app.py

import gradio as gr
import spaces
import torch

from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM
)

from peft import PeftModel


# ============================================================
# CONFIGURATION
# ============================================================

BASE_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"

ADAPTER_MODEL = "dd253B/DhanushAI-0.5B"


# ============================================================
# GLOBAL MODEL
# ============================================================

tokenizer = None
model = None


# ============================================================
# LOAD MODEL
# ============================================================

def load_model():

    global tokenizer
    global model

    if model is not None:
        return

    print("====================================")
    print("Loading DhanushAI...")
    print("====================================")

    # -----------------------------
    # Tokenizer
    # -----------------------------

    print("Loading tokenizer...")

    tokenizer = AutoTokenizer.from_pretrained(
        BASE_MODEL
    )

    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token


    # -----------------------------
    # Base model
    # -----------------------------

    print("Loading base model...")

    base_model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL,
        torch_dtype=torch.float16
    )


    # -----------------------------
    # LoRA adapter
    # -----------------------------

    print("Loading DhanushAI adapter...")

    model = PeftModel.from_pretrained(
        base_model,
        ADAPTER_MODEL
    )


    # -----------------------------
    # Move to GPU
    # -----------------------------

    model = model.to("cuda")

    model.eval()

    print("====================================")
    print("DhanushAI loaded successfully!")
    print("====================================")


# ============================================================
# CHAT FUNCTION
# ============================================================

@spaces.GPU
def chat(message):

    # Load model after ZeroGPU allocation
    load_model()

    if message is None:
        return "Please enter a message."

    message = message.strip()

    if not message:
        return "Please enter a message."


    # -----------------------------
    # Prompt
    # -----------------------------

    prompt = f"""You are DhanushAI, a helpful AI assistant.

User: {message}

Assistant:"""


    # -----------------------------
    # Tokenize
    # -----------------------------

    inputs = tokenizer(
        prompt,
        return_tensors="pt"
    )


    inputs = {
        key: value.to("cuda")
        for key, value in inputs.items()
    }


    # -----------------------------
    # Generate
    # -----------------------------

    with torch.no_grad():

        outputs = model.generate(

            **inputs,

            max_new_tokens=200,

            temperature=0.7,

            top_p=0.9,

            do_sample=True,

            repetition_penalty=1.1
        )


    # -----------------------------
    # Decode
    # -----------------------------

    generated = tokenizer.decode(
        outputs[0],
        skip_special_tokens=True
    )


    # -----------------------------
    # Remove prompt
    # -----------------------------

    if "Assistant:" in generated:

        answer = generated.split(
            "Assistant:",
            1
        )[1].strip()

    else:

        answer = generated.strip()


    return answer


# ============================================================
# GRADIO UI + API
# ============================================================

demo = gr.Interface(

    fn=chat,

    inputs=gr.Textbox(
        label="Message",
        placeholder="Ask DhanushAI something..."
    ),

    outputs=gr.Textbox(
        label="DhanushAI"
    ),

    title="DhanushAI",

    description="My custom AI model",

    api_name="chat"
)


# ============================================================
# START
# ============================================================

demo.launch()