app.py

import streamlit as st
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig

# Title for UI
st.title("💬 DeepSeek Math Assistant")

# Model and Tokenizer Setup
model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
device = "cuda" if torch.cuda.is_available() else "cpu"

@st.cache_resource
def load_model_and_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
        device_map="auto",
        trust_remote_code=True
    )
    return tokenizer, model

tokenizer, model = load_model_and_tokenizer()

# User Input
user_input = st.text_area("Ask a math question:", height=100, placeholder="e.g. What's the integral of x^2?")
generate_button = st.button("Generate Answer")

# Generation Settings
gen_config = GenerationConfig(
    max_length=256,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.2,
    num_beams=5,
    do_sample=True
)

if generate_button and user_input.strip() != "":
    # Add a system prompt to specialize the model
    prompt = f"You are a specialized math assistant. Only answer math-related questions.\nUser: {user_input}\nAssistant:"

    # Tokenize
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    input_ids = inputs["input_ids"]
    attention_mask = inputs.get("attention_mask", None)
    
    output = model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        generation_config=gen_config
    )

    
    # Decode
    response = tokenizer.decode(output[0], skip_special_tokens=True)

    # Extract only the assistant's answer
    if "Assistant:" in response:
        response = response.split("Assistant:")[-1].strip()

    # Show response
    st.markdown(f"**🧠 Answer:** {response}")