From 0b6636cbcbd3d40cb8c36459a3ff0f1f3c9b48a7 Mon Sep 17 00:00:00 2001 From: Qiming Zhang Date: Sat, 11 Jul 2026 02:27:13 -0700 Subject: [PATCH] [XPU]remove is_xxx from moe class and bump up kernels (#48079) Signed-off-by: mayuyuace Co-authored-by: Kunshang Ji --- requirements/xpu.txt | 2 +- vllm/_custom_ops.py | 12 -------- .../layers/fused_moe/experts/xpu_moe.py | 30 ++++--------------- 3 files changed, 7 insertions(+), 37 deletions(-) diff --git a/requirements/xpu.txt b/requirements/xpu.txt index 68b8eb1301..2da3c6654b 100644 --- a/requirements/xpu.txt +++ b/requirements/xpu.txt @@ -18,4 +18,4 @@ torchvision torchcodec >= 0.14 # Required for the torchcodec video decoding backend auto_round_lib>=0.14.0 -vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.10.1/vllm_xpu_kernels-0.1.10.1-cp38-abi3-manylinux_2_28_x86_64.whl +vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.11/vllm_xpu_kernels-0.1.11-cp38-abi3-manylinux_2_28_x86_64.whl diff --git a/vllm/_custom_ops.py b/vllm/_custom_ops.py index 2a4579ae03..627058cb2e 100644 --- a/vllm/_custom_ops.py +++ b/vllm/_custom_ops.py @@ -2277,18 +2277,6 @@ def topk_sigmoid( e_score_correction_bias: torch.Tensor | None = None, routed_scaling_factor: float = 1.0, ) -> None: - if current_platform.is_xpu(): - # xpu doesn't support routed_scaling_factor currently, will revert - # in next vllm-xpu-kernels bumpup - torch.ops._moe_C.topk_sigmoid( - topk_weights, - topk_ids, - token_expert_indices, - gating_output, - renormalize, - e_score_correction_bias, - ) - return torch.ops._moe_C.topk_sigmoid( topk_weights, topk_ids, 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 e8b29bcf2c..cde167e5d3 100644 --- a/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py @@ -62,11 +62,6 @@ class XPUExperts(mk.FusedMoEExpertsModular): max_num_tokens, num_dispatchers, ) - self.is_fp8 = False - self.is_int4 = False - self.is_mxfp4 = False - self.is_block_fp8 = False - self.is_mxfp8 = False self.gemm1_clamp_limit = quant_config.gemm1_clamp_limit self.fused_moe_impl: XpuFusedMoe | None = None @@ -149,17 +144,14 @@ class XPUExperts(mk.FusedMoEExpertsModular): expert_tokens_meta: mk.ExpertTokensMetadata | None, apply_router_weight_on_input: bool, ): - # The kernel takes is_fp8/is_int4/is_mxfp4 as independent booleans. - # In this hierarchy each subclass flips exactly one to True; assert - # the invariant so a future subclass that sets two doesn't silently - # miscompute (kernel-side priority is undocumented). - assert sum([self.is_fp8, self.is_int4, self.is_mxfp4]) <= 1, ( - "XPUExperts: at most one of is_fp8, is_int4, is_mxfp4 may be True; " - f"got is_fp8={self.is_fp8}, is_int4={self.is_int4}, " - f"is_mxfp4={self.is_mxfp4}." - ) if self.fused_moe_impl is None: topk = topk_ids.size(-1) + if ( + self.quant_config is not None + and self.quant_config.weight_quant_dtype == "mxfp4" + ): + w1 = w1.view(torch.float4_e2m1fn_x2) + w2 = w2.view(torch.float4_e2m1fn_x2) self.fused_moe_impl = XpuFusedMoe( w13=w1, w13_scales=self.w1_scale, @@ -172,11 +164,6 @@ class XPUExperts(mk.FusedMoEExpertsModular): num_experts=self.moe_config.num_local_experts, ep_rank=self.moe_config.ep_rank, ep_size=self.moe_config.ep_size, - is_fp8=self.is_fp8, - is_int4=self.is_int4, - is_mxfp4=self.is_mxfp4, - is_mxfp8=self.is_mxfp8, - is_block_fp8=self.is_block_fp8, gemm1_clamp_limit=self.gemm1_clamp_limit, ) assert self.fused_moe_impl is not None @@ -202,7 +189,6 @@ class XPUExpertsFp8(XPUExperts): max_num_tokens, num_dispatchers, ) - self.is_fp8 = True @staticmethod def _supports_quant_scheme( @@ -231,7 +217,6 @@ class XPUExpertsMxFp8(XPUExpertsFp8): num_dispatchers, ) assert quant_config.quant_dtype == "mxfp8" - self.is_mxfp8 = True @staticmethod def _supports_quant_scheme( @@ -259,7 +244,6 @@ class XPUExpertsBlockFp8(XPUExperts): max_num_tokens, num_dispatchers, ) - self.is_block_fp8 = True @staticmethod def _supports_quant_scheme( @@ -298,7 +282,6 @@ class XPUExpertsWNA16(XPUExperts): max_num_tokens, num_dispatchers, ) - self.is_int4 = True @staticmethod def _supports_quant_scheme( @@ -325,7 +308,6 @@ class XPUExpertsMxFp4(XPUExperts): max_num_tokens, num_dispatchers, ) - self.is_mxfp4 = True @staticmethod def _supports_quant_scheme(