[BugFix] Remove redundant device_put. (vllm-project#1099)

Author: Lumosis

tpu_inference/runner/tpu_runner.py

@@ -1315,10 +1315,10 @@ def _prepare_inputs_dp(self, scheduler_output: "VllmSchedulerOutput"):
         seq_lens_cpu = seq_lens
 
         (input_ids, positions, block_tables, query_start_loc, seq_lens,
-         logits_indices, request_distribution, logits_indices) = device_array(
+         logits_indices, request_distribution) = device_array(
              self.mesh,
              (input_ids, positions, block_tables, query_start_loc, seq_lens,
-              logits_indices, request_distribution, logits_indices),
+              logits_indices, request_distribution),
              sharding=data_parallel_attn_sharding,
          )
         # Async scheduling: substitute placeholder tokens for DP