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 12, 02:15 UTC
[Bugfix] Fix adapter_enabled IMA (#29977)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Dec 4, 2025 at 12:51 UTC
dd38ba3a2682d6f73b02bc983a5b0157ed3e5498
1 file changed
+6
-2
vllm/lora/ops/triton_ops/fused_moe_lora_op.py
+6
-2
@@ -96,10 +96,14 @@ def _fused_moe_lora_kernel(
96
slice_id = tl.program_id(axis=1)
97
lora_idx = tl.program_id(axis=2)
98
lora_id = tl.load(lora_ids + lora_idx)
99
- moe_enabled = tl.load(adapter_enabled + lora_id)
100
- if lora_id == -1 or moe_enabled == 0:
99
+
100
+ if lora_id == -1:
101
# Early exit for the no-lora case.
102
return
103
+ moe_enabled = tl.load(adapter_enabled + lora_id)
104
+ if moe_enabled == 0:
105
+ # Early exit for the no moe lora case.
106
+ return
107
max_loras = tl.num_programs(axis=2)
108
grid_k = tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)
109