Add custom inference handler for Maya1 TTS

handler.py

@@ -49,38 +49,52 @@ class EndpointHandler:
         if device_override == "cpu":
             device_map_arg = "cpu"
 
-        # Check if model path has a valid config before attempting to load
-        model_path = path if path else "/repository"
+        # Allow loading from HuggingFace Hub via environment variable
+        # MAYA_MODEL_ID: HuggingFace model ID (e.g., "Plachta/MARS5-TTS" or similar Maya model)
+        model_id = os.getenv("MAYA_MODEL_ID", "")
         
-        # 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"
-            )
+        # Determine model path: use env var if set, otherwise local path
+        if model_id:
+            # Load from HuggingFace Hub
+            model_path = model_id
+            print(f"Loading Maya1 model from HuggingFace Hub: {model_id}")
+        else:
+            # Load from local repository
+            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 REMOTE MODEL: Set MAYA_MODEL_ID to a HuggingFace model ID\n"
+                    f"   Example: MAYA_MODEL_ID=Plachta/MARS5-TTS\n"
+                    f"3. For LOCAL MODEL: 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
+                model_path, torch_dtype=torch_dtype, device_map=device_map_arg
             )
-            self.tokenizer = AutoTokenizer.from_pretrained(path)
+            self.tokenizer = AutoTokenizer.from_pretrained(model_path)
         except Exception as e:
             raise RuntimeError(
-                f"Failed to load Maya1 model from {path}.\n"
+                f"Failed to load Maya1 model from {model_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"
+                f"Please ensure:\n"
+                f"1. The model repository contains valid model files, OR\n"
+                f"2. Set MAYA_MODEL_ID to a valid HuggingFace model ID, OR\n"
+                f"3. Set MAYA_USE_FAKE=1 for testing without a real model"
             )
 
         # determine device from model parameters (safer than using `model.device`)