[XPU] reudce host overhead of XPU MOE (#42915)

Signed-off-by: mayuyuace <qiming1.zhang@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Qiming Zhang
2026-05-23 13:09:34 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent a5bbd81e2e
commit 54d153637b
@@ -23,7 +23,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
from vllm.platforms import current_platform
if current_platform.is_xpu():
from vllm_xpu_kernels.fused_moe_interface import xpu_fused_moe
from vllm_xpu_kernels.fused_moe_interface import XpuFusedMoe
def prepare_fp8_moe_layer_for_xpu(
@@ -50,6 +50,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
self.is_fp8 = False
self.is_mxfp4 = False
self.is_mxfp8 = False
self.fused_moe_impl: XpuFusedMoe | None = None
@property
def expects_unquantized_inputs(self) -> bool:
@@ -132,26 +133,30 @@ class XPUExperts(mk.FusedMoEExpertsModular):
expert_tokens_meta: mk.ExpertTokensMetadata | None,
apply_router_weight_on_input: bool,
):
topk = topk_ids.size(-1)
xpu_fused_moe(
if self.fused_moe_impl is None:
topk = topk_ids.size(-1)
self.fused_moe_impl = XpuFusedMoe(
w13=w1,
w13_scales=self.w1_scale,
w13_bias=self.w1_bias,
w2=w2,
w2_scales=self.w2_scale,
w2_bias=self.w2_bias,
n_experts_per_token=topk,
activation=activation.value,
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_mxfp4=self.is_mxfp4,
is_mxfp8=self.is_mxfp8,
)
assert self.fused_moe_impl is not None
self.fused_moe_impl.apply(
output=output,
hidden_states=hidden_states,
w13=w1,
w13_scales=self.w1_scale,
w13_bias=self.w1_bias,
w2=w2,
w2_scales=self.w2_scale,
w2_bias=self.w2_bias,
topk_weights=topk_weights,
topk_ids=topk_ids,
n_experts_per_token=topk,
activation=activation.value,
num_experts=self.moe_config.num_local_experts,
ep_rank=self.moe_config.ep_rank,
ep_size=self.moe_config.ep_size,
output=output,
is_fp8=self.is_fp8,
is_mxfp4=self.is_mxfp4,
is_mxfp8=self.is_mxfp8,
)