diff --git a/demo/vibevoice_asr_inference_from_file.py b/demo/vibevoice_asr_inference_from_file.py
index bf4f75df..497ca743 100644
--- a/demo/vibevoice_asr_inference_from_file.py
+++ b/demo/vibevoice_asr_inference_from_file.py
@@ -54,12 +54,12 @@ def __init__(
         print(f"Using attention implementation: {attn_implementation}")
         self.model = VibeVoiceASRForConditionalGeneration.from_pretrained(
             model_path,
-            dtype=dtype,
+            torch_dtype=dtype,
             device_map=device if device == "auto" else None,
             attn_implementation=attn_implementation,
             trust_remote_code=True
         )
-        
+
         if device != "auto":
             self.model = self.model.to(device)
         
@@ -443,7 +443,7 @@ def main():
     parser.add_argument(
         "--device", 
         type=str, 
-        default="cuda" if torch.cuda.is_available() else ("xpu" if torch.backends.xpu.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu") ),
+        default="cuda" if torch.cuda.is_available() else ("xpu" if hasattr(torch.backends, 'xpu') and torch.backends.xpu.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu") ),
         choices=["cuda", "cpu", "mps","xpu", "auto"],
         help="Device to run inference on"
     )
@@ -533,7 +533,7 @@ def main():
     # Initialize model
     # Handle MPS device and dtype
     if args.device == "mps":
-        model_dtype = torch.float32  # MPS works better with float32
+        model_dtype = torch.float16  # float16 to fit within MPS memory (~14GB vs ~28GB for float32)
     elif args.device == "xpu":
         model_dtype = torch.float32
     elif args.device == "cpu":
diff --git a/vibevoice/modular/configuration_vibevoice.py b/vibevoice/modular/configuration_vibevoice.py
index 18451136..141f9ce3 100644
--- a/vibevoice/modular/configuration_vibevoice.py
+++ b/vibevoice/modular/configuration_vibevoice.py
@@ -14,17 +14,15 @@
 def _convert_dtype_to_string(config_dict: dict) -> dict:
     """
     Convert torch.dtype objects to their string representation for JSON serialization.
-    
-    This fixes the "Object of type dtype is not JSON serializable" error that occurs
-    when transformers tries to log/serialize the config with torch_dtype as a torch.dtype object.
-    
+    Recursively handles nested dicts.
+
     See: https://github.com/microsoft/VibeVoice/issues/199
     """
-    if "torch_dtype" in config_dict and config_dict["torch_dtype"] is not None:
-        dtype = config_dict["torch_dtype"]
-        if isinstance(dtype, torch.dtype):
-            # Convert torch.dtype to string (e.g., torch.bfloat16 -> "bfloat16")
-            config_dict["torch_dtype"] = str(dtype).replace("torch.", "")
+    for key, value in config_dict.items():
+        if isinstance(value, torch.dtype):
+            config_dict[key] = str(value).replace("torch.", "")
+        elif isinstance(value, dict):
+            config_dict[key] = _convert_dtype_to_string(value)
     return config_dict