Llama4TextMoe: mark as not having backward

torch-ext/moe/layers.py

@@ -25,6 +25,8 @@ def _fix_llama4_experts(hidden_states: torch.Tensor, experts: nn.Module):
 
 
 class Llama4TextMoe(nn.Module):
+    has_backward = False
+
     experts: nn.Module
     router: nn.Linear
     shared_expert: nn.Module