【Paddle Tensor No.26】fix svdvals (#69820)

* fix svdvals

* fix

* fix

* fix

* fix bug in op_gen

* fix svdvals and restore op_gen

Author: aquagull

paddle/phi/kernels/cpu/svdvals_kernel.cc

@@ -107,7 +107,7 @@ void SvdvalsKernel(const Context& dev_ctx,
       0,
       phi::errors::InvalidArgument("The batch size of Input(X) must be > 0."));
   DDim s_dims;
-  if (batches == 1) {
+  if (x_dims.size() <= 2) {
     s_dims = {k};
   } else {
     s_dims = {batches, k};

python/paddle/tensor/linalg.py

@@ -3025,20 +3025,7 @@ def svdvals(x: Tensor, name: str | None = None) -> Tensor:
             Tensor(shape=[2], dtype=float32, place=Place(cpu), stop_gradient=True,
             [8.14753819, 0.78589684])
     """
-    if in_dynamic_or_pir_mode():
-        return _C_ops.svdvals(x)
-    else:
-        check_variable_and_dtype(x, 'dtype', ['float32', 'float64'], 'svdvals')
-        helper = LayerHelper('svdvals', **locals())
-        s = helper.create_variable_for_type_inference(dtype=x.dtype)
-        attrs = {}
-        helper.append_op(
-            type='svdvals',
-            inputs={'X': [x]},
-            outputs={'S': s},
-            attrs=attrs,
-        )
-        return s
+    return _C_ops.svdvals(x)
 
 
 def _conjugate(x):

test/legacy_test/test_svdvals_op.py

@@ -1,4 +1,4 @@
-#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#   Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -15,7 +15,7 @@
 import unittest
 
 import numpy as np
-from op_test import OpTest, skip_check_grad_ci
+from op_test import OpTest
 from utils import dygraph_guard, static_guard
 
 import paddle
@@ -69,13 +69,10 @@ def init_data(self):
         self.outputs = {"s": self._output_data}
 
 
-@skip_check_grad_ci(
-    reason="'check_grad' on singular values is not required for svdvals."
-)
 class TestSvdvalsBigMatrix(TestSvdvalsOp):
     def init_data(self):
         """Generate large input matrix."""
-        self._input_shape = (200, 300)
+        self._input_shape = (40, 40)
         self._input_data = np.random.random(self._input_shape).astype("float64")
         self._output_data = np.linalg.svd(
             self._input_data, compute_uv=False, hermitian=False
@@ -84,7 +81,13 @@ def init_data(self):
         self.outputs = {'s': self._output_data}
 
     def test_check_grad(self):
-        pass
+        self.check_grad(
+            ['x'],
+            ['s'],
+            numeric_grad_delta=0.001,
+            max_relative_error=1e-5,
+            check_pir=True,
+        )
 
 
 class TestSvdvalsAPI(unittest.TestCase):
@@ -103,8 +106,7 @@ def test_dygraph_api(self):
             # Test dynamic graph for svdvals
             s = paddle.linalg.svdvals(x)
             np_s = np.linalg.svd(self.x_np, compute_uv=False, hermitian=False)
-            self.assertTrue(np.allclose(np_s, s.numpy(), rtol=1e-6))
-
+            np.testing.assert_allclose(np_s, s.numpy(), rtol=1e-6)
             # Test with reshaped input
             x_reshaped = x.reshape([-1, 12, 10])
             s_reshaped = paddle.linalg.svdvals(x_reshaped)
@@ -114,8 +116,8 @@ def test_dygraph_api(self):
                     for matrix in self.x_np.reshape([-1, 12, 10])
                 ]
             )
-            self.assertTrue(
-                np.allclose(np_s_reshaped, s_reshaped.numpy(), rtol=1e-6)
+            np.testing.assert_allclose(
+                np_s_reshaped, s_reshaped.numpy(), rtol=1e-6
             )
 
     def test_static_api(self):
@@ -130,7 +132,7 @@ def test_static_api(self):
 
         np_s = np.linalg.svd(self.x_np, compute_uv=False, hermitian=False)
         for r in res:
-            self.assertTrue(np.allclose(np_s, r, rtol=1e-6))
+            np.testing.assert_allclose(np_s, r, rtol=1e-6)
 
     def test_error(self):
         """Test invalid inputs for svdvals"""

test/white_list/op_threshold_white_list.py

@@ -48,6 +48,7 @@
     'lgamma',
     'sparse_attention',
     'svd',
+    'svdvals',
     'matrix_power',
     'cholesky_solve',
     'solve',