From 063ce98fb7104dba93f72d56c094fb8b708bd793 Mon Sep 17 00:00:00 2001
From: zofia <110436990+zufangzhu@users.noreply.github.com>
Date: Fri, 5 Jun 2026 08:36:58 +0800
Subject: [PATCH] [XPU][MoE] support block_fp8_moe on xpu (#42139)

Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
Signed-off-by: zofia <110436990+zufangzhu@users.noreply.github.com>
---
 .buildkite/intel_jobs/test-intel.yaml         |  3 +-
 .../layers/fused_moe/experts/xpu_moe.py       | 31 +++++++++++++++++++
 .../layers/fused_moe/oracle/fp8.py            |  3 +-
 3 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/.buildkite/intel_jobs/test-intel.yaml b/.buildkite/intel_jobs/test-intel.yaml
index 805b7e54f12..63ce93c4810 100644
--- a/.buildkite/intel_jobs/test-intel.yaml
+++ b/.buildkite/intel_jobs/test-intel.yaml
@@ -40,7 +40,8 @@ steps:
         python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192 &&
         python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 &&
         python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel &&
-        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192
+        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192 &&
+        VLLM_XPU_FUSED_MOE_USE_REF=1 python3 examples/basic/offline_inference/generate.py --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 --enforce-eager -tp 2
         '
   - label: "XPU V1 test"
     depends_on:
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 82969dd8e25..94208326461 100644
--- a/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
+++ b/vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
@@ -14,7 +14,9 @@ from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
 )
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     QuantKey,
+    kFp8Dynamic128Sym,
     kFp8DynamicTensorSym,
+    kFp8Static128BlockSym,
     kFp8StaticTensorSym,
     kInt4Static,
     kMxfp4Static,
@@ -62,6 +64,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
         self.is_fp8 = False
         self.is_int4 = False
         self.is_mxfp4 = False
+        self.is_block_fp8 = False
         self.is_mxfp8 = False
         self.fused_moe_impl: XpuFusedMoe | None = None
 
@@ -171,6 +174,7 @@ class XPUExperts(mk.FusedMoEExpertsModular):
                 is_int4=self.is_int4,
                 is_mxfp4=self.is_mxfp4,
                 is_mxfp8=self.is_mxfp8,
+                is_block_fp8=self.is_block_fp8,
             )
         assert self.fused_moe_impl is not None
         self.fused_moe_impl.apply(
@@ -238,6 +242,33 @@ class XPUExpertsMxfp8(XPUExpertsFp8):
         return (weight_key, activation_key) in SUPPORTED_W_A
 
 
+class XPUExpertsBlockFp8(XPUExperts):
+    def __init__(
+        self,
+        moe_config: FusedMoEConfig,
+        quant_config: FusedMoEQuantConfig,
+        max_num_tokens: int | None = None,
+        num_dispatchers: int | None = None,
+    ):
+        super().__init__(
+            moe_config,
+            quant_config,
+            max_num_tokens,
+            num_dispatchers,
+        )
+        self.is_block_fp8 = True
+
+    @staticmethod
+    def _supports_quant_scheme(
+        weight_key: QuantKey | None,
+        activation_key: QuantKey | None,
+    ) -> bool:
+        SUPPORTED_W_A = [
+            (kFp8Static128BlockSym, kFp8Dynamic128Sym),
+        ]
+        return (weight_key, activation_key) in SUPPORTED_W_A
+
+
 class XPUExpertsWNA16(XPUExperts):
     """W4A16 INT4-symmetric MoE backed by `xpu_fused_moe(is_int4=True)`.
 
diff --git a/vllm/model_executor/layers/fused_moe/oracle/fp8.py b/vllm/model_executor/layers/fused_moe/oracle/fp8.py
index 0a2e3846dd9..cce3245b75c 100644
--- a/vllm/model_executor/layers/fused_moe/oracle/fp8.py
+++ b/vllm/model_executor/layers/fused_moe/oracle/fp8.py
@@ -184,11 +184,12 @@ def backend_to_kernel_cls(
 
     elif backend == Fp8MoeBackend.XPU:
         from vllm.model_executor.layers.fused_moe.experts.xpu_moe import (
+            XPUExpertsBlockFp8,
             XPUExpertsFp8,
             XPUExpertsMxfp8,
         )
 
-        return [XPUExpertsFp8, XPUExpertsMxfp8]
+        return [XPUExpertsFp8, XPUExpertsMxfp8, XPUExpertsBlockFp8]
 
     elif backend == Fp8MoeBackend.CPU:
         from vllm.model_executor.layers.fused_moe.experts.cpu_moe import (