dev-python/cuda-tile-bin
NVIDIA CUDA Tile Compiler (binary wheels)
Screenshots
View
Download
BrowseChangeLog
commit 60ac562d920bee7bc47818736879b856cf633b80
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 15:26:20 2026 +0200
dev-python/cuda-tile-bin: new package, 1.3.0
Tier 0 leaf for the vllm CUDA target packaging cycle. NVIDIA's CUDA
Tile compiler — required as an unconditional transitive dep by
flashinfer-python.
PyPI publishes wheel-only releases (no sdist) and the source for
NVIDIA's tile compiler is not public, so a -bin shape is the only
form available. Per-Python wheels are pulled per-impl and unpacked
via `installer`, mirroring runai-model-streamer-bin's shape but
extended for the cp311/cp312/cp313 split that NVIDIA ships.
Upstream caps at requires-python = "<3.14,>=3.10", so PYTHON_COMPAT
is python3_; pkgcheck's PythonCompatUpdate suggesting 3.14
is a known false positive against upstream's own ceiling.
Installs as the namespace package "cuda.tile" (not "cuda_tile") —
the dist name carries an underscore but the import name uses the
"cuda" namespace shared with cuda-bindings / cuda-pathfinder.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 15:26:20 2026 +0200
dev-python/cuda-tile-bin: new package, 1.3.0
Tier 0 leaf for the vllm CUDA target packaging cycle. NVIDIA's CUDA
Tile compiler — required as an unconditional transitive dep by
flashinfer-python.
PyPI publishes wheel-only releases (no sdist) and the source for
NVIDIA's tile compiler is not public, so a -bin shape is the only
form available. Per-Python wheels are pulled per-impl and unpacked
via `installer`, mirroring runai-model-streamer-bin's shape but
extended for the cp311/cp312/cp313 split that NVIDIA ships.
Upstream caps at requires-python = "<3.14,>=3.10", so PYTHON_COMPAT
is python3_; pkgcheck's PythonCompatUpdate suggesting 3.14
is a known false positive against upstream's own ceiling.
Installs as the namespace package "cuda.tile" (not "cuda_tile") —
the dist name carries an underscore but the import name uses the
"cuda" namespace shared with cuda-bindings / cuda-pathfinder.