Fix the memory usage issue of logits in generate() (huggingface#34813)

Author: kjohew

src/transformers/generation/utils.py

@@ -3246,7 +3246,7 @@ def _sample(
 
             # Clone is needed to avoid keeping a hanging ref to outputs.logits which may be very large for first iteration
             # (the clone itself is always small)
-            next_token_logits = outputs.logits.clone()[:, -1, :].float()
+            next_token_logits = outputs.logits[:, -1, :].clone().float()
             next_token_logits = next_token_logits.to(input_ids.device)
 
             # pre-process distribution