app_lora.py

import spaces
import gradio as gr
import torch
import time
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# ============================================================
# CONFIG
# ============================================================

MODEL_NAME = "rikunarita/Qwen3-4B-Thinking-2507-Genius-Coder"
LORA_MODEL_NAME = "rahul7star/Qwen3-4B-Thinking-2509-Genius-Coder-AI"

MAX_INPUT_TOKENS = 4096
MAX_NEW_TOKENS = 4096

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# ============================================================
# LOAD TOKENIZER
# ============================================================

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

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

print("Loading base model...")
base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype="auto",
    device_map="auto",
)
base_model.eval()

print("Base model loaded successfully")

# ============================================================
# OPTIONAL LORA LOAD
# ============================================================

model = base_model
lora_loaded = False

try:
    print("Attempting to load LoRA adapter...")
    model = PeftModel.from_pretrained(
        base_model,
        LORA_MODEL_NAME,
        torch_dtype="auto",
    )
    model.eval()
    lora_loaded = True
    print("✅ LoRA loaded successfully")
except Exception as e:
    print("⚠️ LoRA not loaded:", e)
    model = base_model
    lora_loaded = False

# ============================================================
# SYSTEM PROMPT
# ============================================================

SYSTEM_PROMPT = """You are a professional AI Coding Assistant.
Your responses must be:
- Clear and concise
- Well-structured with headings and bullet points
- Technically accurate
- Written in a formal, professional tone
- Focused on best practices and production-quality code
"""

# ============================================================
# GENERATION FUNCTION
# ============================================================

@spaces.GPU()
def generate_answer(question, max_tokens, use_lora):

    print("\n================ GENERATE ANSWER START ================")

    if not question or not question.strip():
        return "Please enter a valid question."

    try:
        start_time = time.time()

        active_model = model if (use_lora and lora_loaded) else base_model

        messages = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": question.strip()},
        ]

        prompt = tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True,
        )

        inputs = tokenizer(
            prompt,
            return_tensors="pt",
            truncation=True,
            max_length=MAX_INPUT_TOKENS,
        ).to(DEVICE)

        input_token_count = inputs.input_ids.shape[-1]
        print(f"Input tokens: {input_token_count}")

        max_tokens = min(int(max_tokens), MAX_NEW_TOKENS)
        print(f"Final max_new_tokens: {max_tokens}")

        print("🚀 Starting generation...")

        with torch.no_grad():
            output = active_model.generate(
                **inputs,
                max_new_tokens=max_tokens,
                do_sample=False,
                repetition_penalty=1.05,
                use_cache=True,
                pad_token_id=tokenizer.eos_token_id,
                eos_token_id=tokenizer.eos_token_id,
            )

        print("✅ Generation finished")

        generated_tokens = output[0][input_token_count:]

        response = tokenizer.decode(
            generated_tokens,
            skip_special_tokens=True,
        )
        print(response)

        print(f"Generated tokens: {generated_tokens.shape[-1]}")
        print(f"⏱ Total time: {time.time() - start_time:.2f} sec")
        print("================ GENERATE ANSWER END ==================\n")

        return response.strip() or "No output generated."

    except Exception as e:
        import traceback
        traceback.print_exc()

        if torch.cuda.is_available():
            torch.cuda.empty_cache()

        return f"Error occurred: {str(e)}"


# ============================================================
# UI
# ============================================================

with gr.Blocks() as demo:

    gr.Markdown(
        """
        # 🤖 Professional Coding Assistant  
        **Qwen3-4B + Optional LoRA**

        - ⚡ Stable GPU inference  
        - 🧠 Deterministic responses  
        - 💻 Production-quality code  
        """
    )

    question = gr.Textbox(
        label="Your Question",
        placeholder="Explain Quick Sort with complexity and a Python example",
        value="write a python code using pytorch for a simple neural network demo",
        lines=4,
    )

    answer = gr.Markdown(label="AI Response", elem_id="answer_box")

    max_tokens = gr.Slider(
        64, 4096, value=1024, step=32, label="Max New Tokens"
    )

    use_lora = gr.Checkbox(
        value=lora_loaded,
        label="Enable LoRA Adapter"
    )

    with gr.Row():
        submit = gr.Button("Generate Answer", variant="primary")
        copy_btn = gr.Button("📋 Copy Response")
        clear = gr.Button("Clear")

    submit.click(
        fn=generate_answer,
        inputs=[question, max_tokens, use_lora],
        outputs=answer,
    )

    clear.click(
        fn=lambda: ("", ""),
        outputs=[question, answer],
    )

    copy_btn.click(
        fn=None,
        js="""
        () => {
            const el = document.querySelector('#answer_box');
            navigator.clipboard.writeText(el.innerText);
        }
        """,
    )

demo.launch(
    theme=gr.themes.Soft(),
    css="""
    .gradio-container { max-width: 900px !important; margin: auto; }
    textarea { font-size: 14px !important; }
    """,
)