app.py

# from fastapi import FastAPI, Request
# from pydantic import BaseModel
# # from unsloth import FastLanguageModel
# import torch
# import re




# app = FastAPI()

# # Load model once on startup
# model, tokenizer = FastLanguageModel.from_pretrained(
#     model_name = "unsloth/mistral-7b-instruct-v0.3-bnb-4bit",
#     max_seq_length = 2048,
#     dtype = None,
#     load_in_4bit = True,
# )
# FastLanguageModel.for_inference(model)

# class SAPNoteRequest(BaseModel):
#     text: str

# @app.post("/generate_qa")
# def generate_qa(req: SAPNoteRequest):
#     text = req.text
#     match = re.search(r"SAP Note\s*(\d+)", text)
#     sap_note_number = match.group(1) if match else "UNKNOWN"

#     prompt = f"""
#     Generate 20 question-answer pairs based on the following SAP Note.
#     Each question should include the SAP note number {sap_note_number} to clarify context.

#     \"\"\"{text}\"\"\"

#     Q1: question
#     A1: answer

#     ### Response:
#     """
#     inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
#     outputs = model.generate(
#         inputs.input_ids,
#         max_new_tokens=2048,
#         do_sample=True,
#         temperature=0.7,
#         top_p=0.95,
#         repetition_penalty=1.2
#     )
#     output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
#     qa_pairs = output_text.split("### Response:")[-1].strip()
#     return {"qa_pairs": qa_pairs}


### Hugging face code

# import torch
# from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# # Quantization settings
# quantization_config = BitsAndBytesConfig(
#     load_in_4bit=True,
#     bnb_4bit_quant_type="nf4",
#     bnb_4bit_compute_dtype=torch.float16,
# )

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

# 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()