Update modeling_phi.py

fix dtype for inference

modeling_phi.py

@@ -312,7 +312,7 @@ class MoE(nn.Module):
         x = x.repeat_interleave(self.num_experts_per_tok, dim=0)
         y = torch.empty_like(x)
         for i, expert in enumerate(self.mlp):
-            y[flat_expert_indices == i] = expert(x[flat_expert_indices == i])
+            y[flat_expert_indices == i] = expert(x[flat_expert_indices == i]).to(dtype=y.dtype)
         y = (y.view(*expert_weights.shape, -1) * expert_weights.unsqueeze(-1)).sum(dim=1)
         return y.view(*orig_shape)