deform_conv2d kernels are launched on the default CUDA stream

[simonbyrne]

🐛 Describe the bug

The deform_conv2d kernels are launched on the default CUDA stream, not the current pytorch stream

vision/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu

Line 253 in 65eb7b9

deformable_im2col_kernel<scalar_t, int64_t><<<blocks, threads>>>(

pytorch kernels should launch on the stream returned by at::cuda::getCurrentCUDAStream();, e.g.

vision/torchvision/csrc/ops/cuda/roi_pool_kernel.cu

Line 158 in 65eb7b9

cudaStream_t stream = at::cuda::getCurrentCUDAStream();

Versions

It's on the current main branch