Fix DeepSpeedInferenceConfig crash on bool moe backward-compat value

deepspeed/inference/config.py

@@ -313,7 +313,7 @@ def validate_dtype(cls, field_value, values):
     @field_validator("moe")
     def moe_backward_compat(cls, field_value, values):
         if isinstance(field_value, bool):
-            return DeepSpeedMoEConfig(moe=field_value)
+            return DeepSpeedMoEConfig(enabled=field_value)
         return field_value
 
     @field_validator("use_triton")

tests/unit/inference/test_inference_config.py

@@ -42,3 +42,13 @@ def test_json_config(self, tmpdir):
 
         engine = deepspeed.init_inference(torch.nn.Module(), config=config_json)
         assert engine._config.replace_with_kernel_inject
+
+    def test_moe_backward_compat_bool(self):
+        # `moe` accepts a bool for backward compatibility (moe: Union[bool, DeepSpeedMoEConfig]);
+        # it should build a DeepSpeedMoEConfig rather than raising a validation error.
+        from deepspeed.inference.config import DeepSpeedInferenceConfig, DeepSpeedMoEConfig
+
+        for value in (True, False):
+            config = DeepSpeedInferenceConfig(moe=value)
+            assert isinstance(config.moe, DeepSpeedMoEConfig)
+            assert config.moe.enabled == value