handler.py

from typing import Dict, List, Any
import base64
import tempfile
import os
import subprocess
import sys

class EndpointHandler:
    def __init__(self, path=""):
        print("Starting GPU Cloud Environment... (The Monkey Patch Fix)")
        
        try:
            # 1. Install Linux audio drivers just to be 100% safe
            print("Installing OS audio drivers...")
            subprocess.check_call("apt-get update && apt-get install -y libsndfile1 ffmpeg", shell=True)
            
            # 2. THE ULTIMATE FIX: Force PIP to keep the pre-installed PyTorch 2.5.1!
            print("Installing TTS while locking PyTorch versions...")
            subprocess.check_call([
                sys.executable, "-m", "pip", "install", 
                "TTS==0.22.0", "torch==2.5.1", "torchaudio==2.5.1", 
                "--ignore-installed", "blinker", "--no-cache-dir"
            ])
            
            # 3. Apply the Scipy compiler patch
            print("Forcing clean Scipy binary...")
            subprocess.check_call([
                sys.executable, "-m", "pip", "install", "--force-reinstall", 
                "--only-binary", "scipy", "scipy==1.13.0", "--no-cache-dir"
            ])
            print("Dependencies safely locked and installed!")
        except Exception as e:
            print(f"Installation crashed: {e}")
            
        # NOW we safely import the libraries AFTER they are installed
        import torch
        from TTS.api import TTS
        
        # --- THE COQUI BUG FIX (MONKEY PATCH) ---
        # Because these are read-only properties, we have to hack the TTS class itself BEFORE we use it!
        TTS.is_multi_lingual = property(lambda self: False)
        TTS.is_multi_speaker = property(lambda self: False)
        # ----------------------------------------
        
        print("Loading Ormuri Voice into the Cloud GPU...")
        
        model_file = os.path.join(path, "best_model.pth")
        config_file = os.path.join(path, "config.json")
        
        # Load the Coqui TTS model onto the Nvidia GPU
        self.tts = TTS(model_path=model_file, config_path=config_file, progress_bar=False, gpu=True)
        
        print("GPU TTS Engine Ready!")

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        text = data.pop("inputs", "")
        
        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio:
            temp_path = temp_audio.name

        self.tts.tts_to_file(text=text, file_path=temp_path)

        with open(temp_path, "rb") as f:
            audio_bytes = f.read()
        
        base64_audio = base64.b64encode(audio_bytes).decode("utf-8")
        os.remove(temp_path)

        return [{"audio": base64_audio, "sampling_rate": 22050}]