[Bug Fix] Fix view grad when out_grad is not contiguous (PaddlePaddle#76679)

* Fix view grad when out_grad is not contiguous

* feat(phi): add no_need_buffer for view_shape_grad in backward.yaml

Add `no_need_buffer: input` to the view_shape_grad backward operator configuration. This indicates that the gradient computation for view_shape does not require the input tensor buffer, optimizing memory usage during backward passes.

* feat(test): add gradient test for view and transpose operations

Add a new test case `TestViewGrad.test_dygraph` to verify gradient correctness when using `view` followed by `transpose` in dynamic graph mode. The test ensures that the computed gradients match expected values, improving coverage for tensor manipulation operations.

Author: LittleHeroZZZX

paddle/phi/kernels/stride/view_grad_kernel.cc

@@ -15,6 +15,8 @@
 #include "paddle/common/flags.h"
 #include "paddle/phi/backends/all_context.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/funcs/strided_reshape_utils.h"
+#include "paddle/phi/kernels/funcs/strided_utils.h"
 #include "paddle/phi/kernels/view_kernel.h"
 
 COMMON_DECLARE_bool(use_stride_kernel);
@@ -32,8 +34,34 @@ void ViewShapeGradKernel(const Context& dev_ctx,
         "FLAGS_use_stride_kernel is closed. Strided kernel "
         "be called, something wrong has happened!"));
   }
-  ViewShapeStridedKernel<Context>(
-      dev_ctx, out_grad, common::vectorize<int64_t>(input.dims()), input_grad);
+  DDim target_dims = input.dims();
+  DDim target_stride;
+
+  if (ReshapeStride(
+          out_grad.dims(), out_grad.strides(), target_dims, target_stride)) {
+    input_grad->set_meta(out_grad.meta());
+    input_grad->Resize(target_dims);
+    input_grad->set_strides(target_stride);
+    input_grad->set_offset(out_grad.offset());
+    input_grad->ResetHolder(out_grad.Holder());
+    input_grad->ShareInplaceVersionCounterWith(out_grad);
+  } else {
+    DenseTensor contiguous_tmp;
+    DenseTensor tmp_out_grad = out_grad;
+
+    contiguous_tmp.set_meta(tmp_out_grad.meta());
+
+    PD_VISIT_ALL_TYPES(out_grad.dtype(), "ViewShapeGradKernel", ([&] {
+                         phi::StridedTensorContiguous<data_t>(tmp_out_grad,
+                                                              &contiguous_tmp);
+                       }));
+
+    input_grad->set_meta(contiguous_tmp.meta());
+    input_grad->Resize(target_dims);
+    input_grad->set_strides(DenseTensorMeta::calc_strides(target_dims));
+    input_grad->set_offset(0);
+    input_grad->ResetHolder(contiguous_tmp.Holder());
+  }
 }
 
 template <typename Context>

paddle/phi/ops/yaml/backward.yaml

@@ -4044,6 +4044,7 @@
   kernel :
     func : view_shape_grad
   backward : view_shape_double_grad
+  no_need_buffer: input
 
 - backward_op : warpctc_grad
   forward : warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank = 0, bool norm_by_times = false) -> Tensor(loss), Tensor(warpctcgrad)

test/legacy_test/test_stride.py

@@ -1092,5 +1092,20 @@ def func2():
         func2()
 
 
+class TestViewGrad(unittest.TestCase):
+    def test_dygraph(self):
+        paddle.disable_static()
+        x = paddle.randn(2, 12, requires_grad=True)
+
+        y = x.view(2, 3, 4)
+        z = y.transpose(1, 2)
+
+        loss = z.sum()
+        loss.backward()
+
+        x_grad_expected = paddle.full_like(x, 1.0)
+        self.assertEqual((x.grad == x_grad_expected).all(), True)
+
+
 if __name__ == '__main__':
     unittest.main()