app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch
import html  # เพิ่ม html library

# --- Configuration ---
BASE_MODEL_ID = "google/gemma-2-2b-it"
ADAPTER_ID = "Phonsiri/gemma-2-2b-it-grpo-v6-checkpoints"

# --- Load Tokenizer & Model ---
print(f"Loading base model: {BASE_MODEL_ID}...")
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    device_map="auto",
    torch_dtype=torch.float16
)

print(f"Loading adapter: {ADAPTER_ID}...")
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)

def generate(prompt):
    messages = [{"role": "user", "content": prompt}]
    
    formatted_prompt = tokenizer.apply_chat_template(
        messages, 
        tokenize=False,
        add_generation_prompt=True
    )
    
    inputs = tokenizer(formatted_prompt, return_tensors="pt").to(model.device)
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=2048,
            temperature=0.6,
            top_p=0.9,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id
        )
    
    full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Cleaning Logic
    if "model\n" in full_response:
        response = full_response.split("model\n")[-1].strip()
    elif "<start_of_turn>model" in full_response:
        response = full_response.split("<start_of_turn>model")[-1].strip()
    else:
        response = full_response[len(formatted_prompt):].strip()
        if len(response) == 0:
             response = full_response

    response = response.replace("<end_of_turn>", "").strip()
    
    # --- สำคัญ: แก้ไขการแสดงผล Tag ---
    # แปลง < เป็น &lt; เพื่อให้ Gradio ไม่มองว่าเป็น HTML tag ที่ต้องซ่อน
    # หรือใช้วิธีใส่ Code Block ครอบ
    return f"```xml\n{response}\n```"

# --- Gradio UI ---
examples = [
    ["A store sells notebooks for $3 each and pens for $1.50 each. Sarah buys 4 notebooks and 6 pens. How much does she pay in total?"],
    ["If John is taller than Mike, and Mike is taller than Sarah, who is the tallest?"],
    ["Solve for x: 2x + 5 = 15"]
]

# เปลี่ยน Output เป็น Markdown เพื่อให้ render code block สวยๆ
demo = gr.Interface(
    fn=generate,
    inputs=gr.Textbox(
        label="Question", 
        lines=3, 
        placeholder="Ask a math or reasoning question..."
    ),
    outputs=gr.Markdown(  # เปลี่ยนจาก Textbox เป็น Markdown
        label="Reasoning & Answer"
    ),
    title="Gemma-2-2B GRPO (Adapter Version)",
    description=f"Running Adapter: {ADAPTER_ID}\nBase Model: {BASE_MODEL_ID}",
    examples=examples,
    theme=gr.themes.Soft()
)

if __name__ == "__main__":
    demo.launch(share=True)