app.py

import os
os.system("pip install -U bitsandbytes==0.45.3 transformers accelerate torch --no-cache-dir")


import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
from functools import lru_cache

# Define models
BASE_MODEL = "deepseek-ai/deepseek-math-7b-rl"
FINETUNED_MODEL = "LaibaIrfan/emoji_math"

# Load tokenizer and model
@lru_cache()
def load_model():
    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)  # Use base model tokenizer
    base_model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL,
        torch_dtype=torch.float32,
        device_map="cpu"    )
    model = PeftModel.from_pretrained(base_model, FINETUNED_MODEL, device_map="cpu")
    return tokenizer, model

# Load the model
tokenizer, model = load_model()

# Function to generate the result
def generate_result(incorrect_math):
    input_text = f"Incorrect: {incorrect_math}\nCorrect:"

    # Move input to GPU
    inputs = tokenizer(input_text, return_tensors="pt").to("cuda")

    # Generate output on GPU
    output = model.generate(**inputs, max_length=200)

    return tokenizer.decode(output[0], skip_special_tokens=True)

# Gradio Interface
iface = gr.Interface(
    fn=generate_result,
    inputs="text",
    outputs="text",
    title="Emoji Math Solver 🧮",
    description="Enter an emoji-based math equation, and the model will generate the correct answer!"
)

iface.launch(debug=True, share=True, inline=True)

# Function to generate result
def generate_result(incorrect_math):
    input_text = f"Incorrect: {incorrect_math}\nCorrect:"
    inputs = tokenizer(input_text, return_tensors="pt").to("cuda")  # Use GPU if available
    output = model.generate(**inputs, max_length=200)
    return tokenizer.decode(output[0], skip_special_tokens=True)

# Gradio Interface
iface = gr.Interface(
    fn=generate_result,
    inputs="text",
    outputs="text",
    title="Emoji Math Solver 🧮",
    description="Enter an emoji-based math equation, and the model will generate the correct answer!"
)

iface.launch(share=True)