Add custom inference handler for Maya1 TTS

handler.py

@@ -52,17 +52,17 @@ class EndpointHandler:
         # Allow loading from HuggingFace Hub via environment variable
         # MAYA_MODEL_ID: HuggingFace model ID (e.g., "maya-research/maya1" or similar Maya model)
         model_id = os.getenv("MAYA_MODEL_ID", "")
-        
+
         # Determine model path: use env var if set, otherwise local path
         local_path = path if path else "/repository"
-        
+
         # Check if trying to load from self (circular reference)
         # Only block if it's exactly "binhqd/maya1" - other repos ending in /maya1 are fine
         if model_id == "binhqd/maya1":
             print(f"⚠️ MAYA_MODEL_ID is set to {model_id} which is this repository!")
             print("This creates a circular reference. Checking local files instead...")
             model_id = ""  # Force local loading
-        
+
         if model_id:
             # Load from HuggingFace Hub
             model_path = model_id
@@ -70,7 +70,7 @@ class EndpointHandler:
         else:
             # Load from local repository
             model_path = local_path
-            
+
             # Check if this is a valid model directory
             config_path = os.path.join(model_path, "config.json")
             if not os.path.exists(config_path):
@@ -243,10 +243,13 @@ class EndpointHandler:
             waveform = self.snac.decode(snac_codes.to(self.device))
 
         # Extract audio and convert to numpy
-        if waveform.dim() == 3:  # (batch, channels, samples)
-            waveform = waveform.squeeze(0).squeeze(0)  # Remove batch and channel dims
-        elif waveform.dim() == 2:  # (batch, samples)
-            waveform = waveform.squeeze(0)
+        # SNAC outputs shape (batch, samples) or (batch, 1, samples)
+        # Safely remove all size-1 dimensions
+        waveform = waveform.squeeze()  # Remove all dimensions of size 1
+        
+        # Ensure we have a 1D tensor
+        if waveform.dim() > 1:
+            waveform = waveform[0]  # Take first item if still multi-dimensional
 
         waveform = waveform.cpu().numpy()