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, 16:30 UTC
[BugFix] Align fused MoE-LoRA kernel config with actual weight shapes (#34396)
Signed-off-by: Runkai Tao <rt572@physics.rutgers.edu>
Runkai Tao committed
Feb 26, 2026 at 13:03 UTC
a1f53addb132f75704710184f4c1cc4780343329
1 file changed
+5
-1
vllm/lora/layers/fused_moe.py
+5
-1
@@ -83,7 +83,11 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
83
):
84
if envs.VLLM_TUNED_CONFIG_FOLDER:
85
hidden_size = layer.hidden_size
86
- intermediate_size = layer.intermediate_size_per_partition
86
+ intermediate_size = (
87
+ self.w2_lora_a_stacked[0].shape[-1]
88
+ if op_prefix == "w2"
89
+ else self.w13_lora_b_stacked[0].shape[-2]
90
+ )
91
shrink_config = get_lora_op_configs(
92
op_type=f"fused_moe_lora_{op_prefix}_shrink",
93
max_loras=num_loras,