Add custom inference handler for Maya1 TTS

handler.py

@@ -49,10 +49,39 @@ class EndpointHandler:
         if device_override == "cpu":
             device_map_arg = "cpu"
 
-        self.model = AutoModelForCausalLM.from_pretrained(
-            path, torch_dtype=torch_dtype, device_map=device_map_arg
-        )
-        self.tokenizer = AutoTokenizer.from_pretrained(path)
+        # Check if model path has a valid config before attempting to load
+        model_path = path if path else "/repository"
+        
+        # Check if this is a valid model directory
+        config_path = os.path.join(model_path, "config.json")
+        if not os.path.exists(config_path):
+            raise RuntimeError(
+                f"❌ Model configuration not found at: {config_path}\n\n"
+                f"The repository appears to be missing model weights and configuration.\n\n"
+                f"To fix this:\n"
+                f"1. For TESTING: Set environment variable MAYA_USE_FAKE=1 to use fake mode\n"
+                f"2. For PRODUCTION: Upload Maya1 model weights to your repository:\n"
+                f"   - config.json\n"
+                f"   - model.safetensors (or pytorch_model.bin)\n"
+                f"   - tokenizer.json\n"
+                f"   - tokenizer_config.json\n"
+                f"   - generation_config.json (optional)\n\n"
+                f"Current path: {model_path}\n"
+                f"Files found: {os.listdir(model_path) if os.path.exists(model_path) else 'path does not exist'}\n"
+            )
+
+        try:
+            self.model = AutoModelForCausalLM.from_pretrained(
+                path, torch_dtype=torch_dtype, device_map=device_map_arg
+            )
+            self.tokenizer = AutoTokenizer.from_pretrained(path)
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to load Maya1 model from {path}.\n"
+                f"Error: {e}\n\n"
+                f"Please ensure the repository contains valid Maya1 model files.\n"
+                f"For testing without model weights, set MAYA_USE_FAKE=1"
+            )
 
         # determine device from model parameters (safer than using `model.device`)
         try: