Keep API startup independent of model prewarming

app.py

@@ -1,5 +1,6 @@
 import os
 import platform
+import sys
 import shutil
 import tempfile
 import torch
@@ -7,6 +8,13 @@ import torchaudio
 import gradio as gr
 import spaces
 from fastapi import FastAPI
+
+# Hugging Face Spaces should use UTF-8; explicitly configure streams so model
+# libraries cannot inherit a platform-specific charmap encoding.
+for _stream in (sys.stdout, sys.stderr):
+    if hasattr(_stream, "reconfigure"):
+        _stream.reconfigure(encoding="utf-8", errors="backslashreplace")
+
 from fastapi.responses import JSONResponse
 from einops import rearrange
 from huggingface_hub import login
@@ -135,17 +143,8 @@ def load_model(model_name):
     return MODEL_CACHE[model_name]
 
 
-# Pre-warm both models in background so the first user request is fast
-# but the Gradio server starts immediately without waiting for downloads
-import threading
-
-def _prewarm():
-    _log("Pre-warming models...")
-    for _name in ["small-music", "small-sfx"]:
-        load_model(_name)
-    _log("All models ready!")
-
-threading.Thread(target=_prewarm, daemon=True).start()
+# Model loading is lazy: startup must remain healthy even when a model download
+# or initialization fails. The first generation request loads the selected model.
 
 
 def generate_audio(prompt, duration, steps, cfg_scale, seed, model_name):