app.py

import gradio as gr
import torch
import yt_dlp
import os
import subprocess
import json
import time
import langdetect
import uuid
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

# Load Hugging Face Token
HF_TOKEN = os.getenv("HF_TOKEN")

print("Starting the program...")
model_path = "Qwen/Qwen2.5-7B-Instruct"

# **Efficient Model Loading**
bnb_config = BitsAndBytesConfig(load_in_8bit=True)  # Use 8-bit precision to reduce memory usage

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

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path, 
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    quantization_config=bnb_config,  # Load in 8-bit to save memory
    trust_remote_code=True
).to(device).eval()
print("Model successfully loaded.")

def generate_unique_filename(extension):
    return f"{uuid.uuid4()}{extension}"

def cleanup_files(*files):
    for file in files:
        if file and os.path.exists(file):
            os.remove(file)
            print(f"Removed file: {file}")

def download_youtube_audio(url):
    """Downloads audio from a YouTube video and converts it to WAV format."""
    print(f"Downloading audio from YouTube: {url}")
    output_path = generate_unique_filename(".wav")

    ydl_opts = {
        'format': 'bestaudio/best',
        'postprocessors': [{
            'key': 'FFmpegExtractAudio',
            'preferredcodec': 'wav',
            'preferredquality': '192',
        }],
        'outtmpl': output_path,
    }

    try:
        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
            ydl.download([url])
        if os.path.exists(output_path + ".wav"):
            os.rename(output_path + ".wav", output_path)  # Ensure correct naming
    except Exception as e:
        return f"Error downloading audio: {str(e)}"
    
    return output_path if os.path.exists(output_path) else "Download Failed"

def transcribe_audio(file_path):
    """Transcribes audio using `insanely-fast-whisper` and handles large files efficiently."""
    print(f"Starting transcription of file: {file_path}")
    temp_audio = None

    if file_path.endswith(('.mp4', '.avi', '.mov', '.flv')):
        print("Video file detected. Extracting audio using ffmpeg...")
        temp_audio = generate_unique_filename(".wav")
        command = ["ffmpeg", "-i", file_path, "-q:a", "0", "-map", "a", temp_audio]
        subprocess.run(command, check=True)
        file_path = temp_audio  # Use extracted audio file

    output_file = generate_unique_filename(".json")
    command = [
        "insanely-fast-whisper",
        "--file-name", file_path,
        "--device-id", "0",
        "--model-name", "openai/whisper-large-v3",
        "--task", "transcribe",
        "--timestamp", "chunk",
        "--transcript-path", output_file
    ]

    try:
        subprocess.run(command, check=True)
    except Exception as e:
        return f"Error in transcription: {str(e)}"
    
    # Process the JSON file in chunks to avoid memory overflow
    result = []
    try:
        with open(output_file, "r") as f:
            for line in f:
                chunk = json.loads(line.strip())  # Read JSON line by line
                result.append(chunk.get("text", ""))
    except Exception as e:
        return f"Error reading transcription file: {str(e)}"

    cleanup_files(output_file)
    if temp_audio:
        cleanup_files(temp_audio)
    
    return " ".join(result)[:500000]  # Limit transcription size

def generate_summary_stream(transcription):
    """Summarizes the transcription efficiently to avoid memory overflow."""
    detected_language = langdetect.detect(transcription[:1000])  # Detect using a smaller portion

    # Use smaller chunks for processing
    chunk_size = 2000
    transcript_chunks = [transcription[i:i+chunk_size] for i in range(0, len(transcription), chunk_size)]
    summary_result = []

    for chunk in transcript_chunks[:3]:  # Process only the first 3 chunks to avoid OOM
        prompt = f"""Summarize the following video transcription in 150-300 words in {detected_language}:\n{chunk}"""
        input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
        output_ids = model.generate(input_ids, max_length=300)  # Limit output size
        response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
        summary_result.append(response)

    return "\n\n".join(summary_result)

def process_youtube(url):
    """Handles YouTube video processing: downloads audio, transcribes it, and cleans up."""
    if not url:
        return "Please enter a YouTube URL.", None

    audio_file = download_youtube_audio(url)
    if "Error" in audio_file or audio_file == "Download Failed":
        return audio_file, None

    transcription = transcribe_audio(audio_file)
    cleanup_files(audio_file)  # Clean up the downloaded file
    return transcription, None

def process_uploaded_video(video_path):
    """Processes uploaded video file for transcription."""
    transcription = transcribe_audio(video_path)
    return transcription, None

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🎥 Video Transcription and Smart Summary
    Upload a video or provide a YouTube link to get a transcription and AI-generated summary.
    """)

    with gr.Tabs():
        with gr.TabItem("📤 Video Upload"):
            video_input = gr.Video()
            video_button = gr.Button("🚀 Process Video")
        
        with gr.TabItem("🔗 YouTube Link"):
            url_input = gr.Textbox(placeholder="https://www.youtube.com/watch?v=...")
            url_button = gr.Button("🚀 Process URL")

    transcription_output = gr.Textbox(label="📝 Transcription", lines=10, show_copy_button=True)
    summary_output = gr.Textbox(label="📊 Summary", lines=10, show_copy_button=True)
    summary_button = gr.Button("📝 Generate Summary")

    video_button.click(process_uploaded_video, inputs=[video_input], outputs=[transcription_output, summary_output])
    url_button.click(process_youtube, inputs=[url_input], outputs=[transcription_output, summary_output])
    summary_button.click(generate_summary_stream, inputs=[transcription_output], outputs=[summary_output])

demo.launch()