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 4, 07:30 UTC
[XPU] enable is_act_and_mul for xpu (#37481)
Signed-off-by: Chendi Xue <chendi.xue@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Chendi.Xue committed
May 4, 2026 at 20:07 UTC
685bf811d65b58b2f8ef149d7da53dfd0393a912
2 files changed
+6
-3
vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
+2
-1
@@ -62,7 +62,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
62
63
@staticmethod
64
def _supports_no_act_and_mul() -> bool:
65
- return False
65
+ return True
66
67
@staticmethod
68
def _supports_activation(activation: MoEActivation) -> bool:
@@ -70,6 +70,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
70
MoEActivation.SILU,
71
MoEActivation.GELU,
72
MoEActivation.SWIGLUOAI,
73
+ MoEActivation.RELU2_NO_MUL,
74
]
75
76
@staticmethod
vllm/model_executor/layers/fused_moe/layer.py
+4
-2
@@ -538,9 +538,11 @@ class FusedMoE(PluggableLayer):
538
# for heuristic purposes, so it must be initialized first.
539
self.quant_method: FusedMoEMethodBase = _get_quant_method()
540
541
- if not self.moe_config.is_act_and_mul and not current_platform.is_cuda_alike():
541
+ if not self.moe_config.is_act_and_mul and not (
542
+ current_platform.is_cuda_alike() or current_platform.is_xpu()
543
+ ):
544
raise NotImplementedError(
543
- "is_act_and_mul=False is supported only for CUDA and ROCm for now"
545
+ "is_act_and_mul=False is supported only for CUDA and XPU for now"
546
)
547
548
if self.enable_eplb and not self.quant_method.supports_eplb: