diff --git a/grail/infrastructure/delta_checkpoint.py b/grail/infrastructure/delta_checkpoint.py
index 741bd52..42226bb 100644
--- a/grail/infrastructure/delta_checkpoint.py
+++ b/grail/infrastructure/delta_checkpoint.py
@@ -134,6 +134,8 @@ def apply_sparse_delta(
 
     # Infer dtype from base_state if not specified
     if target_dtype is None:
+        if not base_state:
+            raise ValueError("Cannot infer target_dtype: base_state is empty")
         target_dtype = next(iter(base_state.values())).dtype
 
     for name, base_tensor in base_state.items():
diff --git a/tests/unit/infrastructure/test_delta_checkpoint.py b/tests/unit/infrastructure/test_delta_checkpoint.py
index bdbb3bf..c43d561 100644
--- a/tests/unit/infrastructure/test_delta_checkpoint.py
+++ b/tests/unit/infrastructure/test_delta_checkpoint.py
@@ -202,6 +202,16 @@ def test_dtype_conversion(self) -> None:
 
         assert result["weight"].dtype == torch.bfloat16
 
+    def test_empty_base_state_raises_value_error(self) -> None:
+        """Test that empty base_state with target_dtype=None raises ValueError instead of StopIteration."""
+        with pytest.raises(ValueError, match="Cannot infer target_dtype: base_state is empty"):
+            apply_sparse_delta({}, {}, {}, target_dtype=None)
+
+    def test_empty_base_state_with_explicit_dtype(self) -> None:
+        """Test that empty base_state with explicit target_dtype returns empty dict."""
+        result = apply_sparse_delta({}, {}, {}, target_dtype=torch.float32)
+        assert result == {}
+
 
 class TestRoundTrip:
     """Test round-trip: compute delta, apply delta, verify identical."""