# cache_download.py - Run ONCE
from transformers import AutoModelForCausalLM, AutoTokenizer
import os

MODEL_NAME = "microsoft/Phi-3-mini-4k-instruct"
MODEL_NAME = "Qwen/Qwen2-1.5B-Instruct"  # ✅ No tokenizer bugs


print("Caching Phi-3...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, cache_dir="./hf_cache")
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME, 
    cache_dir="./hf_cache",
    torch_dtype="auto",
    device_map="cpu"
)
print("✅ Cached to ./hf_cache/")