diff --git a/lighthouse/feature_extractor/vision_encoders/slowfast.py b/lighthouse/feature_extractor/vision_encoders/slowfast.py
index d1edda6..1949617 100644
--- a/lighthouse/feature_extractor/vision_encoders/slowfast.py
+++ b/lighthouse/feature_extractor/vision_encoders/slowfast.py
@@ -93,8 +93,8 @@ def __call__(
         slowfast_frames: torch.Tensor,
         bsz: int = 45):
         n_chunk = len(slowfast_frames)
-        features = torch.HalfTensor(n_chunk, self.SLOWFAST_FEATURE_DIM,
-                                    device=self._device).fill_(0)
+        features = torch.zeros([n_chunk, self.SLOWFAST_FEATURE_DIM],
+                                device=self._device, dtype=torch.float16)
         n_batch = int(math.ceil(n_chunk / bsz))
         for i in range(n_batch):
             st_idx = i * bsz
@@ -106,5 +106,4 @@ def __call__(
             inputs = self._pack_pathway_output(fast_clip)
             batch_features = self._slowfast_extractor(inputs)
             features[st_idx:ed_idx] = batch_features.half()
-        slowfast_features = features.cpu()
-        return slowfast_features
\ No newline at end of file
+        return features
\ No newline at end of file