change torch dtype

vitpose.py

@@ -17,7 +17,7 @@ class VitPose:
             object_detection_checkpoint="PekingU/rtdetr_r50vd_coco_o365",
             pose_estimation_checkpoint="usyd-community/vitpose-plus-small",
             device="cuda" if torch.cuda.is_available() else "cpu",
-            dtype=torch.bfloat16,
+            dtype=torch.float16,
             compile=True,  # or True to get more speedup
         )
         self.output_video_path = None