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 8, 05:00 UTC
Fix call to moe_mk in modelopt MoE modules (required for LoRA) (#34575)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
danisereb committed
Feb 16, 2026 at 17:33 UTC
ec7d9e67459dd8ca6e2e3e77a40993291cab152c
1 file changed
+10
-10
vllm/model_executor/layers/quantization/modelopt.py
+10
-10
@@ -977,11 +977,11 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
977
978
assert self.moe_mk is not None
979
return self.moe_mk(
980
- x,
981
- layer.w13_weight,
982
- layer.w2_weight,
983
- topk_weights,
984
- topk_ids,
980
+ hidden_states=x,
981
+ w1=layer.w13_weight,
982
+ w2=layer.w2_weight,
983
+ topk_weights=topk_weights,
984
+ topk_ids=topk_ids,
985
activation=layer.activation,
986
global_num_experts=layer.global_num_experts,
987
expert_map=layer.expert_map,
@@ -1549,11 +1549,11 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
1549
else:
1550
assert self.moe_mk is not None
1551
return self.moe_mk(
1552
- x,
1553
- layer.w13_weight,
1554
- layer.w2_weight,
1555
- topk_weights,
1556
- topk_ids,
1552
+ hidden_states=x,
1553
+ w1=layer.w13_weight,
1554
+ w2=layer.w2_weight,
1555
+ topk_weights=topk_weights,
1556
+ topk_ids=topk_ids,
1557
activation=layer.activation,
1558
global_num_experts=layer.global_num_experts,
1559
expert_map=layer.expert_map,