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, 18:15 UTC
[XPU] Enable custom routing functions in IPEX for Llama4 (#28004)
Signed-off-by: frost-intel <frost.mitchell@intel.com>
Frost Mitchell committed
Nov 5, 2025 at 08:39 UTC
6e97eccf5dd5036e26d63141d2bc1a9ea17a2cc8
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/layer.py
+1
-1
@@ -915,7 +915,6 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
915
or logical_replica_count is not None
916
):
917
raise NotImplementedError("Expert load balancing is not supported for XPU.")
918
- assert custom_routing_function is None
918
return layer.ipex_fusion(
919
x,
920
use_grouped_topk,
@@ -924,6 +923,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
923
renormalize,
924
topk_group,
925
num_expert_group,
926
+ custom_routing_function=custom_routing_function,
927
)
928
929
def forward_tpu(