#!/usr/bin/env bash
# Serve Cortex.0 with llama.cpp on an OpenAI-compatible endpoint.
#
# Usage: ./scripts/serve_llamacpp.sh [model.gguf] [port]
set -euo pipefail

MODEL="${1:-Cortex.0-Q8_0.gguf}"
PORT="${2:-8080}"

if ! command -v llama-server >/dev/null 2>&1; then
  echo "llama-server not found. Build llama.cpp: https://github.com/ggml-org/llama.cpp" >&2
  exit 1
fi

if [ ! -f "$MODEL" ]; then
  echo "Model file '$MODEL' not found." >&2
  echo "Download it: hf download Frankenstein-Labs/Cortex.0 Cortex.0-Q8_0.gguf" >&2
  exit 1
fi

# --jinja makes llama.cpp apply the model's own chat template, which is what
# carries the tool-call protocol Flint was trained with.
exec llama-server \
  --model "$MODEL" \
  --port "$PORT" \
  --ctx-size 32768 \
  --n-gpu-layers 999 \
  --jinja \
  --host 127.0.0.1