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 10, 10:15 UTC
[Bugfix] Fix FusedMoE LoRA w2_output_size (#31949)
Signed-off-by: Xin Yang <xyangx@amazon.com>
Xin Yang committed
Jan 8, 2026 at 21:54 UTC
707b240d7e7523c652b8a5b075ffc1ee9e1120ae
1 file changed
+1
-1
vllm/lora/layers/fused_moe.py
+1
-1
@@ -720,7 +720,7 @@ class FusedMoE3DWithLoRA(FusedMoEWithLoRA):
720
"""
721
Full size
722
"""
723
- return self.w2_lora_a_stacked[0].shape[-2]
723
+ return self.base_layer.hidden_size
724
725
@classmethod
726
def can_replace_layer(