Cudnn_convolution_backward is not a member of 'at'

rinkujadhav2013 · June 19, 2023, 4:18pm

Hi,
I’m trying to adapt my code from version 1.10.2 to 1.11.0 (for now). I was using cudnn_convolution_backward in 1.10.2 but I’m unable to use in 1.11.0. It throws an error saying

error: ‘cudnn_convolution_backward’ is not a member of ‘at’

but I can see that the function is in the namespace at even for version 1.11.0:

github.com

pytorch/pytorch/blob/v1.11.0/aten/src/ATen/native/cudnn/ConvShared.cpp#L417


      
              const Tensor& input_t,
              IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation, int64_t groups,
              bool benchmark, bool deterministic, bool allow_tf32)
          {
            return cudnn_convolution_backward_weight(
                "cudnn_convolution_backward_weight",
                weight_size, grad_output_t, input_t,
                padding, stride, dilation, groups, benchmark, deterministic, allow_tf32);
          }
          
          
std::tuple<at::Tensor,at::Tensor> cudnn_convolution_backward(
              const at::Tensor& input, const at::Tensor& grad_output_t, const at::Tensor& weight,
              IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation, int64_t groups,
              bool benchmark, bool deterministic, bool allow_tf32, std::array<bool,2> output_mask) {
          
          
  Tensor grad_output = grad_output_t.contiguous(input.suggest_memory_format());
          
          
  Tensor grad_input, grad_weight;
            if (input.numel() == 0) {
              if (output_mask[0]) {
                grad_input = at::empty_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);

I don’t see the difference between 1.11.0 and 1.10.2:

github.com

pytorch/pytorch/blob/v1.10.2/aten/src/ATen/native/cudnn/ConvShared.cpp#L393


      
              bool benchmark, bool deterministic, bool allow_tf32)
          {
            TensorArg grad_output{ grad_output_t, "grad_output", 1 },
                      weight{ weight_t, "weight", 2 };
            return cudnn_convolution_backward_input(
                "cudnn_convolution_backward_input",
                input_size, grad_output, weight,
                padding, stride, dilation, groups, benchmark, deterministic, allow_tf32);
          }
          
          
std::tuple<at::Tensor,at::Tensor> cudnn_convolution_backward(
              const at::Tensor& input, const at::Tensor& grad_output_t, const at::Tensor& weight,
              IntArrayRef padding, IntArrayRef stride, IntArrayRef dilation, int64_t groups,
              bool benchmark, bool deterministic, bool allow_tf32, std::array<bool,2> output_mask) {
          
          
  Tensor grad_output = grad_output_t.contiguous(input.suggest_memory_format());
          
          
  Tensor grad_input, grad_weight;
            if (input.numel() == 0) {
              if (output_mask[0]) {
                grad_input = at::empty_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);

ptrblck · June 19, 2023, 5:17pm

The cudnn_* methods are not exposed anymore and you should be able to drop the cudnn_ prefix in newer PyTorch versions.

rinkujadhav2013 · June 19, 2023, 9:35pm

Can I also just use the torch.ops.aten.convolution_backward function (cuz I’m using C++ method in Python anyway) for pytorch versions > 1.11.0?