diff --git a/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu b/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu
index 8d3beab888f..f1b0f75b008 100644
--- a/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu
+++ b/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu
@@ -228,7 +228,9 @@ void deformable_im2col(
     int deformable_group,
     bool use_mask,
     at::Tensor data_col) {
-  int64_t num_kernels = (int64_t)n_in_channels * out_h * out_w * parallel_imgs;
+  at::cuda::CUDAGuard device_guard(input.get_device());
+
+  const int64_t num_kernels = (int64_t)n_in_channels * out_h * out_w * parallel_imgs;
 
   const unsigned int threads = GET_THREADS();
   const unsigned int blocks = GET_BLOCKS(threads, num_kernels);
@@ -408,12 +410,14 @@ void compute_grad_input(
     int n_offset_grps,
     bool use_mask,
     at::Tensor grad_im) {
-  int out_h =
+  at::cuda::CUDAGuard device_guard(columns.get_device());
+
+  const int out_h =
       (height + 2 * pad_h - (dilation_h * (weight_h - 1) + 1)) / stride_h + 1;
-  int out_w =
+  const int out_w =
       (width + 2 * pad_w - (dilation_w * (weight_w - 1) + 1)) / stride_w + 1;
 
-  int64_t num_kernels =
+  const int64_t num_kernels =
       (int64_t)channels * weight_h * weight_w * out_h * out_w * parallel_imgs;
 
   const unsigned int threads = GET_THREADS();
@@ -650,11 +654,13 @@ void compute_grad_offset_and_mask(
     bool use_mask,
     at::Tensor grad_offset,
     at::Tensor grad_mask) {
-  int out_h =
+  at::cuda::CUDAGuard device_guard(columns.get_device());
+
+  const int out_h =
       (height + 2 * pad_h - (dilation_h * (weight_h - 1) + 1)) / stride_h + 1;
-  int out_w =
+  const int out_w =
       (width + 2 * pad_w - (dilation_w * (weight_w - 1) + 1)) / stride_w + 1;
-  int64_t num_kernels = (int64_t)out_h * out_w * 2 * weight_h * weight_w *
+  const int64_t num_kernels = (int64_t)out_h * out_w * 2 * weight_h * weight_w *
       n_offset_grps * parallel_imgs;
 
   const unsigned int threads = GET_THREADS();