gpo.zugaina.org

Search Portage & Overlays:

dev-python/nvidia-cutlass-dsl-libs-base

NVIDIA CUTLASS Python DSL — base shared libs

Screenshots

  • nvidia-cutlass-dsl-libs-base-4.5.0
    -* ~amd64
    python_targets_python3_11 python_targets_python3_12 python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: NVIDIA-CUDA   
    Overlay: stuff

ChangeLog

commit 27675dae60cb2c7985bb9680c24ecadf4bf76768
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 17:33:15 2026 +0200

dev-python/nvidia-cutlass-dsl-libs-base: install only files unique to base wheel

The libs-base and libs-cu13 wheels overlap on 217 file paths (out of
~227 each), shipping subtly different *contents* at those paths —
the cu13 build vs the generic build. pip's behaviour when both are
installed is to let cu13 overwrite base; portage's collision-protect
errors instead. Mirror pip's end state by keeping only this wheel's
*unique* files (LICENSE, libcuda_dialect_runtime_static.a, and
python_packages/cutlass/utils/block.py — three files in 4.5.0,
verified against the cu13 wheel's manifest), plus dropping the
shared nvidia_cutlass_dsl.pth so the cu13 ebuild owns that.

The cu13 ebuild then installs everything else, including its own
variants of the 217 shared paths, and the two packages coexist
without collision. End state matches what pip produces with
`pip install nvidia-cutlass-dsl[cu13]`.

commit dbe708080b97e4a09d457dcde1b6eb07170d6613
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 17:22:43 2026 +0200

dev-python/nvidia-cutlass-dsl-libs-base: new package, 4.5.0

Tier 2 of the vllm CUDA target packaging cycle. Base shared libs of
NVIDIA's CUTLASS Python DSL (the parent nvidia-cutlass-dsl
metapackage pulls libs-base unconditionally, plus libs-cu13 under
its cu13 extra).

Wheel-only on PyPI — NVIDIA does not publish source for the binary
DSL libs. Per-Python wheels (cp311 through cp314) installed via
\`installer\` per-impl. NVIDIA-CUDA EULA → RESTRICT="bindist mirror".

Surfaced from the front-loaded audit's discovery that
nvidia-cutlass-dsl is split across three PyPI packages (parent +
libs-base + libs-cu13). Wired against cuda-python so the underlying
driver/runtime bindings come from this overlay's tracked stack
rather than NVIDIA's pip-published cuda-python wheel.