biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 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
3 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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(