[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>
Oxana Korzh committed
Jul 3, 2026 at 13:41 UTC
576bf75d0e79d6cca2440138fedae68a73089073
1 file changed
+5
vllm/model_executor/layers/quantization/quark/quark_moe.py
+5
@@ -1322,6 +1322,11 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
1322
gemm1_clamp_limit=getattr(layer, "swiglu_limit", None),
1323
)
1324
1325
+ @property
1326
+ def supports_eplb(self) -> bool:
1327
+ # AITER shuffle keeps expert dim outermost, so EPLB row moves are layout-safe.
1328
+ return True
1329
+
1330
@property
1331
def is_monolithic(self) -> bool:
1332
if self.moe_kernel is not None: