Update app.py

app.py

@@ -1,4 +1,6 @@
 import os
+import tempfile
+from pathlib import Path
 
 import gradio as gr
 import numpy as np
@@ -9,35 +11,38 @@ from huggingface_hub import login
 from stable_audio_3 import StableAudioModel
 
 
-hf_token = os.environ.get("HF_TOKEN")
+HF_TOKEN = os.environ.get("HF_TOKEN")
 
-if not hf_token:
+if not HF_TOKEN:
     raise RuntimeError("HF_TOKEN Space secret is missing.")
 
-login(token=hf_token)
+login(token=HF_TOKEN)
 
+# Safe diagnostics only. Do not call torch.cuda.get_arch_list(),
+# torch.cuda.get_device_name(), or similar CUDA inspection functions
+# during ZeroGPU startup.
 print("PyTorch:", torch.__version__)
-print("PyTorch CUDA:", torch.version.cuda)
-print("Compiled architectures:", torch.cuda.get_arch_list())
+print("PyTorch CUDA build:", torch.version.cuda)
+print("Loading Stable Audio 3 Small SFX...")
 
+# ZeroGPU emulates CUDA placement during startup and later transfers the
+# packed tensors into the temporary GPU worker.
 model = StableAudioModel.from_pretrained(
     "small-sfx",
     device="cuda",
 )
 
-def requested_gpu_duration(prompt: str, duration: float) -> int:
-    """
-    Reserve a modest amount of GPU time.
+print("Model loaded.")
+
 
-    Stable Audio 3 Small models are extremely fast on modern GPUs,
-    but leave room for initial setup and audio decoding.
-    """
+def requested_gpu_duration(prompt: str, duration: float) -> int:
     del prompt
 
     duration = float(duration)
 
     if duration <= 15:
         return 15
+
     if duration <= 60:
         return 20
 
@@ -66,7 +71,7 @@ def generate_sound(prompt: str, duration: float) -> str:
 
     audio = np.asarray(audio)
 
-    # Remove batch dimension.
+    # Remove a batch dimension: [batch, channels, samples].
     if audio.ndim == 3:
         audio = audio[0]
 
@@ -74,6 +79,12 @@ def generate_sound(prompt: str, duration: float) -> str:
     if audio.ndim == 2 and audio.shape[0] <= 8:
         audio = audio.T
 
+    # Avoid clipping if the model produces samples outside [-1, 1].
+    peak = float(np.max(np.abs(audio))) if audio.size else 0.0
+
+    if peak > 1.0:
+        audio = audio / peak
+
     with tempfile.NamedTemporaryFile(
         suffix=".wav",
         delete=False,
@@ -81,7 +92,7 @@ def generate_sound(prompt: str, duration: float) -> str:
         output_path = Path(output_file.name)
 
     sf.write(
-        output_path,
+        str(output_path),
         audio,
         samplerate=44_100,
         subtype="FLOAT",