commit 5e74c9bc230193b302a220c743e12e883d573336
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 21 21:25:08 2026 +0200

sci-ml/torchcodec: bump to 0.13.0

commit c0669f2fe6e737feaf4f33e1ba078b5c16584b10
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon May 18 19:46:22 2026 +0200

sci-ml/torchcodec: follow upstream github move to meta-pytorch org

pytorch/torchcodec permanently redirects to meta-pytorch/torchcodec.
Update HOMEPAGE, SRC_URI, and metadata.xml remote-id + bugs-to.
Same tarball content at the new URL (Manifest unchanged).

commit f5dd46925d8564c9d7c6c800fbb3d57f2dabc200
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 10:42:01 2026 +0200

sci-ml/torchcodec: new package, add 0.12.0

Decode/encode video and audio into PyTorch tensors via FFmpeg.
Required runtime dep of sci-ml/pyannote-audio (input pipelines).

C++ extension; setup.py drives a custom CMake invocation against
system media-video/ffmpeg (the libtorchcodec_*<MAJOR>.so libs
are built per FFmpeg major present and dlopen'd at runtime). Host
ffmpeg-8.1.1 produces just the FFmpeg-8 lib variant.

CUDAHOSTCXX is pinned to g++-15 unconditionally because Caffe2's
cmake config (pulled in via find_package(Torch)) calls
enable_language(CUDA) whenever it finds /opt/cuda, even on
USE=-cuda builds; nvcc would otherwise reject the active gcc-16.
See the internal CUDA-13 host-compiler note.

Upstream's BUILD_AGAINST_ALL_FFMPEG_FROM_S3 path vendors the FFmpeg
binaries to dodge GPL questions on redistributed wheels; we link
against system media-video/ffmpeg directly via the
I_CONFIRM_THIS_IS_NOT_A_LICENSE_VIOLATION opt-out, which only
matters for redistributed binary wheels (not source builds).

CPU-only build verified; CUDA path gated behind USE=cuda for now,
not yet exercised in this overlay.