[xpu][rocm] Update current_platform.supports_fp8() for TritonExperts (#40132)

Signed-off-by: Hank <hcc.mayday@gmail.com>

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 = [

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:

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"