import gradio as gr
from llama_cpp import Llama

# This pulls your specific Qwen 2.5 Coder GGUF
# 1.5B fits perfectly in the 16GB RAM of a free Space
llm = Llama.from_pretrained(
    repo_id="harishkumar12k/Qwen2.5-Coder-1.5B-Q4_K_M-GGUF",
    filename="qwen2.5-coder-1.5b-q4_k_m.gguf",
    n_ctx=2048,
    n_threads=2 
)

def chat(message, history):
    # Format for Qwen-Coder
    prompt = "system\nYou are a helpful coding assistant.\n"
    for user_msg, assistant_msg in history:
        prompt += f"user\n{user_msg}\nassistant\n{assistant_msg}\n"
    prompt += f"user\n{message}\nassistant\n"

    output = llm(
        prompt,
        max_tokens=512,
        stop=["user", "<|endoftext|>"],
        echo=False
    )
    
    return output["choices"][0]["text"].strip()

demo = gr.ChatInterface(
    fn=chat,
    title="Qwen 2.5 Coder 1.5B GGUF",
    description="Running on HF Spaces CPU via llama-cpp-python"
)

if __name__ == "__main__":
    demo.launch()