Skip to content

[rocm7.1_internal_testing] fix large tensor sort on ROCm #2543

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
merged 1 commit into from
Aug 25, 2025
Merged

[rocm7.1_internal_testing] fix large tensor sort on ROCm #2543

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 5 additions & 3 deletions aten/src/ATen/native/cuda/SortStable.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,8 +225,9 @@ void launch_stable_sort_kernel(
return;
}

int64_t numel_or_intmax =
std::min(numel, static_cast<int64_t>(std::numeric_limits<int>::max()));
const int64_t intmax = static_cast<int64_t>(std::numeric_limits<int>::max());
// On ROCm, std::min -> ::min did not work as expected on when input values >= 2147483648
int64_t numel_or_intmax = numel < intmax ? numel : intmax;
int64_t nsort = self.size(dim);
int64_t nbatch = (numel_or_intmax / nsort) * nsort;
TORCH_CHECK(nbatch > 0, "Cannot sort dimension of length ", nsort);
Expand All @@ -238,7 +239,8 @@ void launch_stable_sort_kernel(
scalar_t* values_ptr = values.mutable_data_ptr<scalar_t>();
int64_t remaining = numel;
while (remaining > 0) {
int64_t n = std::min(remaining, nbatch);
// On ROCm, std::min -> ::min did not work as expected on when input values >= 2147483648
int64_t n = remaining < nbatch ? remaining : nbatch;
Copy link
Collaborator

@jithunnair-amd jithunnair-amd Aug 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In an offline discussion, we discussed why use ternary operator instead of the std::min<int64_t> templated version.

@dnikolaev-amd verified that std::min<int64_t> does not get hipified to ::min<int64_t> it remains std::min<int64_t>.

@jeffdaily said "The only argument I can think of against using std::min<int64_t> in device code is generally not recommended. I remember Andy raised an issue about it. pytorch#128253"

However, the usages being addressed in this PR are in host code IIUC. So is there a reason we should still use ternary operator instead?

int64_t nsegments = n / nsort;

if (nsegments == 1 ||
Expand Down
2 changes: 1 addition & 1 deletion test/test_sort_and_select.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -215,7 +215,7 @@ def test_stable_sort(self, device, dtype):
)

@onlyCUDA
@dtypes(torch.uint8)
@dtypes(torch.float16)
@largeTensorTest("200GB") # Unfortunately 80GB A100 is not large enough
def test_sort_large(self, device, dtype):
t0 = torch.randperm(8192, device=device).to(dtype)
Expand Down