PyTorch CUDA backward kernels report unchecked out-of-bounds writes
Two PyTorch bug reports describe CUDA backward kernels that perform unchecked out-of-bounds atomic writes. In both cases the GPU path can corrupt memory when inputs fall outside the range the kernel assumes.
Adaptive max-pool backward writes past bounds on bad indices
A report against PyTorch (#206062) says aten::adaptive_max_pool2d_backward performs out-of-bounds atomic writes on CUDA when given out-of-range int64 indices. The front end does not validate those indices before the kernel runs. Developers who rely on this backward op should treat untrusted or unchecked indices as a memory-safety risk until the kernel is fixed.
Replication pad 3D backward overflows int32 and silently corrupts GPU memory
A second report (#206059) says aten::replication_pad3d_backward hits an INT32 out-of-range condition when padding reaches 2^31, producing out-of-bounds atomic writes on CUDA. The CPU implementation rejects the same input, while the GPU path silently corrupts memory. The mismatch matters for anyone porting large-padding workloads from CPU to CUDA and expecting equivalent rejection.