Why does C drive memory increase when processing many audio files on the server, but returns to normal after stopping the server?

[ninhle616]

I want to ask why, when I deploy to the server and process around 1,000 audio files, the C drive memory keeps increasing. However, when I stop the server, the C drive memory returns to normal.

#!/usr/bin/env python3
from flask import Flask, request, jsonify
import whisper
import os
import gc
import torch

app = Flask(__name__)

# Load model once at startup
print("Loading Whisper model...")
model = whisper.load_model("tiny")
print("Model loaded!")

@app.route('/transcribe', methods=['POST'])
def transcribe():
    """
    API endpoint để transcribe audio
    POST /transcribe
    Parameters:
    - audio_path: đường dẫn tới file audio
    - language: ngôn ngữ (mặc định: Vietnamese)
    """
    try:
        data = request.get_json()
        audio_path = data.get('audio_path')
        language = data.get('language', 'Vietnamese')

        # Kiểm tra audio_path được cung cấp
        if not audio_path:
            return jsonify({'error': 'audio_path is required'}), 400

        # Kiểm tra file tồn tại
        if not os.path.exists(audio_path):
            return jsonify({'error': f'File not found: {audio_path}'}), 400

        # Kiểm tra file size
        file_size = os.path.getsize(audio_path)
        if file_size == 0:
            return jsonify({'error': f'File is empty: {audio_path}'}), 400

        print(f"File size: {file_size} bytes")

        # Kiểm tra format
        supported_formats = ('.mp3', '.wav', '.m4a', '.flac', '.ogg', '.wma', '.aac')
        if not audio_path.lower().endswith(supported_formats):
            return jsonify({
                'error': f'Invalid audio format. Supported: {supported_formats}. Got: {audio_path}'
            }), 400

        print(f"Transcribing: {audio_path}")

        # Transcribe with fp16=False to reduce memory for tiny model
        result = model.transcribe(audio_path, language=language, fp16=False)

        response = {
            'status': 'success',
            'text': result['text']
        }

        # Explicit cleanup after transcription
        del result
        gc.collect()

        # If using GPU, clear CUDA cache
        if torch.cuda.is_available():
            torch.cuda.empty_cache()

        return jsonify(response)

    except Exception as e:
        error_msg = str(e)
        print(f"Error: {error_msg}")
        # Cleanup on error too
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        return jsonify({'error': error_msg}), 500


@app.route('/health', methods=['GET'])
def health():
    return jsonify({'status': 'ok'})


if __name__ == '__main__':
    # Run on port 5000
    # Consider reducing workers for memory-constrained servers
    app.run(host='0.0.0.0', port=5000, debug=False, threaded=True)

[ninhle616]

set none model.