Update

[ghstack-poisoned]

tensordict/nn/cudagraphs.py

@@ -28,9 +28,9 @@
 from torch.utils._pytree import SUPPORTED_NODES, tree_map
 
 try:
-    from torch.utils._pytree import tree_leaves
+    from torch.utils._pytree import tree_flatten, tree_leaves, tree_unflatten
 except ImportError:
-    from torch.utils._pytree import tree_flatten
+    from torch.utils._pytree import tree_flatten, tree_unflatten
 
     def tree_leaves(pytree):
         """Torch 2.0 compatible version of tree_leaves."""
@@ -293,11 +293,12 @@ def check_tensor_id(name, t0, t1):
 
             def _call(*args: torch.Tensor, **kwargs: torch.Tensor):
                 if self.counter >= self._warmup:
-                    tree_map(
-                        lambda x, y: x.copy_(y, non_blocking=True),
-                        (self._args, self._kwargs),
-                        (args, kwargs),
-                    )
+                    srcs, dests = [], []
+                    for arg_src, arg_dest in zip(
+                        tree_leaves((args, kwargs)), self._flat_tree
+                    ):
+                        self._maybe_copy_onto_(arg_src, arg_dest, srcs, dests)
+                    torch._foreach_copy_(dests, srcs)
                     torch.cuda.synchronize()
                     self.graph.replay()
                     if self._return_unchanged == "clone":
@@ -322,8 +323,13 @@ def _call(*args: torch.Tensor, **kwargs: torch.Tensor):
                     self.counter += self._has_cuda
                     return out
                 else:
-                    args, kwargs = self._args, self._kwargs = tree_map(
-                        self._check_device_and_clone, (args, kwargs)
+                    self._flat_tree, self._tree_spec = tree_flatten((args, kwargs))
+
+                    self._flat_tree = tuple(
+                        self._check_device_and_clone(arg) for arg in self._flat_tree
+                    )
+                    args, kwargs = self._args, self._kwargs = tree_unflatten(
+                        self._flat_tree, self._tree_spec
                     )
 
                     torch.cuda.synchronize()
@@ -360,6 +366,21 @@ def _call(*args: torch.Tensor, **kwargs: torch.Tensor):
         _call_func = functools.wraps(self.module)(_call)
         self._call_func = _call_func
 
+    @staticmethod
+    def _maybe_copy_onto_(src, dest, srcs, dests):
+        if isinstance(src, (torch.Tensor, TensorDictBase)):
+            srcs.append(src)
+            dests.append(dest)
+        try:
+            if src != dest:
+                raise ValueError("Varying inputs must be torch.Tensor subclasses.")
+        except Exception:
+            raise RuntimeError(
+                "Couldn't assess input value. Make sure your function only takes tensor inputs or that "
+                "the input value can be easily checked and is constant. For a better efficiency, avoid "
+                "passing non-tensor inputs to your function."
+            )
+
     @classmethod
     def _check_device_and_clone(cls, x):
         if isinstance(x, torch.Tensor) or is_tensor_collection(x):

test/test_compile.py

@@ -1056,6 +1056,19 @@ def test_td_input_non_tdmodule(self, compiled):
             if i == 5:
                 assert not func._is_tensordict_module
 
+    def test_td_input_non_tdmodule_nontensor(self, compiled):
+        func = lambda x, y: x + y
+        func = self._make_cudagraph(func, compiled)
+        for i in range(10):
+            assert func(torch.zeros(()), 1.0) == 1.0
+            if i == 5:
+                assert not func._is_tensordict_module
+        if torch.cuda.is_available():
+            with pytest.raises(
+                ValueError, match="Varying inputs must be torch.Tensor subclasses."
+            ):
+                func(torch.zeros(()), 2.0)
+
 
 if __name__ == "__main__":
     args, unknown = argparse.ArgumentParser().parse_known_args()