dev-python/nvdiffrast
Modular primitives for high-performance differentiable rendering
Screenshots
View
Download
BrowseChangeLog
commit cfa3b4e0a2fd0fe3fc90c66156a6cebec889be14
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 16 18:23:29 2026 +0200
dev-python/nvdiffrast: add a CUDA arch fallback
The ebuild probed the visible GPU and otherwise left
TORCH_CUDA_ARCH_LIST unset, on the assumption that cpp_extension would
pick a fallback list. It does not: with no device to query it assembles
an empty architecture list and then indexes it, so the build ends in
IndexError. That hits any build on a machine with no GPU, or where the
devices are absent, such as a container or a build host.
Fall back to compute capability 7.5 with PTX and say so. PTX is what
makes the fallback useful: such a build is by definition running
somewhere without the target GPU, and a bare 7.5 cubin would not load
on a newer card. An explicit list is still honoured.
commit 924583f1c3844867bd1dca8432af3091cbb87725
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 12 16:15:41 2026 +0200
dev-python/nvdiffrast: enable Python 3.14
The extension uses Torch C++ bindings without a Python upper bound, and
its dependency stack supports Python 3.14.
commit faab5038a03707e10c7bf5fcf17d77a60fa8798d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 05:23:18 2026 +0200
dev-python/nvdiffrast: trim 0.4.0 comment prose
commit 0aa66535b261e50e33fc55a90d578787842f5621
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Aug 17 17:13:56 2026 +0200
dev-python/nvdiffrast: build for the native GPU arch
The cuda compile forced TORCH_CUDA_ARCH_LIST to a fixed 8.0 8.6 8.9 9.0
list, so every host built four Ampere-through-Hopper targets regardless
of its GPU, and a machine outside that window got no runnable kernel by
default. Probe the native compute capability with nvcc's device query
and build only that; an explicit TORCH_CUDA_ARCH_LIST still wins, and
when no GPU is visible at build time the value is left unset so torch
falls back to its full arch list.
commit d19a02571b208ca16d173a545b1425ef6bfba12d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 17:36:39 2026 +0200
dev-python/nvdiffrast: enforce build requirements
The 0.4.0 pyproject requires setuptools 64 or newer, wheel, and ninja. Declare the
host build requirements explicitly so isolated extension builds receive the backend
toolchain upstream specifies.
commit f411a3096d979346a26553f43bd8a9ef42631990
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 10:08:53 2026 +0200
dev-python/nvdiffrast: new package, add 0.4.0
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 16 18:23:29 2026 +0200
dev-python/nvdiffrast: add a CUDA arch fallback
The ebuild probed the visible GPU and otherwise left
TORCH_CUDA_ARCH_LIST unset, on the assumption that cpp_extension would
pick a fallback list. It does not: with no device to query it assembles
an empty architecture list and then indexes it, so the build ends in
IndexError. That hits any build on a machine with no GPU, or where the
devices are absent, such as a container or a build host.
Fall back to compute capability 7.5 with PTX and say so. PTX is what
makes the fallback useful: such a build is by definition running
somewhere without the target GPU, and a bare 7.5 cubin would not load
on a newer card. An explicit list is still honoured.
commit 924583f1c3844867bd1dca8432af3091cbb87725
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 12 16:15:41 2026 +0200
dev-python/nvdiffrast: enable Python 3.14
The extension uses Torch C++ bindings without a Python upper bound, and
its dependency stack supports Python 3.14.
commit faab5038a03707e10c7bf5fcf17d77a60fa8798d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 05:23:18 2026 +0200
dev-python/nvdiffrast: trim 0.4.0 comment prose
commit 0aa66535b261e50e33fc55a90d578787842f5621
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Aug 17 17:13:56 2026 +0200
dev-python/nvdiffrast: build for the native GPU arch
The cuda compile forced TORCH_CUDA_ARCH_LIST to a fixed 8.0 8.6 8.9 9.0
list, so every host built four Ampere-through-Hopper targets regardless
of its GPU, and a machine outside that window got no runnable kernel by
default. Probe the native compute capability with nvcc's device query
and build only that; an explicit TORCH_CUDA_ARCH_LIST still wins, and
when no GPU is visible at build time the value is left unset so torch
falls back to its full arch list.
commit d19a02571b208ca16d173a545b1425ef6bfba12d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 17:36:39 2026 +0200
dev-python/nvdiffrast: enforce build requirements
The 0.4.0 pyproject requires setuptools 64 or newer, wheel, and ninja. Declare the
host build requirements explicitly so isolated extension builds receive the backend
toolchain upstream specifies.
commit f411a3096d979346a26553f43bd8a9ef42631990
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 10:08:53 2026 +0200
dev-python/nvdiffrast: new package, add 0.4.0