[XPU][Minor] format moe kernel name and add in kernel list (#44771)
Signed-off-by: Yan Ma <yan.ma@intel.com>
This commit is contained in:
@@ -113,7 +113,7 @@ if HAS_TRITON:
|
||||
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
|
||||
XPUExperts,
|
||||
XPUExpertsFp8,
|
||||
XPUExpertsMXFp4,
|
||||
XPUExpertsMxFp4,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import (
|
||||
fused_experts,
|
||||
@@ -143,7 +143,9 @@ if HAS_TRITON:
|
||||
"TritonOrDeepGemmExperts",
|
||||
"XPUExperts",
|
||||
"XPUExpertsFp8",
|
||||
"XPUExpertsMXFp4",
|
||||
"XPUExpertsBlockFp8",
|
||||
"XPUExpertsMxFp8",
|
||||
"XPUExpertsMxFp4",
|
||||
]
|
||||
else:
|
||||
# Some model classes directly use the custom ops. Add placeholders
|
||||
|
||||
@@ -213,7 +213,7 @@ class XPUExpertsFp8(XPUExperts):
|
||||
return (weight_key, activation_key) in SUPPORTED_W_A
|
||||
|
||||
|
||||
class XPUExpertsMxfp8(XPUExpertsFp8):
|
||||
class XPUExpertsMxFp8(XPUExpertsFp8):
|
||||
def __init__(
|
||||
self,
|
||||
moe_config: FusedMoEConfig,
|
||||
@@ -305,7 +305,7 @@ class XPUExpertsWNA16(XPUExperts):
|
||||
return (weight_key, activation_key) == (kInt4Static, None)
|
||||
|
||||
|
||||
class XPUExpertsMXFp4(XPUExperts):
|
||||
class XPUExpertsMxFp4(XPUExperts):
|
||||
def __init__(
|
||||
self,
|
||||
moe_config: FusedMoEConfig,
|
||||
|
||||
@@ -186,10 +186,10 @@ def backend_to_kernel_cls(
|
||||
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
|
||||
XPUExpertsBlockFp8,
|
||||
XPUExpertsFp8,
|
||||
XPUExpertsMxfp8,
|
||||
XPUExpertsMxFp8,
|
||||
)
|
||||
|
||||
return [XPUExpertsFp8, XPUExpertsMxfp8, XPUExpertsBlockFp8]
|
||||
return [XPUExpertsFp8, XPUExpertsMxFp8, XPUExpertsBlockFp8]
|
||||
|
||||
elif backend == Fp8MoeBackend.CPU:
|
||||
from vllm.model_executor.layers.fused_moe.experts.cpu_moe import (
|
||||
|
||||
@@ -206,9 +206,9 @@ def backend_to_kernel_cls(
|
||||
return [AiterExperts]
|
||||
|
||||
elif backend == Mxfp4MoeBackend.XPU:
|
||||
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import XPUExpertsMXFp4
|
||||
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import XPUExpertsMxFp4
|
||||
|
||||
return [XPUExpertsMXFp4]
|
||||
return [XPUExpertsMxFp4]
|
||||
|
||||
elif backend == Mxfp4MoeBackend.CPU:
|
||||
from vllm.model_executor.layers.fused_moe.experts.cpu_moe import CPUExpertsMxfp4
|
||||
|
||||
+3
-3
@@ -22,7 +22,7 @@ from vllm.model_executor.layers.fused_moe.experts.marlin_moe import (
|
||||
MarlinExperts,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
|
||||
XPUExpertsMXFp4,
|
||||
XPUExpertsMxFp4,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import (
|
||||
Mxfp4MoeBackend,
|
||||
@@ -54,8 +54,8 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
|
||||
self.experts_cls = CutlassExpertsMxfp4
|
||||
elif current_platform.is_xpu():
|
||||
self.mxfp4_backend = Mxfp4MoeBackend.XPU
|
||||
self.experts_cls = XPUExpertsMXFp4
|
||||
logger.info_once("Using XPUExpertsMXFp4 for MXFP4 MoE on XPU platform")
|
||||
self.experts_cls = XPUExpertsMxFp4
|
||||
logger.info_once("Using XPUExpertsMxFp4 for MXFP4 MoE on XPU platform")
|
||||
else:
|
||||
logger.info_once("Using MarlinExperts for MXFP4 MoE")
|
||||
self.experts_cls = MarlinExperts
|
||||
|
||||
Reference in New Issue
Block a user