diff options
| author |   Sv. Lockal <lockalsash@gmail.com> | 2024-07-25 09:27:07 +0000 |
|---|---|---|
| committer |   Alfredo Tupone <tupone@gentoo.org> | 2024-07-27 21:08:42 +0200 |
| commit | ca2e68ab5527af67bf95684068758b4efbd5b5a5 (patch) | |
| tree | d73b47f8656dfc94f45836bd1741ebeafa4b7d06 /app-shells/mcfly | |
| parent | sys-cluster/kube-scheduler: drop 1.27.12 (diff) | |
| download | gentoo-ca2e68ab5527af67bf95684068758b4efbd5b5a5.tar.gz gentoo-ca2e68ab5527af67bf95684068758b4efbd5b5a5.tar.bz2 gentoo-ca2e68ab5527af67bf95684068758b4efbd5b5a5.zip | |
sci-libs/caffe2: update dependencies to fix rocm flag
pytorch 2.3.0 introduced 2 new direct dependencies: hipBLASLt and aotriton.
pytorch uses hipBLASLt to perform gemm operation on datacenter AMD Instinct GPUs. For other GPUs pytorch fallbacks to hipBLAS.
caffe2-2.3.x ebuilds now contain a patch to optionally disable this dependency, when none AMDGPU_TARGETS="gfx90a gfx940 gfx941 gfx942" is used.
pytorch uses aotriton to perform FlashAttention operation.
caffe2-2.3.x ebuilds now contain a patch which fully disables aotriton dependency, as there is no such package yet.
Technically aotriton can be compiled (with minor patches), but I suggest to wait for next releases.
It is a massive burden, as it depends on forked triton and forked clang (merge with upstream is not expected anytime soon).
aotriton is usually distributed as a huge static (!) library (but in next release library will be shared).
Minor fixes added for compatibility with libc++ (used in experimental llvm Gentoo profile), however other ebuilds also require minor patches
(in other words: right now ROCm ecosystem can be compiled with libc++, but only by people with experience in C++).
Closes: https://bugs.gentoo.org/931046
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Signed-off-by: Alfredo Tupone <tupone@gentoo.org>
Diffstat (limited to 'app-shells/mcfly')
0 files changed, 0 insertions, 0 deletions