app.py

from transformers import pipeline, AutoModelForSequenceClassification, AutoTokenizer
import gradio as gr
import torch
from concurrent.futures import ThreadPoolExecutor
from threading import Lock

# Global cache and thread lock for thread-safe caching
CACHE_SIZE = 100
prediction_cache = {}
cache_lock = Lock()

def load_model(model_name):
    """
    Loads the model with 8-bit quantization if a GPU is available.
    On CPU, it loads the full model.
    """
    if torch.cuda.is_available():
        # Use 8-bit quantization and auto device mapping for GPU inference.
        model = AutoModelForSequenceClassification.from_pretrained(
            model_name, load_in_8bit=True, device_map="auto"
        )
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        device = 0  # GPU index
    else:
        # CPU fallback: do not use quantization.
        model = AutoModelForSequenceClassification.from_pretrained(model_name)
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        device = -1

    return pipeline("text-classification", model=model, tokenizer=tokenizer, device=device)

# Load both models concurrently atartup.

with ThreadPoolExecutor() as executor:
    sentiment_future = executor.submit(load_model, "cardiffnlp/twitter-roberta-base-sentiment")
    emotion_future = executor.submit(load_model, "bhadresh-savani/bert-base-uncased-emotion")

sentiment_pipeline = sentiment_future.result()
emotion_pipeline = emotion_future.result()

def analyze_text(text):
    # Check cache first (thread-safe)
    with cache_lock:
        if text in prediction_cache:
            return prediction_cache[text]
    
    try:
        # Run both model inferences in parallel.
        with ThreadPoolExecutor() as executor:
            future_sentiment = executor.submit(sentiment_pipeline, text)
            future_emotion = executor.submit(emotion_pipeline, text)
            sentiment_result = future_sentiment.result()[0]
            emotion_result = future_emotion.result()[0]
        
        # Format the output with rounded scores.
        result = {
            "Sentiment": {sentiment_result['label']: round(sentiment_result['score'], 4)},
            "Emotion": {emotion_result['label']: round(emotion_result['score'], 4)}
        }
    except Exception as e:
        result = {"error": str(e)}
    
    # Update cache with protection.
    with cache_lock:
        if len(prediction_cache) >= CACHE_SIZE:
            prediction_cache.pop(next(iter(prediction_cache)))
        prediction_cache[text] = result
    
    return result

# Define the Gradio interface.
demo = gr.Interface(
    fn=analyze_text,
    inputs=gr.Textbox(placeholder="Enter your text here...", label="Input Text"),
    outputs=gr.JSON(label="Analysis Results"),
    title="🚀 Fast Sentiment & Emotion Analysis",
    description="An optimized application using quantized models (when available) and parallel processing for fast inference.",
    examples=[
        ["I'm thrilled to start this new adventure!"],
        ["This situation is making me really frustrated."],
        ["I feel so heartbroken and lost."]
    ],
    theme="soft",
    allow_flagging="never"
)

# Warm up the models to reduce first-call latency.
_ = analyze_text("Warming up models...")

if __name__ == "__main__":
    # In Spaces, binding to 0.0.0.0 is required.
    demo.launch(server_name="0.0.0.0", server_port=7860)