app.py

import gradio as gr
from transformers import pipeline, AutoTokenizer, BitsAndBytesConfig, AutoModelForCausalLM
import torch
import spaces

MODEL_PATH = "benhaotang/phi4-qwq-sky-t1"
MODEL_URL = f"https://huggingface.co/{MODEL_PATH}"

def load_model():
    bnb_config = BitsAndBytesConfig(
        load_in_8bit=False,
        llm_int8_enable_fp32_cpu_offload=True
    )
    
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_PATH,
        device_map="auto",
        torch_dtype=torch.float16,
        offload_folder="offload_folder",
        quantization_config=bnb_config
    )
    
    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
    
    pipe = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        device_map="auto",
    )
    
    return pipe

pipe = load_model()

@spaces.GPU(duration=110)
def generate_response(prompt, max_length=1024):
    # Create messages with system prompt
    messages = [
        {"role": "system", "content": "You are a helpful AI assistant. You always think step by step."},
        {"role": "user", "content": prompt}
    ]
    
    outputs = pipe(messages, max_new_tokens=max_length)

    # print("Raw output:", outputs) #removed after debugging output format problem is done
    # Extract just the assistant's response
    try:
        # outputs[0]["generated_text"] is already a list of messages
        message_list = outputs[0]["generated_text"]
        # Get the last message (assistant's response)
        assistant_message = message_list[-1]
        if assistant_message["role"] == "assistant":
            return assistant_message["content"]
    except Exception as e:
        # print(f"Error extracting response: {e}")
        # If extraction fails, return the raw output
        return str(outputs[0]["generated_text"])
        
    return outputs[0]["generated_text"]

# Example with proper line breaks
example_prompt = """For a scalar field theory with interaction Lagrangian $\mathcal{L}_{int} = g\phi^3 + \lambda\phi^4$:

1. Enumerate all possible 1-loop Feynman diagrams contributing to the scalar propagator

2. For each diagram, write down its loop contribution

3. Provide Mathematica code to calculate these loop amplitudes with dimensional regularization at $d=4-\epsilon$

Please explain your reasoning step by step."""

demo = gr.Interface(
    fn=generate_response,
    inputs=[
        gr.Textbox(
            label="Enter your question",
            placeholder="Ask me anything...",
            lines=5
        ),
    ],
    outputs=gr.Textbox(label="Response", lines=10),
    title="benhaotang/phi4-qwq-sky-t1",
    description=f""" To achieve CoT and science reasoning on small scale with a merge of CoT finetuned phi4 model.
    
Model: [benhaotang/phi4-qwq-sky-t1]({MODEL_URL})""",
    examples=[
        [example_prompt]  # Now using the formatted example
    ]
)

demo.launch()