cpu for fatser whisper

app.py

@@ -29,16 +29,24 @@ logging.basicConfig(
 logger = logging.getLogger(__name__)
 
 # Configuration - Auto-detect GPU
-if torch.cuda.is_available():
-    DEVICE = "cuda"
-    COMPUTE_TYPE = "float16"  # Use float16 for GPU
+# Note: faster-whisper uses ctranslate2 which doesn't work well with ZeroGPU,
+# so we always use CPU for Whisper. TTS will use GPU when available.
+if torch.cuda.is_available() and not SPACES_AVAILABLE:
+    # Only use GPU for local CUDA setups, not ZeroGPU
+    TTS_DEVICE = "cuda"
     logger_temp = logging.getLogger(__name__)
-    logger_temp.info(f"🚀 GPU detected! Using CUDA with {torch.cuda.get_device_name(0)}")
+    logger_temp.info(f"🚀 GPU detected! Using CUDA with {torch.cuda.get_device_name(0)} for TTS")
 else:
-    DEVICE = "cpu"
-    COMPUTE_TYPE = "int8"  # Use int8 for CPU
+    TTS_DEVICE = "cpu"
     logger_temp = logging.getLogger(__name__)
-    logger_temp.info("Running on CPU")
+    if SPACES_AVAILABLE:
+        logger_temp.info("🚀 Running on ZeroGPU - TTS will use GPU inside decorated function")
+    else:
+        logger_temp.info("Running on CPU")
+
+# Whisper always uses CPU (ctranslate2 compatibility)
+WHISPER_DEVICE = "cpu"
+WHISPER_COMPUTE_TYPE = "int8"
 
 # Set temp directory to writable location
 os.environ['TMPDIR'] = '/tmp'
@@ -58,23 +66,25 @@ tts_model = None
 # ==================== Model Loading ====================
 
 def load_models():
-    """Load models on startup"""
+    """Load models (lazy loading for ZeroGPU compatibility)"""
     global whisper_model, tts_model
     
     if whisper_model is None:
         logger.info("Loading Whisper model...")
         whisper_model = WhisperModel(
             "small",
-            device=DEVICE,
-            compute_type=COMPUTE_TYPE,
+            device=WHISPER_DEVICE,
+            compute_type=WHISPER_COMPUTE_TYPE,
             cpu_threads=4
         )
         logger.info("✅ Whisper model loaded!")
     
     if tts_model is None:
         logger.info("Loading TTS model...")
-        tts_model = ChatterboxMultilingualTTS.from_pretrained(device=DEVICE)
-        logger.info("✅ TTS model loaded!")
+        # In ZeroGPU, determine device at runtime
+        tts_device = "cuda" if (SPACES_AVAILABLE and torch.cuda.is_available()) else TTS_DEVICE
+        tts_model = ChatterboxMultilingualTTS.from_pretrained(device=tts_device)
+        logger.info(f"✅ TTS model loaded on {tts_device}!")
     
     return whisper_model, tts_model
 
@@ -560,10 +570,13 @@ def create_interface():
 # ==================== Main ====================
 
 if __name__ == "__main__":
-    # Load models at startup
-    logger.info("Initializing models...")
-    load_models()
-    logger.info("Models loaded successfully!")
+    # Load models at startup (except in ZeroGPU where GPU isn't available yet)
+    if not SPACES_AVAILABLE:
+        logger.info("Initializing models...")
+        load_models()
+        logger.info("Models loaded successfully!")
+    else:
+        logger.info("Running in ZeroGPU mode - models will be loaded on first request")
     
     # Create and launch interface
     # .queue() is essential for long-running tasks like model generation