#!/usr/bin/env bash
set -euo pipefail

export PATH="$HOME/.local/bin:${PATH}"

MODEL_REPO="empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF"
MODEL_FILE="Qwythos-9B-Claude-Mythos-5-1M-MTP-Q4_K_M.gguf"
MODEL_DIR="${MODEL_DIR:-/tmp/models}"
MODEL_PATH="${MODEL_DIR}/${MODEL_FILE}"

# Install the prebuilt llama.cpp binary. On this free CPU tier the installer
# probes the hardware and downloads the optimized CPU build.
if ! command -v llama >/dev/null 2>&1; then
  echo "==> Installing prebuilt llama.cpp via llama.app ..."
  curl -LsSf https://llama.app/install.sh | sh
fi

# Fetch the GGUF model at runtime (no token needed, the model is public).
mkdir -p "${MODEL_DIR}"
if [ ! -s "${MODEL_PATH}" ]; then
  echo "==> Downloading ${MODEL_FILE} ..."
  curl -fL --retry 3 --retry-delay 5 \
    -o "${MODEL_PATH}" \
    "https://huggingface.co/${MODEL_REPO}/resolve/main/${MODEL_FILE}"
fi

echo "==> Launching llama serve (web UI on :8080) ..."
exec llama serve \
  -m "${MODEL_PATH}" \
  --spec-type draft-mtp \
  --spec-draft-n-max 6 \
  -c 16384 \
  --host 0.0.0.0 \
  --port 8080