after aligned d_stride :::get error ::::RuntimeError: CUDA error: an illegal memory access was encountered

[SWALLOWhajnal]

I specifically aligned d_stride, but it still reported a memory problem

RuntimeError: CUDA error: an illegal memory access was encountered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions

[SWALLOWhajnal]

my batch_size=1 It's working.
my batch_size=2 It's RuntimeError: CUDA error: an illegal memory access was encountered
my batch_size=3 It's working.

[SWALLOWhajnal]

Error from dcnv4_col2im_cuda.cuh

---

int shm_size = sizeof(opmath_t) * (G * block_multiplier * K) * 2;
if(!check_backward_warpp(d_stride, D)){
shm_size = sizeof(opmath_t) * ((G * block_multiplier * K) * 2 + G * block_multiplier * blockdimX * 3);
}
cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize,
shm_size);

kernel<<<num_blocks, num_threads, shm_size, stream>>>(
value, p_offset, grad_output, G, D, Q, kernel_h, kernel_w, stride_h,
stride_w, pad_h, pad_w, dilation_h, dilation_w, height_in, width_in,
height_out, width_out, offset_scale, remove_center, block_multiplier,
grad_im, grad_offset, padded_offset_dim);

cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
printf("error in dcnv4_im2col_cuda: %s\n", cudaGetErrorString(err));
printf("launch arguments: gridDim=(%d, %d, %d), blockDim=(%d, %d, %d), "
"shm_size=%d\n\n",
num_blocks.x, num_blocks.y, num_blocks.z, num_threads.x,
num_threads.y, num_threads.z, shm_size);
AT_ASSERTM(false, "kernel launch error");
}
}

error in dcnv4_im2col_cuda: an illegal memory access was encountered
launch arguments: gridDim=(8192, 1, 1), blockDim=(12, 16, 1), shm_size=3456。
/DCNv4_op/src/cuda/dcnv4_col2im_cuda.cuh":470, please report a bug to PyTorch. kernel launch error

I can't fix this error.

[Q-M-E]

I have encountered the same issue.