Something dumb in the ordering of destruction

111254 · June 28, 2020, 7:54am

I am facing some Bug of Cuda destruction:https://github.com/pytorch/pytorch/issues/40425
Is there anything related with “dumb in the ordering of destruction”?The
The fbcode setting is meaning what?

github.com

pytorch/pytorch/blob/master/aten/src/ATen/cudnn/Handle.cpp#L12


#include <ATen/cuda/detail/DeviceThreadHandles.h>
#include <c10/cuda/CUDAStream.h>

namespace at { namespace native {
namespace {

void createCuDNNHandle(cudnnHandle_t *handle) {
  AT_CUDNN_CHECK(cudnnCreate(handle));
}

void destroyCuDNNHandle(cudnnHandle_t handle) {
// this is because of something dumb in the ordering of
// destruction. Sometimes atexit, the cuda context (or something)
// would already be destroyed by the time this gets destroyed. It
// happens in fbcode setting. @colesbury and I decided to not destroy
// the handle as a workaround.
//   - @soumith
#ifdef NO_CUDNN_DESTROY_HANDLE
#else
    cudnnDestroy(handle);
#endif

ptrblck · June 29, 2020, 8:06am

Are you running into any issues regarding the order of the handle destructions?
If so, this should be fixed in this PR. Could you install the latest nightly binary and retry it?

If you are not seeing an issue, you could ignore the ifdef as it’s used only in some internal builds, if I’m not mistaken.

111254 · June 29, 2020, 9:45am

Hi @ptrblck, i am very glad to see you reply this.
Have you ever see my question in github?
#40425
Are you sure my question have something to do with #39345?
I am trying #39345 Now.