import gradio as gr
from llama_cpp import Llama
from huggingface_hub import hf_hub_download

# --- THIS IS THE DOWNLOAD STEP ---
# The Space will download the model automatically when it starts.
# We use Bartowski's optimized version which is small and fast.
repo_id = "bartowski/Qwen2.5-VL-7B-Instruct-GGUF"
filename = "Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf"

print(f"⬇️ Downloading {filename} from Hugging Face Cloud...")
model_path = hf_hub_download(
    repo_id=repo_id,
    filename=filename, 
    repo_type="model"
)
print("✅ Download Complete! Loading model...")

# --- LOAD THE MODEL ---
# n_gpu_layers=0 means we force it to run on the CPU (Free Tier)
llm = Llama(
    model_path=model_path,
    n_ctx=4096,       # How much it remembers
    n_gpu_layers=0,   # 0 = CPU
    verbose=True
)

# --- CHAT INTERFACE ---
def chat(message, history):
    # Qwen-specific format
    prompt = f"<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n"
    
    output = llm(
        prompt, 
        max_tokens=512, 
        stop=["<|im_end|>"], 
        echo=False
    )
    return output['choices'][0]['text']

gr.ChatInterface(chat).launch(server_name="0.0.0.0", server_port=7860)