sci-libs/aotriton
Ahead of Time (AOT) Triton Math Library
Screenshots
-
aotriton-0.14.2b-* ~amd64+amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1103 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 +amdgpu_targets_gfx1250
View 
Download 
Browse License: MIT Overlay: stuff -
aotriton-0.13b-r1-* ~amd64+amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1102 +amdgpu_targets_gfx1103 +amdgpu_targets_gfx1150 +amdgpu_targets_gfx1151 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 +amdgpu_targets_gfx1250
View Download Browse License: MIT Overlay: stuff
ChangeLog
commit d70659c7de850d1909acc3615d6e1efc267eb166
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 26 18:29:07 2026 +0200
sci-libs/aotriton: add 0.14.2b
0.14 adds flash-attention kernels written in FlyDSL for gfx950,
gfx1201 and gfx1250. Its CMake refuses the stock FlyDSL wheel while
third_party/flydsl-llvm.txt pins a ROCm LLVM fork commit, because
FlyDSL built against upstream LLVM miscompiles. A patch adds
AOTRITON_USE_SYSTEM_FLYDSL: the build venv sees dev-python/flydsl
through --system-site-packages, and CMake requires the pinned commit
in its installed version (+llvm<sha12>) instead of upstream's
wheel-name cache key. The FlyDSL kernel sources come from the same
v0.3.1 tarball instead of a configure-time clone.
ROCM_PATH is set only for the kernel build, to flydsl's toolkit tree
(ld.lld and llvm-readelf of the pinned LLVM). hip-config-amd.cmake
derives its clang root from ROCM_PATH, so setting it at configure
time breaks hip::host.
CMake 3.27 is now the floor. The iostream patch still applies and is
shared with 0.13b.
gfx1150 has no FlyDSL kernel variant, so a gfx1150 build does not
exercise that path; it was checked with gfx1201 and Triton kernels
skipped.
commit e72acddb42fc375bb23f6f74bcf5b21d513cefc8
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 04:50:09 2026 +0200
sci-libs/aotriton: trim 0.13b-r1 comment prose
commit 7166916c00fb2d52f5b3e39ef9a9a07705596b5f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 23:38:51 2026 +0200
sci-libs/aotriton: relax the dev-util/hip cap to <11
The <dev-util/hip-7.3 cap was inherited from sci-libs/aotriton-bin, where it
tracks which *-rocmX.X-shared.tar.gz shim archives upstream publishes -- a
property of the binary distribution that has no analogue in a source build.
Carried over here it only excluded the ROCm 10.0 stack, leaving the from-source
half of virtual/aotriton unsolvable while aotriton-bin-0.13b, packaging the
same upstream release, already allows it.
Upstream imposes no bound of its own: CMakeLists.txt at the 0.13b tag calls
find_package(hip REQUIRED) with no version argument. Match the -bin sibling at
the same version rather than dropping the cap outright, so the ebuild still
claims only what has been verified.
commit 9601208e8c5f8bdd39dcf63a50d532718244ffb8
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 23 16:26:35 2026 +0200
sci-libs/aotriton: new package
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 26 18:29:07 2026 +0200
sci-libs/aotriton: add 0.14.2b
0.14 adds flash-attention kernels written in FlyDSL for gfx950,
gfx1201 and gfx1250. Its CMake refuses the stock FlyDSL wheel while
third_party/flydsl-llvm.txt pins a ROCm LLVM fork commit, because
FlyDSL built against upstream LLVM miscompiles. A patch adds
AOTRITON_USE_SYSTEM_FLYDSL: the build venv sees dev-python/flydsl
through --system-site-packages, and CMake requires the pinned commit
in its installed version (+llvm<sha12>) instead of upstream's
wheel-name cache key. The FlyDSL kernel sources come from the same
v0.3.1 tarball instead of a configure-time clone.
ROCM_PATH is set only for the kernel build, to flydsl's toolkit tree
(ld.lld and llvm-readelf of the pinned LLVM). hip-config-amd.cmake
derives its clang root from ROCM_PATH, so setting it at configure
time breaks hip::host.
CMake 3.27 is now the floor. The iostream patch still applies and is
shared with 0.13b.
gfx1150 has no FlyDSL kernel variant, so a gfx1150 build does not
exercise that path; it was checked with gfx1201 and Triton kernels
skipped.
commit e72acddb42fc375bb23f6f74bcf5b21d513cefc8
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 04:50:09 2026 +0200
sci-libs/aotriton: trim 0.13b-r1 comment prose
commit 7166916c00fb2d52f5b3e39ef9a9a07705596b5f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 23:38:51 2026 +0200
sci-libs/aotriton: relax the dev-util/hip cap to <11
The <dev-util/hip-7.3 cap was inherited from sci-libs/aotriton-bin, where it
tracks which *-rocmX.X-shared.tar.gz shim archives upstream publishes -- a
property of the binary distribution that has no analogue in a source build.
Carried over here it only excluded the ROCm 10.0 stack, leaving the from-source
half of virtual/aotriton unsolvable while aotriton-bin-0.13b, packaging the
same upstream release, already allows it.
Upstream imposes no bound of its own: CMakeLists.txt at the 0.13b tag calls
find_package(hip REQUIRED) with no version argument. Match the -bin sibling at
the same version rather than dropping the cap outright, so the ebuild still
claims only what has been verified.
commit 9601208e8c5f8bdd39dcf63a50d532718244ffb8
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 23 16:26:35 2026 +0200
sci-libs/aotriton: new package