diff --git a/vllm/v1/worker/block_table.py b/vllm/v1/worker/block_table.py index bf38e88f0..5662fc350 100644 --- a/vllm/v1/worker/block_table.py +++ b/vllm/v1/worker/block_table.py @@ -91,8 +91,7 @@ class BlockTable: # block_size. block_table_indices = (req_indices * self.max_num_blocks_per_req + positions // self.block_size) - block_table_cpu = self.get_cpu_tensor() - block_numbers = block_table_cpu.flatten()[block_table_indices].numpy() + block_numbers = self.block_table_np.ravel()[block_table_indices] block_offsets = positions % self.block_size np.add(block_numbers * self.block_size, block_offsets,