diff --git a/funasr/bin/train_ds.py b/funasr/bin/train_ds.py
index b32881d99..d1f25973e 100644
--- a/funasr/bin/train_ds.py
+++ b/funasr/bin/train_ds.py
@@ -40,6 +40,37 @@ try:
 except:
     deepspeed = None
 
+_DISTRIBUTED_TRAIN_CONF_KEYS = (
+    "use_ddp",
+    "use_fsdp",
+    "use_deepspeed",
+    "deepspeed_config",
+)
+
+
+def _resolve_distributed_config(kwargs, world_size):
+    """Resolve distributed settings with top-level values taking precedence."""
+    train_conf = dict(kwargs.get("train_conf") or {})
+
+    def get_setting(name, default):
+        if name in kwargs:
+            return kwargs[name]
+        return train_conf.get(name, default)
+
+    use_fsdp = get_setting("use_fsdp", False)
+    use_deepspeed = get_setting("use_deepspeed", False)
+    deepspeed_config = get_setting("deepspeed_config", "")
+    if use_deepspeed and use_fsdp:
+        raise ValueError("use_deepspeed and use_fsdp cannot be enabled at the same time")
+
+    trainer_conf = {
+        key: value
+        for key, value in train_conf.items()
+        if key not in _DISTRIBUTED_TRAIN_CONF_KEYS
+    }
+    use_ddp = world_size > 1 and not use_deepspeed and not use_fsdp
+    return use_ddp, use_fsdp, use_deepspeed, deepspeed_config, trainer_conf
+
 
 @hydra.main(config_name=None, version_base=None)
 def main_hydra(kwargs: DictConfig):
@@ -83,9 +114,13 @@ def main(**kwargs):
     if local_rank == 0:
         tables.print()
 
-    use_ddp = world_size > 1
-    use_fsdp = kwargs.get("use_fsdp", False)
-    use_deepspeed = kwargs.get("use_deepspeed", False)
+    (
+        use_ddp,
+        use_fsdp,
+        use_deepspeed,
+        deepspeed_config,
+        trainer_conf,
+    ) = _resolve_distributed_config(kwargs, world_size)
     if use_deepspeed:
         logging.info(f"use_deepspeed: {use_deepspeed}")
         deepspeed.init_distributed(dist_backend=kwargs.get("backend", "nccl"))
@@ -143,10 +178,12 @@ def main(**kwargs):
         world_size=world_size,
         use_ddp=use_ddp,
         use_fsdp=use_fsdp,
+        use_deepspeed=use_deepspeed,
+        deepspeed_config=deepspeed_config,
         device=kwargs["device"],
         excludes=kwargs.get("excludes", None),
         output_dir=kwargs.get("output_dir", "./exp"),
-        **kwargs.get("train_conf"),
+        **trainer_conf,
     )
 
     model = trainer.warp_model(model, **kwargs)
diff --git a/tests/test_train_ds_distributed_config.py b/tests/test_train_ds_distributed_config.py
new file mode 100644
index 000000000..20537d160
--- /dev/null
+++ b/tests/test_train_ds_distributed_config.py
@@ -0,0 +1,62 @@
+import unittest
+
+from funasr.bin.train_ds import _resolve_distributed_config
+
+
+class TestDistributedConfig(unittest.TestCase):
+    def test_nested_train_conf_is_supported(self):
+        use_ddp, use_fsdp, use_deepspeed, config, trainer_conf = (
+            _resolve_distributed_config(
+                {
+                    "train_conf": {
+                        "use_deepspeed": True,
+                        "deepspeed_config": "nested_ds.json",
+                        "log_interval": 10,
+                    }
+                },
+                world_size=2,
+            )
+        )
+
+        self.assertFalse(use_ddp)
+        self.assertFalse(use_fsdp)
+        self.assertTrue(use_deepspeed)
+        self.assertEqual(config, "nested_ds.json")
+        self.assertEqual(trainer_conf, {"log_interval": 10})
+
+    def test_top_level_values_override_train_conf(self):
+        use_ddp, use_fsdp, use_deepspeed, config, trainer_conf = (
+            _resolve_distributed_config(
+                {
+                    "use_deepspeed": False,
+                    "deepspeed_config": "top_level_ds.json",
+                    "train_conf": {
+                        "use_deepspeed": True,
+                        "deepspeed_config": "nested_ds.json",
+                        "use_ddp": True,
+                        "log_interval": 10,
+                    },
+                },
+                world_size=2,
+            )
+        )
+
+        self.assertTrue(use_ddp)
+        self.assertFalse(use_fsdp)
+        self.assertFalse(use_deepspeed)
+        self.assertEqual(config, "top_level_ds.json")
+        self.assertEqual(trainer_conf, {"log_interval": 10})
+
+    def test_deepspeed_and_fsdp_are_mutually_exclusive(self):
+        with self.assertRaisesRegex(ValueError, "cannot be enabled"):
+            _resolve_distributed_config(
+                {
+                    "use_deepspeed": True,
+                    "train_conf": {"use_fsdp": True},
+                },
+                world_size=2,
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()