Skip to content

[release/2.12] [ROCm] derive the hipcub compatible cccl version from 'HIPCUB_CCCL_VERSION' #3441

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

Draft
jithunnair-amd wants to merge 2 commits into
base: release/2.12
Choose a base branch
from
+18 −5
Draft

[release/2.12] [ROCm] derive the hipcub compatible cccl version from 'HIPCUB_CCCL_VERSION' #3441

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2fb28e9
[ROCm] derive the hipcub compatible cccl version from 'HIPCUB_CCCL_VE…
Naraenda Jul 8, 2026
5aa0a1e
resolve merge conflict (#21)
stanleytsang-amd Jul 16, 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: 5 additions & 1 deletion aten/src/ATen/cuda/cub.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,7 +70,11 @@
#define ATEN_CUB_MAXIMUM() NO_ROCM(at_cuda_detail)ROCM_HIPCUB(::cub)::Max()
#endif

#if defined(USE_ROCM)
// Pre ROCm 8.0 we require backporting of FpLimits specialization
// for 'c10::BFloat16'. However, ROCm 8.0 lifts the compatible
// CUB version to 3.x which uses libhipcxx and derives numerical
// limits from there.
#if defined(USE_ROCM) && CUB_VERSION < 300000

// backport https://github.com/NVIDIA/cub/pull/306 for c10::BFloat16

Expand Down
14 changes: 10 additions & 4 deletions aten/src/ATen/cuda/cub_definitions.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,14 +1,20 @@
#pragma once

#if !defined(USE_ROCM)
#include <cuda.h> // for CUDA_VERSION
#endif

#if !defined(USE_ROCM)
#include <cuda.h> // for CUDA_VERSION
#include <cub/version.cuh>
#else
// Check if we can find HIPCUB_CCCL_VERSION.
#include <hipcub/hipcub_version.hpp>
// Older versions of hipCUB do not support the CUB V3 API.
#if defined(HIPCUB_CCCL_VERSION) && HIPCUB_CCCL_VERSION >= 300000
#define CUB_VERSION HIPCUB_CCCL_VERSION
#else
// If we cannot find a compatible CCCL version, fallback to
// a very old version.
#define CUB_VERSION 200001
#endif
#endif

// cub support for CUB_WRAPPED_NAMESPACE is added to cub 1.13.1 in:
// https://github.com/NVIDIA/cub/pull/326
Expand Down
3 changes: 3 additions & 0 deletions cmake/public/LoadHIP.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -228,6 +228,9 @@ if(HIP_FOUND)

# Optional components.
find_package_and_print_version(hipsparselt) # Will be required when ready.
# ROCm 8.0 and later requires libhipcxx! This should be marked as
# 'REQUIRED' once minimal ROCm version is bumped to 8.0 or later.
find_package_and_print_version(libhipcxx)

list(REMOVE_DUPLICATES ROCM_INCLUDE_DIRS)

Expand Down