From 576bf75d0e79d6cca2440138fedae68a73089073 Mon Sep 17 00:00:00 2001
From: Oxana Korzh <okorzh@amd.com>
Date: Fri, 3 Jul 2026 13:41:52 -0600
Subject: [PATCH] [AMD][EPLB] Enable EPLB for Quark OCP MXFP4 MoE (#47220)

Signed-off-by: okorzh-amd <okorzh-amd@users.noreply.github.com>
Co-authored-by: okorzh-amd <okorzh-amd@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
---
 vllm/model_executor/layers/quantization/quark/quark_moe.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/vllm/model_executor/layers/quantization/quark/quark_moe.py b/vllm/model_executor/layers/quantization/quark/quark_moe.py
index ebe37d8dc5..bce888415c 100644
--- a/vllm/model_executor/layers/quantization/quark/quark_moe.py
+++ b/vllm/model_executor/layers/quantization/quark/quark_moe.py
@@ -1322,6 +1322,11 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
                 gemm1_clamp_limit=getattr(layer, "swiglu_limit", None),
             )
 
+    @property
+    def supports_eplb(self) -> bool:
+        # AITER shuffle keeps expert dim outermost, so EPLB row moves are layout-safe.
+        return True
+
     @property
     def is_monolithic(self) -> bool:
         if self.moe_kernel is not None: