#!/bin/bash
# Script to pre-download the Llama model on a login node (which has internet access)
# Run this BEFORE submitting the Slurm job

set -euo pipefail

MODEL_NAME="${1:-meta-llama/Llama-3.2-3B-Instruct}"
HF_HOME="${HF_HOME:-/scratch/user/hangxiao/huggingface_cache}"

echo "Downloading model: ${MODEL_NAME}"
echo "Cache directory: ${HF_HOME}"

export HF_HOME
export TRANSFORMERS_CACHE="${HF_HOME}"
export HF_TOKEN="${HF_TOKEN:-}"

mkdir -p "${HF_HOME}"

source /sw/eb/sw/Anaconda3/2024.02-1/etc/profile.d/conda.sh
conda activate /scratch/user/hangxiao/.conda/envs/paper_impact

# Use huggingface-cli to download the model (simpler and faster)
if command -v huggingface-cli &> /dev/null; then
    echo "Using huggingface-cli to download model..."
    huggingface-cli download "${MODEL_NAME}" \
        --cache-dir "${HF_HOME}" \
        --token "${HF_TOKEN:-}" \
        --local-dir-use-symlinks False
else
    # Fallback: Use Python to download via transformers
    echo "huggingface-cli not found, using transformers library..."
    python <<EOF
from transformers import AutoTokenizer, AutoModelForCausalLM
import os

model_name = "${MODEL_NAME}"
cache_dir = "${HF_HOME}"

print(f"Downloading tokenizer for {model_name}...")
tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    cache_dir=cache_dir,
    token=os.getenv("HF_TOKEN")
)

print(f"Downloading model config and files for {model_name}...")
# Just download config and tokenizer files - vLLM will handle model weights
# This ensures all metadata is cached
try:
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        cache_dir=cache_dir,
        token=os.getenv("HF_TOKEN"),
        torch_dtype="auto"
    )
except Exception as e:
    print(f"Note: Full model download had issue: {e}")
    print("This is okay - vLLM will download weights as needed")

print(f"Model files downloaded successfully to {cache_dir}")
EOF
fi

echo "Model download complete!"