[ROCm][CI] skip moe weight padding for eplb (#49350)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
This commit is contained in:
Divakar Verma
2026-07-22 10:18:01 -05:00
committed by GitHub
parent 37e370fe93
commit 53c2f20dd9
@@ -139,10 +139,13 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
def _maybe_pad_weight(self, weight: torch.Tensor) -> torch.Tensor:
# Pad the weight tensor. This is an optimization on ROCm platform, which
# can benefit from tensors located far enough from one another in memory
# can benefit from tensors located far enough from one another in memory.
# Skip padding when EPLB is enabled because EPLB requires contiguous
# weights for the view/rearrangement operations.
if (
envs.VLLM_ROCM_MOE_PADDING
and current_platform.is_rocm()
and not self.moe.moe_parallel_config.enable_eplb
and weight.stride(-1) == 1
and (weight.stride(-2) * weight.element_size()) % 512 == 0
):