[misc] fix: validate DataProtoItem consistency

tests/test_protocol_on_cpu.py

@@ -23,6 +23,7 @@
 
 from verl import DataProto
 from verl.protocol import (
+    DataProtoItem,
     deserialize_single_tensor,
     deserialize_tensordict,
     serialize_single_tensor,
@@ -152,6 +153,39 @@ def test_tensor_dict_constructor():
         data = DataProto.from_dict(tensors={"obs": obs, "act": act}, num_batch_dims=3)
 
 
+def test_dataproto_item_consistency():
+    data = DataProto.from_dict(
+        tensors={"obs": torch.randn(2, 3)},
+        non_tensors={"label": ["a", "b"]},
+        meta_info={"split": "train"},
+    )
+
+    item = data[0]
+
+    assert item.batch.batch_size == torch.Size([])
+    assert item.non_tensor_batch == {"label": "a"}
+    assert item.meta_info == {"split": "train"}
+
+
+def test_dataproto_item_rejects_batched_tensordict():
+    batch = TensorDict({"obs": torch.randn(1, 3)}, batch_size=[1])
+
+    with pytest.raises(AssertionError, match="must represent a single item"):
+        DataProtoItem(batch=batch)
+
+
+@pytest.mark.parametrize(
+    ("kwargs", "message"),
+    [
+        ({"non_tensor_batch": []}, "non_tensor_batch must be a dict"),
+        ({"meta_info": []}, "meta_info must be a dict"),
+    ],
+)
+def test_dataproto_item_rejects_non_dict_metadata(kwargs, message):
+    with pytest.raises(AssertionError, match=message):
+        DataProtoItem(**kwargs)
+
+
 def test_tensor_dict_make_iterator():
     obs = torch.randn(100, 10)
     labels = [random.choice(["abc", "cde"]) for _ in range(100)]

verl/protocol.py

@@ -308,11 +308,26 @@ def collate_fn(x: list["DataProtoItem"]):
 
 @dataclass
 class DataProtoItem:
-    # TODO(zhangchi.usc1992) add consistency check
     batch: TensorDict = None
     non_tensor_batch: dict = field(default_factory=dict)
     meta_info: dict = field(default_factory=dict)
 
+    def __post_init__(self):
+        self.check_consistency()
+
+    def check_consistency(self):
+        """Validate the unbatched item contract used by ``DataProto.__getitem__``."""
+
+        if self.batch is not None:
+            assert len(self.batch.batch_size) == 0, (
+                "DataProtoItem.batch must represent a single item with num_batch_dims=0, "
+                f"got batch_size={self.batch.batch_size}"
+            )
+        assert isinstance(self.non_tensor_batch, dict), (
+            f"DataProtoItem.non_tensor_batch must be a dict, got {type(self.non_tensor_batch)}"
+        )
+        assert isinstance(self.meta_info, dict), f"DataProtoItem.meta_info must be a dict, got {type(self.meta_info)}"
+
 
 @dataclass
 class DataProto: