app.py

import gradio as gr
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer, pipeline

# Load ONNX model
def create_fast_summarizer():
    model = ORTModelForSeq2SeqLM.from_pretrained(
        "onnx-community/bart-large-cnn-ONNX",
        encoder_file_name="encoder_model_q4.onnx",
        decoder_file_name="decoder_model_q4.onnx",
        provider="CPUExecutionProvider",
        use_io_binding=True
    )
    tokenizer = AutoTokenizer.from_pretrained(
        "onnx-community/bart-large-cnn-ONNX",
        use_fast=True
    )
    return pipeline(
        "summarization",
        model=model,
        tokenizer=tokenizer,
        device=-1
    )

summarizer = create_fast_summarizer()

# Summarize function with prompt +tuned params
def summarize_text(text):
    prompt = "Summarize the key events, including casualties and political context:\n" + text
    result = summarizer(
        prompt,
        max_length=160,
        min_length=55,
        do_sample=False,
        num_beams=6,
        length_penalty=1.5,
        no_repeat_ngram_size=3,  # Prevent repetition
        clean_up_tokenization_spaces=True,
        early_stopping=True
    )
    return result[0]['summary_text']

# Build Gradio interface
app = gr.Interface(
    fn=summarize_text,
    inputs=gr.Textbox(lines=15, placeholder="Paste your text here..."),
    outputs="text",
    title="ONNX Summarizer 🚀",
    description="Paste any news or article text and get a concise, context-rich summary."
)

app.launch(mcp_server=True,share=True)