A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Core] direct indexing on self.block_table_np in compute_slot_mapping (#22940)
Signed-off-by: linzebing <linzebing1995@gmail.com>
Zebing Lin committed
Aug 15, 2025 at 15:12 UTC
6e670778cdd87c282c42002a2304cb0a4a165904
1 file changed
+1
-2
vllm/v1/worker/block_table.py
+1
-2
@@ -91,8 +91,7 @@ class BlockTable:
91
# block_size.
92
block_table_indices = (req_indices * self.max_num_blocks_per_req +
93
positions // self.block_size)
94
- block_table_cpu = self.get_cpu_tensor()
95
- block_numbers = block_table_cpu.flatten()[block_table_indices].numpy()
94
+ block_numbers = self.block_table_np.ravel()[block_table_indices]
95
block_offsets = positions % self.block_size
96
np.add(block_numbers * self.block_size,
97
block_offsets,