Remove unnecessary copy of gradients in util (#144329)

Summary:
No need to copy gradients to CPU too

X-link: pytorch/pytorch#144329
Approved by: https://github.com/awgu, https://github.com/cyyever

Reviewed By: clee2000

Differential Revision: D67943148

fbshipit-source-id: dea67477dbc0a791cf4a8c72ca7c9e4d4381cc54

userbenchmark/dynamo/dynamobench/_dynamo/utils.py

@@ -2492,7 +2492,7 @@ def to_tensor(t):
                 return True
             score = torch.nn.functional.cosine_similarity(ref, res, dim=0, eps=1e-6)
             if score < 0.99:
-                log.warning("Similarity score=%s", score.cpu().detach().item())
+                log.warning("Similarity score=%s", score.detach().cpu().item())
             return score >= 0.99
         else:
             if not exact_dtype: