[NPU][Fix Issue]: Send expert weights contiguous tensor across cards during EPLB rebalance (#32001)

2026-07-27 09:20:28 +08:00
parent 3863612023
commit a358374ae9
1 changed files with 1 additions and 1 deletions
@@ -332,7 +332,7 @@ class NPUW8A8Int8MoEMethod(_NPUMoEMethodBase):

        # Process weight
        weight: torch.Tensor = getattr(layer, f"{weight_prefix}_weight")
-        weight.data = npu_format_cast(weight.data.transpose(1, 2))
+        weight.data = npu_format_cast(weight.data.transpose(1, 2).contiguous())

        # Set dispatcher output dtype
        if weight_prefix == "w13":