app.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer, GenerationConfig, BitsAndBytesConfig
import gradio as gr

# Authenticate using token from environment
hf_token = os.getenv("HF_TOKEN")
login(token=hf_token)

# Use quantization for low-memory GPU inference
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4"
)

model_name = "mistralai/Mistral-7B-Instruct-v0.3"

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=quantization_config,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Define generation function
def generate_qa(text):
    prompt = f"""### Instruction:
Based on the following SAP Note, generate exactly 20 unique and informative question-answer pairs.
Each question must refer to the SAP note number from text if additional context is needed.
Only output the pairs in the format:
Q1: ...
A1: ...
...
Q20: ...
A20: ...

### Input:
{text}

### Response:
"""
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        input_ids=inputs.input_ids,
        attention_mask=inputs.attention_mask,
        max_new_tokens=2500,
        do_sample=True,
        temperature=0.9,
        top_p=0.95,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.eos_token_id
    )

    output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    qa_pairs = output_text.split("### Response:")[-1].strip()
    return qa_pairs

# Define Gradio UI
demo = gr.Interface(
    fn=generate_qa,
    inputs=gr.Textbox(lines=20, label="SAP Note Text"),
    outputs=gr.Textbox(lines=25, label="Generated Q&A Pairs"),
    title="Mistral Q&A Generator for SAP Notes",
    description="Upload or paste SAP Note content to generate 20 question-answer pairs."
)

demo.launch()