[ROCm][Regression] Remove tensor creation that harms performance on ROCm (vllm-project#20741)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Signed-off-by: Patrick von Platen <patrick.v.platen@gmail.com>

Author: gshtras

vllm/platforms/rocm.py

@@ -247,10 +247,6 @@ def set_device(cls, device: torch.device) -> None:
         Set the device for the current platform.
         """
         torch.cuda.set_device(device)
-        # With this trick we can force the device to be set eagerly
-        # see https://github.com/pytorch/pytorch/issues/155668
-        # for why and when it is needed
-        _ = torch.zeros(1, device=device)
 
     @classmethod
     @lru_cache(maxsize=8)