Fail fast when GRPO combines Liger with the MoE auxiliary loss

tests/test_grpo_trainer.py

@@ -1006,6 +1006,25 @@ def test_liger_kernel_with_peft_prompt_learning_raises(self):
                 peft_config=PromptTuningConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=8),
             )
 
+    @require_liger_kernel
+    def test_init_fails_with_moe_aux_loss_and_liger(self):
+        dataset = load_dataset("trl-internal-testing/zen", "standard_prompt_only", split="train")
+
+        # The MoE auxiliary loss is on by default; it is incompatible with the Liger fused loss.
+        training_args = GRPOConfig(
+            output_dir=self.tmp_dir,
+            use_liger_kernel=True,
+            report_to="none",
+        )
+
+        with pytest.raises(ValueError, match="does not support the Mixture-of-Experts load-balancing auxiliary loss"):
+            GRPOTrainer(
+                model="trl-internal-testing/tiny-Qwen3MoeForCausalLM",
+                reward_funcs="trl-internal-testing/tiny-Qwen2ForSequenceClassification-2.5",
+                args=training_args,
+                train_dataset=dataset,
+            )
+
     @require_peft
     def test_train_peft_model(self):
         model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-Qwen2ForCausalLM-2.5", dtype="float32")

trl/trainer/grpo_trainer.py

@@ -792,6 +792,12 @@ def get_reward(environments, _env_type=env_type, **kwargs):
         is_moe = getattr(text_config, "output_router_logits", None) is not None
         self.aux_loss_enabled = is_moe and args.router_aux_loss_coef != 0.0
         self.router_aux_loss_coef = args.router_aux_loss_coef
+        if self.aux_loss_enabled and self.use_liger_kernel:
+            raise ValueError(
+                "Liger GRPO loss does not support the Mixture-of-Experts load-balancing auxiliary loss, because it "
+                "fuses the loss without materializing the router logits. Either set `router_aux_loss_coef` to `0.0` "
+                "to disable the auxiliary loss, or set `use_liger_kernel` to False."
+            )
         self.scale_rewards = args.scale_rewards
         self.importance_sampling_level = args.importance_sampling_level
         self.off_policy_mask_threshold = args.off_policy_mask_threshold