app.py

import os
import gc
import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import AutoModelForCausalLMWithValueHead
from safetensors.torch import load_file
import logging
from huggingface_hub import login

# Set up logging
login(token=os.environ.get("LA_NAME"))

# Constants
THRESHOLD = 2  # From Plan2Align

# Initialize device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Load models once
print("Loading models...")
model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="auto",
    torch_dtype=torch.float16
)

class RewardModel:
    def __init__(self, device, tokenizer, torch_dtype=torch.float16):
        self.device = device
        self.tokenizer = tokenizer
        
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
            
        # Set chat template if not already set
        if not hasattr(self.tokenizer, 'chat_template') or self.tokenizer.chat_template is None:
            # Using Llama 3's default chat template
            self.tokenizer.chat_template = "<|begin_of_text|>{% for message in messages %}{{'<|start_header_id|>' + message['role'] + '<|end_header_id|>\n' + message['content'] + '<|eot_id|>'}}{% endfor %}"
        
        print("Loading reward model...")
        self.RM = AutoModelForCausalLMWithValueHead.from_pretrained(
            "ray24724919/plan2align_rm",
            device_map={"": 0},  # Force model to stay on GPU
            torch_dtype=torch_dtype
        )
        self.RM.eval()
        print("Reward model loaded successfully!")
    
    def _create_single_message(self, language, source, translation):
        return [
            {
                "role": "system",
                "content": "You are a helpful translator and only output the result."
            },
            {
                "role": "user",
                "content": f"### Translate this from Chinese to {language}, Chinese:\n{source}\n### {language}:"
            },
            {
                "role": "assistant",
                "content": translation
            }
        ]
    
    def _process_inputs(self, messages):
        try:
            input_ids = self.tokenizer.apply_chat_template(
                messages,
                add_generation_prompt=False,
                return_tensors="pt",
                padding=True,
                truncation=True
            )
            
            attention_mask = torch.ones_like(input_ids)
            
            input_ids = input_ids.to(self.device)
            attention_mask = attention_mask.to(self.device)
            
            if len(input_ids.shape) == 1:
                input_ids = input_ids.unsqueeze(0)
                attention_mask = attention_mask.unsqueeze(0)
            
            return {
                "input_ids": input_ids,
                "attention_mask": attention_mask
            }
            
        except Exception as e:
            logging.error(f"Error processing inputs: {str(e)}")
            raise
    
    def reward_fn(self, language, source, translations):
        try:
            all_rewards = []
            for translation in translations:
                messages = self._create_single_message(language, source, translation)
                inputs = self._process_inputs(messages)
                with torch.no_grad():
                    outputs = self.RM(**inputs, return_value=True)
                    rewards = outputs[2]
                reward = rewards[0, -1].cpu().item()
                all_rewards.append(reward)
            return all_rewards
        except Exception as e:
            logging.error(f"Error in reward_fn: {str(e)}")
            raise

    def get_len(self, language, translations):
        try:
            len_ = 0
            for translation in translations:
                l = self.tokenizer(translation, return_tensors="pt").input_ids.to(device).shape[-1]
                len_ += l
            return len_
        except Exception as e:
            logging.error(f"Error in get_len: {str(e)}")
            raise

# Create reward model instance with the already loaded tokenizer
reward_model = RewardModel(device, tokenizer, torch_dtype=torch.float16)
print("Models loaded successfully!")

# Memory management function
def clear_cache():
    """Clear CUDA cache and run garbage collection to free memory"""
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    gc.collect()
    return "Cache cleared"

# Helper functions from Plan2Align
def rm_predict_preference(source, translation0, translation1, language="English"):
    translations = [translation0, translation1]
    for t_i in range(len(translations)):
        translations[t_i] = ''.join(translations[t_i]).replace('</s>',' ')
    rewards = reward_model.reward_fn(language, source.replace('</s>',' '), translations)
    best_index = rewards.index(max(rewards))
    return best_index

def rm_find_best_translation(source, translations, language="English"):
    copy_translations = translations.copy()
    
    if len(translations) < 2:
        return translations[0] if translations else None
    
    for t_i in range(len(translations)):
        translations[t_i] = ''.join(translations[t_i]).replace('</s>',' ')
    
    rewards = reward_model.reward_fn(language, ''.join(source).replace('</s>',' '), translations)
    
    print(rewards)
    
    best_index = rewards.index(max(rewards))

    print(f"Total translations length = {len(translations)}, and best translation index is: {best_index}")

    if rewards[best_index] >= THRESHOLD:
        return copy_translations[best_index]
    else:
        return None

def translate_chinese_to_english(chinese_text, target_language="English"):
    # Generate multiple translations
    translations = []
    
    # Generate three different translations with different system prompts
    system_prompts = [
        "You are a meticulous translator. Provide a literal, word-for-word translation that preserves the structure and meaning of each individual word.",
        "You are a professional translator. Deliver a clear, formal, and precise translation that faithfully conveys the original meaning.",
        "You are a creative and expressive translator. Render the text in a vivid and imaginative way, as if narrating a captivating story."
    ]
    
    for prompt in system_prompts:
        messages = [
            {"role": "system", "content": prompt},
            {"role": "user", "content": f"Translate the following Chinese text to {target_language}:\n\n{chinese_text}"}
        ]
        
        inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device)
        
        outputs = model.generate(
            inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.9,
            do_sample=True
        )
        
        translation = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
        translations.append(translation)
    
    # Get rewards for all translations
    rewards = reward_model.reward_fn(target_language, chinese_text.replace('</s>',' '), 
                                    [t.replace('</s>',' ') for t in translations])
    
    # Find the best translation
    best_index = rewards.index(max(rewards))
    best_translation = translations[best_index]
    
    # Return all information
    return {
        "best_translation": best_translation,
        "best_reward": rewards[best_index],
        "all_translations": translations,
        "all_rewards": rewards,
        "best_index": best_index
    }

# Updated process_text function with cache clearing
def process_text(text, target_language="English"):
    if not text.strip():
        return "Please enter some text to translate.", "", "", "", ""
    
    try:
        result = translate_chinese_to_english(text, target_language)
        
        # Format the candidate translations with their rewards
        candidates = []
        for i, (trans, reward) in enumerate(zip(result["all_translations"], result["all_rewards"])):
            marker = "★ " if i == result["best_index"] else ""
            candidates.append(f"{marker}Candidate {i+1} (Reward: {reward:.4f}):\n{trans}\n")
        
        candidates_text = "\n".join(candidates)
        
        # Clear cache after processing
        clear_cache()
        
        return (
            result["best_translation"],
            f"{result['best_reward']:.4f}",
            candidates_text,
            f"Candidate {result['best_index']+1}",
            "Yes" if result["best_reward"] >= THRESHOLD else "No"
        )
    except Exception as e:
        # Clear cache even if there's an error
        clear_cache()
        return f"Error: {str(e)}", "", "", "", ""

# Define available target languages - only the supported ones
target_languages = [
    "English", "Russian", "German", "Japanese", "Korean"
]

# Create an enhanced Gradio interface
with gr.Blocks(title="Test-Time Machine Translation with Plan2Align")
    gr.Markdown("This demo uses the Plan2Align approach to translate Chinese text to your chosen language, showing how the reward model evaluates different translation candidates. Paper: https://arxiv.org/pdf/2502.20795.")
    
    with gr.Row():
        with gr.Column(scale=1):
            source_text = gr.Textbox(
                label="Chinese Text",
                placeholder="Enter Chinese text here...",
                lines=5
            )
            target_language = gr.Dropdown(
                choices=target_languages,
                value="English",
                label="Target Language"
            )
            translate_button = gr.Button("Translate")
            clear_button = gr.Button("Clear Memory Cache")
        
        with gr.Column(scale=2):
            with gr.Tab("Best Translation"):
                best_translation = gr.Textbox(
                    label="Best Translation",
                    lines=5,
                    interactive=False
                )
                best_reward = gr.Textbox(
                    label="Reward Score",
                    interactive=False
                )
                best_candidate = gr.Textbox(
                    label="Best Candidate",
                    interactive=False
                )
                meets_threshold = gr.Textbox(
                    label="Meets Quality Threshold",
                    interactive=False
                )
            
            with gr.Tab("All Candidates"):
                all_candidates = gr.Textbox(
                    label="All Translation Candidates with Rewards",
                    lines=15,
                    interactive=False
                )
            
            cache_status = gr.Textbox(
                label="Cache Status",
                value="Ready",
                interactive=False
            )
    
    # Set up the translation flow
    translate_button.click(
        fn=process_text,
        inputs=[source_text, target_language],
        outputs=[best_translation, best_reward, all_candidates, best_candidate, meets_threshold]
    )
    
    # Add manual cache clearing button
    clear_button.click(
        fn=clear_cache,
        inputs=[],
        outputs=[cache_status]
    )
    
    # Examples with more complex sentences in Traditional Chinese about Taiwan
    gr.Examples(
        examples=[
            ["夜市文化豐富多彩，從士林夜市到饒河街夜市，提供各種美食、遊戲和購物體驗，吸引了無數遊客。", "English"],
            ["台北101曾經是世界最高的建築物，它不僅是台灣的地標，也象徵著經濟成就和創新精神。", "Russian"],
            ["阿里山日出和森林鐵路是台灣最著名的自然景觀之一，每年吸引數十萬遊客前來欣賞雲海和壯麗的日出。", "German"],
            ["珍珠奶茶起源於台灣，現已成為全球流行的飲品，展現了飲食文化對世界的影響力。", "Japanese"],
            ["原住民文化擁有豐富的傳統和藝術表現形式，包括歌舞、編織和木雕，反映了與自然和諧共處的生活智慧。", "Korean"]
        ],
        inputs=[source_text, target_language],
        outputs=[best_translation, best_reward, all_candidates, best_candidate, meets_threshold],
        fn=process_text
    )
    
    gr.Markdown("## How It Works")
    gr.Markdown("""
    1. The system generates three different translations using different translation styles:
       - Literal: A word-for-word translation preserving structure
       - Professional: A clear, formal translation
       - Creative: A vivid, expressive translation
    
    2. The reward model evaluates each translation and assigns a score
    
    3. The translation with the highest reward score is selected as the best
    
    4. A translation meets the quality threshold if its reward score is ≥ 2.0
    """)

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