xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 04:54:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

[XPU] Enable Expert parallel for MoE models (#28263)

Browse Source 
Signed-off-by: Yan Ma <yan.ma@intel.com>
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Kunshang Ji 2025-11-08 08:33:11 +08:00 committed by GitHub
parent 811df41ee9
commit 1aaecda078
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 6 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/layers/fused_moe/layer.py
View File

@ -642,10 +642,12 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
if current_platform.is_xpu():
import intel_extension_for_pytorch as ipex
ep_rank_start = self.moe.ep_rank * self.moe.num_local_experts
layer.ipex_fusion = ipex.llm.modules.GatedMLPMOE(
layer.w13_weight,
layer.w2_weight,
use_prepack=True,
experts_start_id=ep_rank_start,
)
elif current_platform.is_cpu():
from vllm.model_executor.layers.fused_moe import cpu_fused_moe

2
vllm/model_executor/layers/quantization/ipex_quant.py
View File

@ -399,6 +399,7 @@ class XPUFp8MoEMethod(FusedMoEMethodBase):
layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
import intel_extension_for_pytorch as ipex
ep_rank_start = self.moe.ep_rank * self.moe.num_local_experts
layer.ipex_fusion = ipex.llm.modules.GatedMLPMOE(
layer.w13_weight,
layer.w2_weight,
@ -407,6 +408,7 @@ class XPUFp8MoEMethod(FusedMoEMethodBase):
a1_scale_inv=layer.w13_input_scale,
a2_scale_inv=layer.w2_input_scale,
use_prepack=True,
experts_start_id=ep_rank_start,
)
def get_fused_moe_quant_config(

2
vllm/model_executor/layers/quantization/mxfp4.py
View File

@ -1113,6 +1113,7 @@ class IpexMxfp4MoEMethod(Mxfp4MoEMethod):
layer.w13_weight.data = layer.w13_weight.data.view(torch.int32)
layer.w2_weight.data = layer.w2_weight.data.view(torch.int32)
ep_rank_start = self.moe_config.ep_rank * self.moe_config.num_local_experts
layer.ipex_fusion = ipex.llm.modules.GatedMLPMOE(
layer.w13_weight,
layer.w2_weight,
@ -1121,6 +1122,7 @@ class IpexMxfp4MoEMethod(Mxfp4MoEMethod):
w13_bias=layer.w13_bias,
w2_bias=layer.w2_bias,
is_mxfp4=True,
experts_start_id=ep_rank_start,
)
def apply(