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 9, 14:45 UTC
Set splitk=1 for fused-moe-lora expand kernel (#32882)
Signed-off-by: Danielle Robinson <dmmaddix@amazon.com> Co-authored-by: Danielle Robinson <dmmaddix@amazon.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Danielle Robinson committed
Jan 25, 2026 at 22:52 UTC
ee484b3f4b0c061d2612ea7c0cb40b44baf680c0
1 file changed
+1
-1
vllm/lora/ops/triton_ops/fused_moe_lora_op.py
+1
-1
@@ -351,7 +351,7 @@ def _fused_moe_lora_expand(
351
"GROUP_SIZE_M": group_size_m,
352
"num_warps": num_warps,
353
"num_stages": num_stages,
354
- "SPLIT_K": split_k, # Set split_k = 1 for expand calls
354
+ "SPLIT_K": 1, # Set split_k = 1 for expand calls
355
"USE_GDC": use_gdc,
356
"launch_pdl": use_gdc, # triton kernel metadata
357
}