feat(DeviceProcessor): Enhance tensor processing with device detection and float dtype conversion

- Improved the _process_tensor method to preserve GPU placement for tensors already on a GPU, facilitating multi-GPU training scenarios.
- Introduced a new _detect_device method in TokenizerProcessor to ensure tokenized tensors match the device of existing tensors in transitions.
- Added comprehensive unit tests to validate the functionality of device detection and float dtype conversion across various scenarios.

src/lerobot/processor/device_processor.py

@@ -66,9 +66,26 @@ class DeviceProcessor:
             self._target_float_dtype = None
 
     def _process_tensor(self, tensor: torch.Tensor) -> torch.Tensor:
-        """Process a tensor by moving to device and optionally converting float dtype."""
-        # Move to device first
-        tensor = tensor.to(self.device, non_blocking=self.non_blocking)
+        """Process a tensor by moving to device and optionally converting float dtype.
+
+        If the tensor is already on a GPU and we're configured for a GPU, it preserves
+        that GPU placement (useful for multi-GPU training with Accelerate).
+        Otherwise, it moves to the configured device.
+        """
+        # Determine target device
+        if tensor.is_cuda and self._device.type == "cuda":
+            # Both tensor and target are on GPU - preserve tensor's GPU placement
+            # This handles multi-GPU scenarios where Accelerate has already placed
+            # tensors on the correct GPU for each process
+            target_device = tensor.device
+        else:
+            # Either tensor is on CPU, or we're configured for CPU
+            # In both cases, use the configured device
+            target_device = self._device
+
+        # Only move if necessary
+        if tensor.device != target_device:
+            tensor = tensor.to(target_device, non_blocking=self.non_blocking)
 
         # Convert float dtype if specified and tensor is floating point
         if self._target_float_dtype is not None and tensor.is_floating_point():

src/lerobot/processor/tokenizer_processor.py

@@ -134,9 +134,19 @@ class TokenizerProcessor:
         if task is None:
             return transition
 
-        # Tokenize the task
+        # Tokenize the task (creates CPU tensors)
         tokenized_prompt = self._tokenize_text(task)
 
+        # Detect device from existing tensors in the transition
+        target_device = self._detect_device(transition)
+
+        # Move tokenized tensors to match the device of other data
+        if target_device is not None:
+            tokenized_prompt = {
+                k: v.to(target_device) if isinstance(v, torch.Tensor) else v
+                for k, v in tokenized_prompt.items()
+            }
+
         # Get or create observation dict
         observation = transition.get(TransitionKey.OBSERVATION)
         if observation is None:
@@ -153,6 +163,45 @@ class TokenizerProcessor:
         transition[TransitionKey.OBSERVATION.value] = observation  # type: ignore[misc]
         return transition
 
+    def _detect_device(self, transition: EnvTransition) -> torch.device | None:
+        """Detect device from existing tensors in the transition.
+
+        This allows the tokenized tensors to match the device of other data,
+        which is especially important for multi-GPU training with Accelerate.
+
+        Args:
+            transition: The transition to search for existing tensors.
+
+        Returns:
+            The device of the first tensor found, or None if no tensors exist.
+        """
+        # Check observation tensors first (most likely to exist)
+        observation = transition.get(TransitionKey.OBSERVATION)
+        if observation:
+            for value in observation.values():
+                if isinstance(value, torch.Tensor):
+                    return value.device
+
+        # Check action tensor
+        action = transition.get(TransitionKey.ACTION)
+        if isinstance(action, torch.Tensor):
+            return action.device
+
+        # Check other tensor fields
+        for key in [TransitionKey.REWARD, TransitionKey.DONE, TransitionKey.TRUNCATED]:
+            value = transition.get(key)
+            if isinstance(value, torch.Tensor):
+                return value.device
+
+        # Check complementary data for tensors
+        complementary_data = transition.get(TransitionKey.COMPLEMENTARY_DATA)
+        if complementary_data:
+            for value in complementary_data.values():
+                if isinstance(value, torch.Tensor):
+                    return value.device
+
+        return None  # No tensors found, keep on CPU
+
     def _tokenize_text(self, text: str | list[str]) -> dict[str, torch.Tensor]:
         """Tokenize text using the configured tokenizer.