app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import gradio as gr
import re

# Check if GPU is available
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load the model and tokenizer with GPU optimizations
model = AutoModelForCausalLM.from_pretrained(
    "AlanYky/phi-3.5_tweets_instruct",
    torch_dtype=torch.float32,
    trust_remote_code=True,
)

tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-mini-instruct")

# Define the pipeline with device optimization
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device=0 if device == "cuda" else -1  # Set device for pipeline
)

# Define generation arguments
generation_args = {
    "max_new_tokens": 70,
    "return_full_text": False,
    "temperature": 0.7,
    "top_k": 20,
    "top_p": 0.8,
    "repetition_penalty": 1.2,
    "do_sample": True,
}


# Function for tweet generation
def generate_tweet(tweet_idea):
    max_length = 50  # Set a maximum input length
    if len(tweet_idea) > max_length:
        return f"Error: Tweet idea exceeds the maximum allowed length of {max_length} characters."

    prompt = f"Could you generate a tweet about {tweet_idea}?"
    messages = [{"role": "user", "content": prompt}]
    output = pipe(messages, **generation_args)
    return clean_tweet(output[0]['generated_text'])


def generate_tweet_with_rag(tweet_idea):
    prompt = f"Could you generate a tweet about {tweet_idea}?"
    messages = [{"role": "user", "content": prompt}]
    output = pipe(messages, **generation_args)
    return output[0]['generated_text']


def clean_tweet(tweet):
    # Remove trailing sequences of non-ASCII characters (like `����������������`)
    cleaned_tweet = tweet.replace("�", "")

    cleaned_tweet = re.sub(r"\(\)", "", cleaned_tweet)

    # change more than one dot to one dot
    cleaned_tweet = re.sub(r"\.{4,}", "...", cleaned_tweet)

    # change more than two question marks to two question marks
    cleaned_tweet = re.sub(r"\?{4,}", "???", cleaned_tweet)

    # change more than two ! to two !
    cleaned_tweet = re.sub(r"!{4,}", "!!!", cleaned_tweet)

    # Remove repeated double quotes at the end
    cleaned_tweet = re.sub(r'"+$', '', cleaned_tweet)

    return cleaned_tweet.strip()

custom_css = """
#header {
    text-align: center;
    margin-top: 20px;
}
#input-box, #output-box {
    margin: 0 auto;
    width: 80%;
}
#generate-button {
    margin: 10px auto;
    display: block;
}
"""


with gr.Blocks(css=custom_css) as demo:
    # Add a title with the X.com logo
    gr.Markdown(
        """
        <div id="header">
            <img src="https://upload.wikimedia.org/wikipedia/commons/thumb/e/e3/X.com_logo.svg/1024px-X.com_logo.svg.png" 
                 alt="X.com Logo" width="100">
            <h1 style="font-size: 2.5em; margin: 0;">Tweet Generator</h1>
            <p style="font-size: 1.2em; color: gray;">
                Powered by <b>AlanYky/phi-3.5_tweets_instruct</b>
            </p>
        </div>
        """
    )

    # Center the input and output components
    instruction_input = gr.Textbox(
        label="Tweet Idea",
        placeholder="Enter your tweet idea (It can be a topic, hashtag, sentence, or any format)...",
        lines=2,
        elem_id="input-box"
    )
    generate_button = gr.Button("Generate", elem_id="generate-button")
    output_box = gr.Textbox(
        label="Generated Tweet",
        placeholder="Your tweet will appear here.",
        lines=3,
        elem_id="output-box"
    )

    # Connect the button to the generate function
    generate_button.click(generate_tweet, inputs=instruction_input, outputs=output_box)

print("Model loaded on:", next(model.parameters()).device)

demo.launch()