Skip to content

Fix int32 overflow in CUDA Gather kernel for large tensors #28108

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
justinchuby merged 2 commits into from
Apr 17, 2026
Merged

Fix int32 overflow in CUDA Gather kernel for large tensors #28108

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
de878dc
Fix int32 overflow in CUDA Gather kernel for large tensors
justinchuby Apr 17, 2026
27540d8
Update onnxruntime/core/providers/cuda/tensor/gather_impl.cu
justinchuby Apr 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions onnxruntime/core/providers/cuda/tensor/gather_impl.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,6 @@ __global__ void _GatherKernel(
T* output_data,
const CUDA_LONG N) {
CALCULATE_ELEMENTWISE_INDEX_OR_EXIT(id, N);
CUDA_LONG input_index = 0;
int input_block_index, block_offset;
output_block_size.divmod(id, input_block_index, block_offset);
int indices_index, offset;
Expand All @@ -47,7 +46,10 @@ __global__ void _GatherKernel(
return;
}

input_index = input_block_index * input_block_size + idx * block_size.d_ + offset;
// Use int64_t to avoid overflow when the input tensor has more than
// INT32_MAX elements (e.g. a [262144, 8960] embedding table = 2.35B).
int64_t input_index = static_cast<int64_t>(input_block_index) * input_block_size +
idx * block_size.d_ + offset;
output_data[id] = input_data[input_index];
}

Expand Down
Loading