_LLM_MODEL=local # export FREESTREAM_AGENT=llm set -e MODE="${1:-mlx}" PORT="${PORT:-8080}" case "$MODE" in mlx) # python3.10+ with Apple Silicon; 4-bit quant runs fast on M-series pip install -q "mlx-lm>=0.19" || true exec python -m mlx_lm.server \ --model "${MODEL:-mlx-community/Qwen2.5-7B-Instruct-4bit}" \ --host 0.0.0.0 --port "$PORT" ;; cpp) # brew install llama.cpp exec llama-server \ -hf "${MODEL_HF:-bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M}" \ --host 0.0.0.0 --port "$PORT" \ -c "${CONTEXT_SIZE:-4096}" ${API_KEY:+--api-key "$API_KEY"} ;; *) echo "usage: serve_local_llm.sh [mlx|cpp]" >&2 exit 2 ;; esac