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 7, 20:15 UTC
[LoRA] Update LoRA expand kernel block_n calculation (#32621)
Signed-off-by: Xin Yang <xyangx@amazon.com>
Xin Yang committed
Feb 23, 2026 at 23:17 UTC
c870eb9e0f001ed5d07e6b6e2eb5e500a080a717
1 file changed
+1
-1
vllm/lora/ops/triton_ops/utils.py
+1
-1
@@ -251,7 +251,7 @@ def get_lora_op_configs(
251
else:
252
default = {
253
"block_m": 64,
254
- "block_n": max(64, next_power_of_2(128 // num_slices)),
254
+ "block_n": 64 if num_slices > 1 else 128,
255
"block_k": 16,
256
"num_warps": 4,
257
"num_ctas": 1,