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 13, 17:00 UTC
[BugFix] Fix FusedMoELoRA + ModularKernel Integration (#28237)
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com> Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Varun Sundar Rabindranath committed
Nov 6, 2025 at 17:53 UTC
ca6f755d2416bb1bd1b924b40481a41b2b0eaf85
1 file changed
+3
-3
vllm/lora/layers/fused_moe.py
+3
-3
@@ -25,6 +25,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
25
modular_triton_fused_moe,
26
try_get_optimal_moe_config,
27
)
28
+from vllm.model_executor.layers.fused_moe.layer import FusedMoEModularMethod
29
30
31
class FusedMoEWithLoRA(BaseLayerWithLoRA):
@@ -280,10 +281,9 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):
281
self.base_layer, fused_experts.moe_sum
282
)
283
283
- self.base_layer.quant_method.old_fused_experts = (
284
- self.base_layer.quant_method.fused_experts
284
+ self.base_layer.quant_method = FusedMoEModularMethod(
285
+ self.base_layer.quant_method, m_fused_moe_fn
286
)
286
- self.base_layer.quant_method.fused_experts = m_fused_moe_fn
287
288
def create_lora_weights(
289
self,