A high-throughput and memory-efficient inference and serving engine for LLMs
[XPU][Minor] format moe kernel name and add in kernel list (#44771)
Signed-off-by: Yan Ma <yan.ma@intel.com>
Yan Ma committed
Jun 8, 2026 at 13:58 UTC
54c660c3a6c30c79e6457a162a19f4ed68042743
5 files changed
+13
-11
vllm/model_executor/layers/fused_moe/__init__.py
+4
-2
@@ -113,7 +113,7 @@ if HAS_TRITON:
113
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
114
XPUExperts,
115
XPUExpertsFp8,
116
- XPUExpertsMXFp4,
116
+ XPUExpertsMxFp4,
117
)
118
from vllm.model_executor.layers.fused_moe.fused_moe import (
119
fused_experts,
@@ -143,7 +143,9 @@ if HAS_TRITON:
143
"TritonOrDeepGemmExperts",
144
"XPUExperts",
145
"XPUExpertsFp8",
146
- "XPUExpertsMXFp4",
146
+ "XPUExpertsBlockFp8",
147
+ "XPUExpertsMxFp8",
148
+ "XPUExpertsMxFp4",
149
]
150
else:
151
# Some model classes directly use the custom ops. Add placeholders
vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
+2
-2
@@ -213,7 +213,7 @@ class XPUExpertsFp8(XPUExperts):
213
return (weight_key, activation_key) in SUPPORTED_W_A
214
215
216
-class XPUExpertsMxfp8(XPUExpertsFp8):
216
+class XPUExpertsMxFp8(XPUExpertsFp8):
217
def __init__(
218
self,
219
moe_config: FusedMoEConfig,
@@ -305,7 +305,7 @@ class XPUExpertsWNA16(XPUExperts):
305
return (weight_key, activation_key) == (kInt4Static, None)
306
307
308
-class XPUExpertsMXFp4(XPUExperts):
308
+class XPUExpertsMxFp4(XPUExperts):
309
def __init__(
310
self,
311
moe_config: FusedMoEConfig,
vllm/model_executor/layers/fused_moe/oracle/fp8.py
+2
-2
@@ -186,10 +186,10 @@ def backend_to_kernel_cls(
186
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
187
XPUExpertsBlockFp8,
188
XPUExpertsFp8,
189
- XPUExpertsMxfp8,
189
+ XPUExpertsMxFp8,
190
)
191
192
- return [XPUExpertsFp8, XPUExpertsMxfp8, XPUExpertsBlockFp8]
192
+ return [XPUExpertsFp8, XPUExpertsMxFp8, XPUExpertsBlockFp8]
193
194
elif backend == Fp8MoeBackend.CPU:
195
from vllm.model_executor.layers.fused_moe.experts.cpu_moe import (
vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
+2
-2
@@ -206,9 +206,9 @@ def backend_to_kernel_cls(
206
return [AiterExperts]
207
208
elif backend == Mxfp4MoeBackend.XPU:
209
- from vllm.model_executor.layers.fused_moe.experts.xpu_moe import XPUExpertsMXFp4
209
+ from vllm.model_executor.layers.fused_moe.experts.xpu_moe import XPUExpertsMxFp4
210
211
- return [XPUExpertsMXFp4]
211
+ return [XPUExpertsMxFp4]
212
213
elif backend == Mxfp4MoeBackend.CPU:
214
from vllm.model_executor.layers.fused_moe.experts.cpu_moe import CPUExpertsMxfp4
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py
+3
-3
@@ -22,7 +22,7 @@ from vllm.model_executor.layers.fused_moe.experts.marlin_moe import (
22
MarlinExperts,
23
)
24
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
25
- XPUExpertsMXFp4,
25
+ XPUExpertsMxFp4,
26
)
27
from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
28
Mxfp4MoeBackend,
@@ -54,8 +54,8 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
54
self.experts_cls = CutlassExpertsMxfp4
55
elif current_platform.is_xpu():
56
self.mxfp4_backend = Mxfp4MoeBackend.XPU
57
- self.experts_cls = XPUExpertsMXFp4
58
- logger.info_once("Using XPUExpertsMXFp4 for MXFP4 MoE on XPU platform")
57
+ self.experts_cls = XPUExpertsMxFp4
58
+ logger.info_once("Using XPUExpertsMxFp4 for MXFP4 MoE on XPU platform")
59
else:
60
logger.info_once("Using MarlinExperts for MXFP4 MoE")
61
self.experts_cls = MarlinExperts