fix (#6098)

Co-authored-by: tastelikefeet <[email protected]>

Author: tastelikefeet

swift/trainers/sequence_parallel/ulysses.py

@@ -354,7 +354,7 @@ def moe_aux_loss_hook(module, args, kwargs, output):
             if isinstance(router_logits, tuple):
                 compute_device = router_logits[0].device
                 router_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in router_logits], dim=0)
-            router_logits, _ = GatherLoss.apply(router_logits, None, self.sp_group)
+            router_logits, _ = GatherLoss.apply(router_logits, None)
             router_logits = router_logits.reshape(self.sp_world_size, num_layers, sp_len,
                                                   -1).transpose(0, 1).reshape(num_layers, self.sp_world_size * sp_len,
                                                                               -1)