app.py

import gradio as gr
from huggingface_hub import InferenceClient, login
import tensorflow as tf
from transformers import AutoTokenizer, TFAutoModelForCausalLM
import os

# Set up mixed precision and distribution strategy
policy = tf.keras.mixed_precision.Policy('mixed_bfloat16')
tf.keras.mixed_precision.set_global_policy(policy)
strategy = tf.distribute.MultiWorkerMirroredStrategy()

# Log into Hugging Face
login(os.environ.get("hf_token"))

# Load tokenizer and model
name = "WICKED4950/GPT2-InstEsther0.28eV3.1"
tokenizer = AutoTokenizer.from_pretrained(name)
tokenizer.pad_token = tokenizer.eos_token

with strategy.scope():
    model = TFAutoModelForCausalLM.from_pretrained(name)

# Raw Prediction Function
def raw_pred(input, model, tokenizer, max_length=1024, temperature=0.2):
    input_ids = tokenizer.encode(input, return_tensors='tf')

    # Initialize variables
    generated_ids = input_ids
    stop_token_id = tokenizer.encode("<|SOH|>", add_special_tokens=False)[0]
    all_generated_tokens = []  # To store generated token IDs
    tokens_yielded = []  # To store tokens as they are yielded

    with strategy.scope():
        for _ in range(max_length // 1):  # Generate in chunks of 3 tokens
            # Generate tokens
            outputs = model.generate(
                generated_ids,
                max_length=generated_ids.shape[1] + 1,
                temperature=temperature,
                pad_token_id=tokenizer.eos_token_id,
                eos_token_id=stop_token_id,
                do_sample=True,
                num_return_sequences=1
            )

            # Get the newly generated tokens
            new_tokens = outputs[0, -1:]
            generated_ids = outputs  # Update the generated_ids with the new tokens

            # Store and yield the generated tokens
            all_generated_tokens.extend(new_tokens.numpy().tolist())
            tokens_text = tokenizer.decode(new_tokens, skip_special_tokens=False)
            tokens_yielded.append(tokens_text)
            yield tokens_text

            # Stop if stop token is encountered
            if stop_token_id in new_tokens.numpy():
                final_text = tokenizer.decode(all_generated_tokens, skip_special_tokens=False)
                yield "<|Clean|>" + final_text
                break

# Response Handler Function
def respond(message, history):
    give_mod = ""
    history = history[-3:]  # Limit history to last 3 exchanges
    for chunk in history:
        give_mod += f"<|SOH|>{chunk[0]}<|SOB|>{chunk[1]}"
    give_mod += f"<|SOH|>{message.capitalize()}<|SOB|>"
    print(give_mod)

    response = ""
    for token in raw_pred(give_mod, model, tokenizer):
        if "<|Clean|>" in token:
            response = token
            print(response)
        else:
            response += token
        yield response.replace("<|SOH|>", "").replace("<|Clean|>", "")

# Gradio Chat Interface Setup
demo = gr.ChatInterface(
    fn=respond,
    title="Chat with Esther",  # Title of the app
    description="A friendly chatbot ready to help and chat with you! 😊 (Everything you say will be stored and will be used for further improving the model)",  # Description of the app
    theme="compact",  # Choose the theme
)

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