commit 4a6c268f4e6175a21f7f31e859b675d14499aa4a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Jun 18 08:00:57 2026 +0200

dev-python/triton-bin: add 3.7.1

commit 7a6588a2044164910e43e4d206cf8e79d9d7b330
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 14:27:25 2026 +0200

dev-python/triton-bin: add 3.7.0

torch-2.12.0 pairs with triton 3.7.0 (pytorch
.ci/docker/triton_version.txt); the existing 3.6.0 is torch-2.11.0's
triton. humming-kernels (unversioned triton-bin dep) picks 3.7.0 up to
match the torch-2.12 stack. 3.6.0 stays for vllm and comfyui, which
still pin ~3.6.0 -- those pins predate the torch-2.12 bump and want
re-evaluating against 3.7.0, but that needs a build-verify pass.

commit 994121024687b6fd737cb5548664f4f8e03e81ab
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Jun 14 09:26:23 2026 +0200

dev-python/triton-bin: add 3.6.0

GPU programming language and compiler, packaged from the upstream PyPI
wheels (cp312/cp313/cp314, x86_64). Triton builds from an LLVM/MLIR
source tree that is impractical to compile in-tree, so the binary wheel
is the supported form, mirroring dev-python/cuda-tile-bin. The wheel is
self-contained (no python RDEPEND) and bundles its own LLVM plus the
NVIDIA ptxas toolchain.

3.6.0 is the Triton release that pytorch-2.11.0 pairs with
(pytorch .ci/docker/triton_version.txt), which the source-built
sci-ml/pytorch needs for its torch.compile/inductor backend. Pulled by
dev-python/vllm on CUDA, whose @triton.jit kernels otherwise fail at
first GPU inference.

Drop Triton's proton / proton-viewer profiler console scripts from
/usr/bin: the names are generic enough to collide (e.g. with Valve
Proton) and torch/vllm do not use them; the triton.profiler.proton
module stays importable.