make new parameter

Signed-off-by: Varun Sundar Rabindranath <[email protected]>

Author: Varun Sundar Rabindranath

vllm/model_executor/layers/fused_moe/layer.py

@@ -1249,8 +1249,12 @@ def _maybe_make_contiguous(
                 # do nothing.
                 return p
             assert p.ndim == 3
-            p.data = torch.transpose(p.data, 1, 2)
-            return p
+            # Do not update the layer paramater as the layer's MoE operations would
+            # expect the parameter's tensor to the same shape / stride. Instead,
+            # make a new torch.nn.Parameter that is used just in the context of
+            # EPLB.
+            return torch.nn.Parameter(torch.transpose(p.data, 1, 2),
+                                      requires_grad=False)
 
         weights = list(self.named_parameters())
         weights = [(name, _maybe_make_contiguous(name, p)) for name, p in weights]