[ROCm][CI] skip moe weight padding for eplb (#49350)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Jul 22, 2026 at 11:18 UTC
53c2f20dd9f1ead9c4a086e5103af7f8a74681a2
1 file changed
+4
-1
vllm/model_executor/layers/fused_moe/unquantized_fused_moe_method.py
+4
-1
@@ -139,10 +139,13 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
139
140
def _maybe_pad_weight(self, weight: torch.Tensor) -> torch.Tensor:
141
# Pad the weight tensor. This is an optimization on ROCm platform, which
142
- # can benefit from tensors located far enough from one another in memory
142
+ # can benefit from tensors located far enough from one another in memory.
143
+ # Skip padding when EPLB is enabled because EPLB requires contiguous
144
+ # weights for the view/rearrangement operations.
145
if (
146
envs.VLLM_ROCM_MOE_PADDING
147
and current_platform.is_rocm()
148
+ and not self.moe.moe_parallel_config.enable_eplb
149
and weight.stride(-1) == 1
150
and (weight.stride(-2) * weight.element_size()) % 512 == 0
151
):