diff --git a/doc/prediction.rst b/doc/prediction.rst
index 853254bba..39ca3f739 100644
--- a/doc/prediction.rst
+++ b/doc/prediction.rst
@@ -54,8 +54,8 @@ After 1.4 release, we added a new parameter called ``strict_shape``, one can set
   Output is a 4-dim array with ``(n_samples, n_iterations, n_classes, n_trees_in_forest)``
   as shape.  ``n_trees_in_forest`` is specified by the ``numb_parallel_tree`` during
   training.  When strict shape is set to False, output is a 2-dim array with last 3 dims
-  concatenated into 1.  When using ``apply`` method in scikit learn interface, this is set
-  to False by default.
+  concatenated into 1.  Also the last dimension is dropped if it eqauls to 1. When using
+  ``apply`` method in scikit learn interface, this is set to False by default.
 
 
 Other than these prediction types, there's also a parameter called ``iteration_range``,
diff --git a/src/c_api/c_api_utils.h b/src/c_api/c_api_utils.h
index ecf5b1f34..918830945 100644
--- a/src/c_api/c_api_utils.h
+++ b/src/c_api/c_api_utils.h
@@ -96,6 +96,10 @@ inline void CalcPredictShape(bool strict_shape, PredictionType type, size_t rows
       forest = std::max(static_cast<decltype(forest)>(1), forest);
       shape[3] = forest;
       *out_dim = shape.size();
+    } else if (chunksize == 1) {
+      *out_dim = 1;
+      shape.resize(*out_dim);
+      shape.front() = rows;
     } else {
       *out_dim = 2;
       shape.resize(*out_dim);
diff --git a/tests/python/test_predict.py b/tests/python/test_predict.py
index 7562fc609..d451cd831 100644
--- a/tests/python/test_predict.py
+++ b/tests/python/test_predict.py
@@ -75,6 +75,11 @@ def run_predict_leaf(predictor):
     first = sliced[0, ...]
 
     assert np.prod(first.shape) == classes * num_parallel_tree * ntree_limit
+
+    # When there's only 1 tree, the output is a 1 dim vector
+    booster = xgb.train({"tree_method": "hist"}, num_boost_round=1, dtrain=m)
+    assert booster.predict(m, pred_leaf=True).shape == (rows, )
+
     return leaf