[BugFix] Fix device transfer for collectors with init_random_frames mixed devices

ghstack-source-id: 1684399
Pull Request resolved: #2704

Author: Vincent Moens

torchrl/collectors/collectors.py

@@ -1148,6 +1148,21 @@ def rollout(self) -> TensorDictBase:
                     and self._frames < self.init_random_frames
                 ):
                     self.env.rand_action(self._shuttle)
+                    if (
+                        self.policy_device is not None
+                        and self.policy_device != self.env_device
+                    ):
+                        # TODO: This may break with exclusive / ragged lazy stacks
+                        self._shuttle.apply(
+                            lambda name, val: val.to(
+                                device=self.policy_device, non_blocking=True
+                            )
+                            if name in self._policy_output_keys
+                            else val,
+                            out=self._shuttle,
+                            named=True,
+                            nested_keys=True,
+                        )
                 else:
                     if self._cast_to_policy_device:
                         if self.policy_device is not None: