use non_blocking h2d in ForwardBatch.prepare_mlp_sync_batch. (#11605)

python/sglang/srt/model_executor/forward_batch_info.py

@@ -734,9 +734,8 @@ class ForwardBatch:
             self.encoder_lens = self._pad_tensor_to_size(self.encoder_lens, bs)
         self.positions = self._pad_tensor_to_size(self.positions, num_tokens)
         self.global_num_tokens_cpu = global_num_tokens
-        self.global_num_tokens_gpu = self.global_num_tokens_gpu.new_tensor(
-            global_num_tokens
-        )
+        global_num_tokens_pinned = torch.tensor(global_num_tokens, pin_memory=True)
+        self.global_num_tokens_gpu.copy_(global_num_tokens_pinned, non_blocking=True)
 
         if self.mrope_positions is not None:
             self.mrope_positions = self._pad_tensor_to_size(self.mrope_positions, bs)