#!/usr/bin/env bash
# Download the Huihui Qwen3.8-27B AWQ W4A16 checkpoint (auto-round INT4, ~20 GB)
# into the Hugging Face cache that vLLM reads (HF_HOME=/storage-models/vllm).
#
# Usage:
#   ./download-model.sh
#   HF_HOME=/some/other/cache ./download-model.sh
#
# Re-running is cheap: huggingface_hub resumes/verifies already-downloaded files.
set -euo pipefail

REPO="shawnw3i/Huihui-Qwen3.8-27B-abliterated-AWQ-MTP"
export HF_HOME="${HF_HOME:-/storage-models/vllm}"

SNAPSHOT_GLOB="$HF_HOME/hub/models--${REPO//\//--}/snapshots/*/config.json"
if compgen -G "$SNAPSHOT_GLOB" >/dev/null; then
    echo "Already present in $HF_HOME/hub ($(du -sh "$HF_HOME/hub/models--${REPO//\//--}" | cut -f1))"
    echo "Nothing to do. (Delete that cache entry to force a re-download.)"
    exit 0
fi

if command -v hf >/dev/null 2>&1; then
    hf download "$REPO"
elif command -v huggingface-cli >/dev/null 2>&1; then
    huggingface-cli download "$REPO"
else
    echo "No Hugging Face CLI found. Install it first:" >&2
    echo "  pip install -U 'huggingface_hub[cli]'" >&2
    exit 1
fi

echo
du -sh "$HF_HOME/hub/models--${REPO//\//--}"
echo "Done. Next: docker compose up -d  (first start compiles kernels / captures CUDA graphs)"