scripts/setup_llamacpp_stack.sh

#!/usr/bin/env bash
# One-shot setup for Well-Tuned + Llama Champion + Off the Grid badge stack.
set -euo pipefail
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
cd "$ROOT"

echo "==> 1/4 Export fine-tuned LoRA to GGUF"
python3 scripts/export_lora_gguf.py

echo "==> 2/4 Download Nemotron 4B base GGUF (skip if present)"
GGUF_DIR="${PMS_GGUF_DIR:-${ROOT}/models/nemotron}"
GGUF_FILE="${PMS_GGUF_FILE:-NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf}"
mkdir -p "$GGUF_DIR"
if [[ ! -f "${GGUF_DIR}/${GGUF_FILE}" ]]; then
  python3 -c "
from huggingface_hub import hf_hub_download
hf_hub_download('nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF', '${GGUF_FILE}', local_dir='${GGUF_DIR}')
"
fi

echo "==> 3/4 Build llama-server (skip if present)"
VENDOR_BIN="${ROOT}/vendor/llama.cpp/build/bin/llama-server"
if [[ ! -x "$VENDOR_BIN" ]]; then
  sudo apt-get update -qq
  sudo apt-get install -y -qq build-essential g++ libstdc++-12-dev cmake git
  CC=gcc CXX=g++ cmake -S vendor/llama.cpp -B vendor/llama.cpp/build \
    -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DBUILD_SHARED_LIBS=OFF
  cmake --build vendor/llama.cpp/build -j"$(nproc)" --target llama-server
fi

echo "==> Done. Start stack:"
echo "  Terminal 1: ./scripts/start_llamacpp.sh"
echo "  Terminal 2: PMS_INFERENCE_BACKEND=llamacpp python3 app.py"
echo "  Verify:     curl -s localhost:7860/api/health | python3 -m json.tool"