From 54c660c3a6c30c79e6457a162a19f4ed68042743 Mon Sep 17 00:00:00 2001 From: Yan Ma Date: Mon, 8 Jun 2026 13:58:16 +0800 Subject: [PATCH] [XPU][Minor] format moe kernel name and add in kernel list (#44771) Signed-off-by: Yan Ma --- vllm/model_executor/layers/fused_moe/__init__.py | 6 ++++-- vllm/model_executor/layers/fused_moe/experts/xpu_moe.py | 4 ++-- vllm/model_executor/layers/fused_moe/oracle/fp8.py | 4 ++-- vllm/model_executor/layers/fused_moe/oracle/mxfp4.py | 4 ++-- .../compressed_tensors_moe_w4a4_mxfp4.py | 6 +++--- 5 files changed, 13 insertions(+), 11 deletions(-) diff --git a/vllm/model_executor/layers/fused_moe/__init__.py b/vllm/model_executor/layers/fused_moe/__init__.py index 4f8627a97c..86eefbf7fc 100644 --- a/vllm/model_executor/layers/fused_moe/__init__.py +++ b/vllm/model_executor/layers/fused_moe/__init__.py @@ -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 diff --git a/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py b/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py index 9420832646..00829a0f70 100644 --- a/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py @@ -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, diff --git a/vllm/model_executor/layers/fused_moe/oracle/fp8.py b/vllm/model_executor/layers/fused_moe/oracle/fp8.py index cce3245b75..2477973268 100644 --- a/vllm/model_executor/layers/fused_moe/oracle/fp8.py +++ b/vllm/model_executor/layers/fused_moe/oracle/fp8.py @@ -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 ( diff --git a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py index a506eaffd0..8759fc8e60 100644 --- a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py +++ b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py @@ -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 diff --git a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py index a960af7d5f..906b0727b1 100644 --- a/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py +++ b/vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_w4a4_mxfp4.py @@ -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