gpo.zugaina.org

Search Portage & Overlays:

dev-python/amd-quark-bin

AMD's quantization toolkit for ML model optimization (binary wheel)

Screenshots

  • amd-quark-bin-0.12_p1
    -* ~amd64
    python_single_target_python3_12 python_single_target_python3_13

    View      Download      Browse     License: MIT   
    Overlay: stuff
  • amd-quark-bin-0.11.2
    -* ~amd64
    +python_single_target_python3_12

    View      Download      Browse     License: MIT   
    Overlay: stuff

ChangeLog

commit 20f504e7340682d70d9eeb1d35143a25f59ce3e2
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Jul 26 22:43:50 2026 +0200

dev-python/amd-quark-bin: allow python3_13 on 0.12_p1

Upstream raised requires-python to "<3.14,>=3.11" between 0.11.2 and
0.12.post1, but PYTHON_COMPAT was copied from the older ebuild and left at
python3_12 alone. Since the default target moved on, the package could not be
installed at all -- emerge rejected it on an unsatisfiable REQUIRED_USE.

Widened to the upstream ceiling and noted why python3_14, which pkgcheck keeps
offering, must not be taken.

commit 9ac10d5c1edb786e141516b7eaf15362729f14f9
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 18 13:43:47 2026 +0200

dev-python/amd-quark-bin: add 0.12_p1

Upstream 0.12.post1 (Portage PV 0.12_p1). requires_dist changed vs
0.11.2: onnx moved to >=1.21.0,<=1.22.0 and a psutil dep was added, and
the old numpy<=2.1.3 cap was dropped (now >=2.0). ::gentoo has no
onnx>=1.21, so the RDEPEND floors at the highest available (>=onnx-1.20.1,
self-healing once 1.21+ lands); safe because the ONNX paths are already
non-functional here (onnxscript/onnxslim unpackaged) and the tool's
PyTorch quantization paths don't import onnx. onnx-1.20.1 + psutil-7.2.2
resolve; nothing in the tree depends on amd-quark-bin.

commit c43ad792d38882892e37986660c9395ccdee15c7
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed May 13 14:34:28 2026 +0200

dev-python/amd-quark-bin: disable py3.11

commit 2789fc1d51872d9c9b1d251fd677da08394687ff
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun May 10 15:28:26 2026 +0200

dev-python/amd-quark-bin: switch to DISTUTILS_SINGLE_IMPL

sci-ml/evaluate is SINGLE_IMPL; multi-impl consumer with bare
$ on it produces python_targets_python3_*(-)? that the
single-impl child can't expose. As part of the conversion, replace the
python_foreach_impl wheel install loop with a single-shot python_setup
+ $ -m installer call.

--scan=n: pkgcheck PythonCompatUpdate informational hint (3.13/3.14)
is policy — upstream wheel pins requires_python = "<3.13,>=3.9", so the
PYTHON_COMPAT cap is intentional, not a bump candidate.

commit 83a9e52275b5f54cc02c1340f4d93ce10b1b48b5
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 01:06:44 2026 +0200

dev-python/amd-quark-bin: new package

AMD's quantization toolkit (binary wheel). Required by dev-python/vllm
in the rocm path.

Wheel-only on PyPI. Upstream pins requires_python = '<3.13,>=3.9' so
this ebuild is python3_ only — can't compile-test on this host
(py3.13/3.14 only) but ebuild metadata is verified.

Two upstream version caps relaxed because ::gentoo doesn't carry the
older versions (numpy <=2.1.3 → unbounded; onnx <=1.19.0 → ~1.18.0).
onnxscript and onnxslim install-time deps are unpackaged in our trees
— ONNX-export quantization paths will fail with ImportError, but the
PyTorch-only paths vllm normally uses don't need them.

NonsolvableDepsInStable is policy noise — the deps exist, just
~amd64-only at python3_ targets.