[fix]修复零消耗引入的cudagraph模式数据准备cpu耗时问题

vllm/attention/backends/utils.py

@@ -242,8 +242,8 @@ class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
                     input_block_tables[i, :len(block_table)] = block_table
             # block_tables = torch.from_numpy(input_block_tables).to(
             #     device, non_blocking=True)
-            block_tables = async_tensor_h2d(input_block_tables.tolist(), torch.int32,
-                                    device, self.runner.pin_memory)
+            block_tables = torch.from_numpy(input_block_tables).pin_memory().to(
+                device, non_blocking=True)
 
         else:
             block_tables = make_tensor_with_pad(