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

repo_id="${REPO_ID:-Blackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF}"
quant="${QUANT:-Q4_K_M}"
download_dir="${MODEL_DIR:-./Qwen3.8-27B-ABLITERATED-GGUF}"
context_size="${CTX_SIZE:-16384}"
gpu_layers="${GPU_LAYERS:-999}"
host="${HOST:-0.0.0.0}"
port="${PORT:-8080}"
enable_vision="${ENABLE_VISION:-0}"
projector_type="${MMPROJ_TYPE:-Q8_0}"

model_file="Qwen3.8-27B-ABLITERATED-${quant}.gguf"
model_path="${download_dir}/${model_file}"

if ! command -v hf >/dev/null 2>&1; then
  echo "Missing 'hf' CLI. Install it with: pip install -U huggingface_hub"
  exit 1
fi

if ! command -v llama-server >/dev/null 2>&1; then
  echo "Missing llama-server. Install a current llama.cpp build and place llama-server on PATH."
  exit 1
fi

mkdir -p "${download_dir}"
hf download "${repo_id}" "${model_file}" --local-dir "${download_dir}"

args=(
  -m "${model_path}"
  -ngl "${gpu_layers}"
  -fa on
  --jinja
  --host "${host}"
  --port "${port}"
  -c "${context_size}"
  --temp 1.0
  --top-p 0.95
  --top-k 20
)

if [[ "${enable_vision}" == "1" ]]; then
  projector_file="mmproj-Qwen3.8-27B-ABLITERATED-${projector_type}.gguf"
  projector_path="${download_dir}/${projector_file}"
  hf download "${repo_id}" "${projector_file}" --local-dir "${download_dir}"
  args+=(--mmproj "${projector_path}")
fi

exec llama-server "${args[@]}"