commit 88b1f9a021ba750837fe269d50b9e5c8e786d29d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Oct 14 23:12:03 2026 +0200

dev-python/flash-attn: build with C++20 on PyTorch 2.13+

PyTorch 2.14's ATen headers require C++20, but setup.py pins
-std=c++17, so the extension no longer compiled against the tree's
PyTorch. Backport upstream's fix, which selects C++20 from the
installed torch version. Revbump.

commit 3ab8657fdd41badc99c5836fac7718d2219f5c18
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Oct 14 14:15:39 2026 +0200

dev-python/flash-attn: fix CUTLASS staging in 2.8.3_p1

The sdist bundles csrc/cutlass without its collective builders, so the
old rmdir failed silently and the fetched commit was nested inside the
partial copy instead of replacing it. setup.py also imports torch,
which sci-ml/pytorch provides, not sci-ml/caffe2. Revbump.

commit f3a8117506e21094310f27eb23315606a4c098c7
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Oct 14 09:48:15 2026 +0200

dev-python/flash-attn: add 2.8.3_p2

Backports upstream's weights_only=True torch.load fix
(CVE-2026-31253) and an int32 overflow fix; the CUTLASS submodule
commit is unchanged.

commit b11bddc17a3a712865d690ad42faa6c95b8dd6d1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 12 17:44:29 2026 +0200

dev-python/flash-attn: enable Python 3.14

Upstream requires Python 3.9 or newer, and the Torch CUDA extension
dependency stack supports Python 3.14.

commit d9af0891b1e0f73d9672c32570635eaa2fc99503
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 05:30:16 2026 +0200

dev-python/flash-attn: trim 2.8.3_p1 comment prose

commit 4e2ebdbca4a5da09108f1ec09d57969920ea3213
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 16:40:29 2026 +0200

dev-python/flash-attn: enforce CUDA build dependencies

The exact 2.8.3.post1 setup imports wheel and psutil during extension construction,
requires CUDA 11.7 or newer, and builds against PyTorch CUDA APIs. It also consumes
FLASH_ATTN_CUDA_ARCHS directly, so use that interface instead of an ignored PyTorch
variable.

commit 41c6cc60fe28b27cf770092112bdb8af84c811af
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 20:45:35 2026 +0200

dev-python/flash-attn: drop 2.8.3

commit 0aafac8cb2029acf9eee4280f0081bde1534dd2b
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 20:44:15 2026 +0200

dev-python/flash-attn: add 2.8.3_p1

commit e59a6b5d89df7a1e0f72333cedd2fb8e336c40fa
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 10:09:16 2026 +0200

dev-python/flash-attn: new package, add 2.8.3

Provides flash_attn_varlen_func, required by the TRELLIS sparse-attention path.
The sdist omits its CUTLASS submodule, so a second SRC_URI stages CUTLASS at
the pinned commit (== 4.0.0) into csrc/cutlass during src_prepare. MAX_JOBS is
held at 2: the backward hdim128/256 kernels OOM this host at higher job counts.