Update app.py

app.py

@@ -39,20 +39,39 @@ import shutil
 
 import spaces
 import torch
+
 torch._dynamo.config.suppress_errors = True
 torch._dynamo.config.disable = True
 
 _original_tensor_to = torch.Tensor.to
-def _coerce_cuda_str(self, *args, **kwargs):
-    if args and isinstance(args[0], torch.device) and args[0].type == "cuda":
-        args = ("cuda", *args[1:])
-    if kwargs.get("device") is not None:
-        dev = kwargs["device"]
-        if isinstance(dev, torch.device) and dev.type == "cuda":
-            kwargs = {**kwargs, "device": "cuda"}
+
+
+def _is_cuda_target(x):
+    return (
+        x == "cuda"
+        or (isinstance(x, torch.device) and x.type == "cuda")
+        or (isinstance(x, str) and x.startswith("cuda"))
+        or (isinstance(x, int) and x == 0)
+    )
+
+
+def _spaces_safe_to(self, *args, **kwargs):
+    """ZeroGPU emulates bare .to('cuda'), but LTX-2 uses non_blocking/copy."""
+    if args and _is_cuda_target(args[0]):
+        # keep dtype if it was passed positionally after the device
+        new_args = ("cuda",) + args[1:]
+        new_kwargs = {k: v for k, v in kwargs.items() if k not in ("non_blocking", "copy")}
+        return _original_tensor_to(self, *new_args, **new_kwargs)
+
+    if kwargs.get("device") is not None and _is_cuda_target(kwargs["device"]):
+        new_kwargs = {k: v for k, v in kwargs.items() if k not in ("non_blocking", "copy")}
+        new_kwargs["device"] = "cuda"
+        return _original_tensor_to(self, *args, **new_kwargs)
+
     return _original_tensor_to(self, *args, **kwargs)
 
-torch.Tensor.to = _coerce_cuda_str
+
+torch.Tensor.to = _spaces_safe_to
 
 import gradio as gr
 import numpy as np