app.py

# app.py for your Gradio Space

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

# --- 1. Configuration ---
# This is the ID of your LoRA adapter repository on the Hub
hub_adapter_id = "Prashasst/Sushruta-P3.8Q" # Make sure this is correct!

# Define the quantization configuration for efficient inference
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

# Your system prompt from training
SYSTEM_PROMPT = """You are Sushruta-P3.8Q, a specialized medical AI assistant from Prashasst's AI Labs. You were created and fine-tuned by Prashasst Dongre to serve as a reliable and accessible educational tool for the public. Your primary purpose is to help any user understand complex medical topics by providing clear, logical, step-by-step analyses in a <think>...</think><solution>...</solution> format."""

# --- 2. Load the Model and Tokenizer ---
print("Loading base model and tokenizer...")
# This single command downloads the base Phi-3 model, quantizes it,
# and applies your LoRA adapters from the Hub automatically.
model = AutoModelForCausalLM.from_pretrained(
    hub_adapter_id,
    quantization_config=quantization_config,
    device_map="auto",
    trust_remote_code=True,
    low_cpu_mem_usage=True
)
tokenizer = AutoTokenizer.from_pretrained(hub_adapter_id)
print("Model loaded successfully!")

# --- 3. Define the Prediction Function ---
def generate_response(user_question):
    """
    This function takes a user's question, formats the prompt,
    runs it through the model, and returns the clean response.
    """
    # Format the prompt using the official Phi-3 chat template
    prompt = f"<|system|>\n{SYSTEM_PROMPT}<|end|>\n<|user|>\n{user_question}<|end|>\n<|assistant|>\n"

    # Tokenize the input
    inputs = tokenizer(prompt, return_tensors="pt", return_attention_mask=False).to("cuda")

    # Generate the response
    outputs = model.generate(**inputs, max_new_tokens=1024)
    
    # Decode and clean up the response
    response_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Return only the assistant's part of the response
    return response_text.split("<|assistant|>")[1]

# --- 4. Create the Gradio Interface ---
with gr.Blocks(theme='soft') as demo:
    gr.Markdown("# Sushruta-P3.8Q: Your Medical AI Assistant")
    gr.Markdown("Created by Prashasst Dongre.")
    
    with gr.Row():
        question_box = gr.Textbox(label="Enter your medical question here")
        submit_button = gr.Button("Ask Sushruta")
    
    answer_box = gr.Markdown(label="Sushruta's Analysis:")

    submit_button.click(generate_response, inputs=question_box, outputs=answer_box)

# --- 5. Make the Interface Publicly Accessible ---
# This makes the Gradio app accessible as a web page and an API
demo.queue().launch(share=True)