[BUG Fix] Fix cumsum dtype bug (PaddlePaddle#74830)

Author: DrRyanHuang

python/paddle/tensor/math.py

@@ -4230,15 +4230,19 @@ def cumsum(
         flatten = True
     else:
         flatten = False
-    if dtype is not None and x.dtype != convert_np_dtype_to_dtype_(dtype):
-        x = cast(x, dtype)
-    elif isinstance(x, paddle.Tensor) and x.dtype in [
-        paddle.uint8,
-        paddle.int8,
-        paddle.int16,
-        paddle.int32,
-    ]:
-        x = cast(x, "int64")
+
+    if dtype is None:
+        if x.dtype in [
+            paddle.uint8,
+            paddle.int8,
+            paddle.int16,
+            paddle.int32,
+        ]:
+            x = cast(x, "int64")
+    else:
+        dtype = convert_np_dtype_to_dtype_(dtype)
+        if x.dtype != dtype:
+            x = cast(x, dtype)
 
     if in_dynamic_or_pir_mode():
         if axis is None:

test/legacy_test/test_cumsum_op.py

@@ -28,6 +28,7 @@
 import paddle.inference as paddle_infer
 from paddle import base
 from paddle.base import core
+from paddle.framework import convert_np_dtype_to_dtype_
 
 
 class TestCumsumOp(unittest.TestCase):
@@ -273,6 +274,13 @@ def run_cases(self):
         z = np.cumsum(data_np, axis=-2)
         np.testing.assert_array_equal(z, y.numpy())
 
+        # test data type
+        data_np = np.arange(12).reshape(3, 4).astype(np.int16)
+        data = paddle.to_tensor(data_np)
+        y = paddle.cumsum(data, axis=0, dtype='int32')
+        z = np.cumsum(data_np, axis=0, dtype="int32")
+        np.testing.assert_equal(convert_np_dtype_to_dtype_(z.dtype), y.dtype)
+
     def run_static_uint8(self, use_gpu=False):
         with paddle.static.program_guard(paddle.static.Program()):
             data_np = np.random.random((100, 100)).astype(np.uint8)
@@ -281,6 +289,7 @@ def run_static_uint8(self, use_gpu=False):
             y2 = paddle.cumsum(x, axis=0)
             y3 = paddle.cumsum(x, axis=-1)
             y4 = paddle.cumsum(x, axis=-2)
+            y5 = paddle.cumsum(x, axis=-1, dtype='int32')
             place = base.CUDAPlace(0) if use_gpu else base.CPUPlace()
             exe = base.Executor(place)
             exe.run(paddle.static.default_startup_program())
@@ -291,6 +300,7 @@ def run_static_uint8(self, use_gpu=False):
                     y2,
                     y3,
                     y4,
+                    y5,
                 ],
             )
             z = np.cumsum(data_np)
@@ -301,6 +311,8 @@ def run_static_uint8(self, use_gpu=False):
             np.testing.assert_allclose(z, out[2], rtol=1e-05)
             z = np.cumsum(data_np, axis=-2)
             np.testing.assert_allclose(z, out[3], rtol=1e-05)
+            z = np.cumsum(data_np, axis=-1, dtype="int32")
+            np.testing.assert_equal(z.dtype, out[4].dtype)
 
     def run_static_int8(self, use_gpu=False):
         with paddle.static.program_guard(paddle.static.Program()):
@@ -310,7 +322,7 @@ def run_static_int8(self, use_gpu=False):
             y2 = paddle.cumsum(x, axis=0)
             y3 = paddle.cumsum(x, axis=-1)
             y4 = paddle.cumsum(x, axis=-2)
-
+            y5 = paddle.cumsum(x, axis=-1, dtype='int16')
             place = base.CUDAPlace(0) if use_gpu else base.CPUPlace()
             exe = base.Executor(place)
             exe.run(paddle.static.default_startup_program())
@@ -321,6 +333,7 @@ def run_static_int8(self, use_gpu=False):
                     y2,
                     y3,
                     y4,
+                    y5,
                 ],
             )
             z = np.cumsum(data_np)
@@ -331,6 +344,8 @@ def run_static_int8(self, use_gpu=False):
             np.testing.assert_allclose(z, out[2], rtol=1e-05)
             z = np.cumsum(data_np, axis=-2)
             np.testing.assert_allclose(z, out[3], rtol=1e-05)
+            z = np.cumsum(data_np, axis=-1, dtype="int16")
+            np.testing.assert_equal(z.dtype, out[4].dtype)
 
     def run_static_int16(self, use_gpu=False):
         with paddle.static.program_guard(paddle.static.Program()):
@@ -883,22 +898,6 @@ def test_check_grad(self):
 create_test_bf16_class(TestSumOpReverseExclusive)
 
 
-class BadInputTest(unittest.TestCase):
-    def test_error(self):
-        paddle.enable_static()
-        with paddle.static.program_guard(
-            paddle.static.Program(), paddle.static.Program()
-        ):
-
-            def test_bad_x():
-                data = [1, 2, 4]
-                result = paddle.cumsum(data, axis=0)
-
-            with self.assertRaises(TypeError):
-                test_bad_x()
-        paddle.disable_static()
-
-
 class TestTensorAxis(unittest.TestCase):
     def setUp(self):
         paddle.seed(2022)