From a4905133f375ec721be441e7ec4a3f923daa28f3 Mon Sep 17 00:00:00 2001
From: Hank_ <37239608+ILikeIneine@users.noreply.github.com>
Date: Wed, 22 Apr 2026 19:39:40 +0800
Subject: [PATCH] [xpu][rocm] Update `current_platform.supports_fp8()` for
 TritonExperts (#40132)

Signed-off-by: Hank <hcc.mayday@gmail.com>
---
 .../layers/fused_moe/fused_moe.py             | 19 +------------------
 vllm/platforms/rocm.py                        |  2 +-
 vllm/platforms/xpu.py                         |  4 ++++
 3 files changed, 6 insertions(+), 19 deletions(-)

diff --git a/vllm/model_executor/layers/fused_moe/fused_moe.py b/vllm/model_executor/layers/fused_moe/fused_moe.py
index 9218d0aff7..3b12f29493 100644
--- a/vllm/model_executor/layers/fused_moe/fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/fused_moe.py
@@ -1952,24 +1952,7 @@ class TritonExperts(mk.FusedMoEExpertsModular):
         weight_key: QuantKey | None,
         activation_key: QuantKey | None,
     ) -> bool:
-        p = current_platform
-        if p.is_rocm():
-            from vllm.platforms.rocm import on_gfx9, on_gfx12x
-
-            is_rocm_on_gfx9 = on_gfx9()
-            is_rocm_on_gfx12x = on_gfx12x()
-        else:
-            is_rocm_on_gfx9 = False
-            is_rocm_on_gfx12x = False
-
-        device_supports_fp8 = (
-            is_rocm_on_gfx9
-            or is_rocm_on_gfx12x
-            or (p.is_cuda() and p.has_device_capability((8, 9)))
-            or p.is_xpu()
-        )
-
-        if not device_supports_fp8:
+        if not current_platform.supports_fp8():
             return (weight_key, activation_key) == (None, None)
 
         SUPPORTED_W_A = [
diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
index 89714f00f6..0801c85242 100644
--- a/vllm/platforms/rocm.py
+++ b/vllm/platforms/rocm.py
@@ -800,7 +800,7 @@ class RocmPlatform(Platform):
 
     @classmethod
     def supports_fp8(cls) -> bool:
-        return any(gfx in _GCN_ARCH for gfx in ["gfx94", "gfx95", "gfx12"])
+        return on_gfx9() or on_gfx12x()
 
     @classmethod
     def is_fp8_fnuz(cls) -> bool:
diff --git a/vllm/platforms/xpu.py b/vllm/platforms/xpu.py
index aa67341973..d52ba23243 100644
--- a/vllm/platforms/xpu.py
+++ b/vllm/platforms/xpu.py
@@ -323,6 +323,10 @@ class XPUPlatform(Platform):
             )
         return "vllm.distributed.device_communicators.xpu_communicator.XpuCommunicator"  # noqa
 
+    @classmethod
+    def supports_fp8(cls) -> bool:
+        return True
+
     @classmethod
     def get_default_ir_op_priority(
         cls, vllm_config: "VllmConfig"