dtype

mindspore/ccsrc/plugin/device/gpu/kernel/nn/kl_div_loss_grad_kernel.cc

@@ -72,6 +72,12 @@ bool KLDivLossGradGpuKernelMod::LaunchKernel(const std::vector<AddressPtr> &inpu
 
 std::vector<std::pair<KernelAttr, KLDivLossGradGpuKernelMod::KLDivLossLaunchFunc>>
   KLDivLossGradGpuKernelMod::func_list_ = {
+    {KernelAttr()
+       .AddInputAttr(kNumberTypeFloat64)
+       .AddInputAttr(kNumberTypeFloat64)
+       .AddInputAttr(kNumberTypeFloat64)
+       .AddOutputAttr(kNumberTypeFloat64),
+     &KLDivLossGradGpuKernelMod::LaunchKernel<double>},
     {KernelAttr()
        .AddInputAttr(kNumberTypeFloat32)
        .AddInputAttr(kNumberTypeFloat32)

tests/st/ops/gpu/test_kl_div_op.py

@@ -88,3 +88,27 @@ def test_kl_div_loss_grad():
                   -0.03094601, -0.14319494]
 
     assert np.allclose(dx[0].asnumpy(), dx1_expect)
+
+
+@pytest.mark.level1
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_kl_div_loss_grad_float64():
+    """
+    Feature: Test KLDivLossGrad.
+    Description: Test KLDivLossGrad op with float inputs.
+    Expectation: The result match to expect.
+    """
+    np.random.seed(42)
+    prediction = np.random.rand(20).astype(np.float64)
+    target = np.random.rand(20).astype(np.float64)
+    sens = np.random.rand(20).astype(np.float64)
+    grad = Grad(Net())
+    dx = grad(Tensor(prediction), Tensor(target), Tensor(sens))
+
+    dx1_expect = [-0.07466945, -0.06907414, -0.01004642, -0.3331403, -0.11802178, -0.52019656,
+                  -0.06224053, -0.2674369, -0.32387912, -0.00858657, -0.58906615, -0.13217884,
+                  -0.06111591, -0.8490888, -0.57735133, -0.7452407, -0.02695603, -0.01914206,
+                  -0.03094601, -0.14319494]
+
+    assert np.allclose(dx[0].asnumpy(), dx1_expect)