Fix meta tensor: add device_map=None and use torch_dtype parameter

acestep/handler.py

@@ -492,7 +492,8 @@ class AceStepHandler:
                             acestep_v15_checkpoint_path,
                             trust_remote_code=True,
                             attn_implementation=candidate,
-                            dtype=self.dtype,
+                            torch_dtype=self.dtype,
+                            device_map=None,  # Disable automatic device mapping
                             low_cpu_mem_usage=False,  # Disable meta device for custom model with assertions
                         )
                         attn_implementation = candidate