install.sh's CUDA branch allows torch>=2.4,<2.12.0 and the cu12x/cu13x wheel
indexes top out at torch 2.11.0, so a fresh NVIDIA install lands on torch 2.11.
studio/backend/requirements/extras-no-deps.txt pinned torchcodec==0.10.0, which
upstream pairs with torch 2.10 exactly, and torchcodec publishes no
Requires-Dist: torch, so pip cannot notice. The guard that exists to catch this,
_torchcodec_version_mismatch_hint, looks the torch minor up in
_TORCH_TORCHCODEC_MINORS and returns None on a miss, and the table had no "2.11"
row: it was silent on precisely the version where the mismatch now happens.
Add the 2.11 row (mirrored into scripts/notebook_validator.py), add an
audio-torch211 extra pinning the 0.11 line, and select torchcodec at install
time from the venv's torch minor instead of a flat pin. A requirements file
cannot branch on the installed torch version, so _select_torchcodec_spec follows
the existing _select_torchao_spec pattern. It runs after the final torch repair,
which can still move torch onto another minor.
torchcodec 0.11 is the release upstream pairs with torch 2.11 exactly and the
newest one published on the cu128 index install.sh resolves torch 2.11.0 from,
since 0.12 dropped CUDA 12.8. From 0.12 torchcodec is ABI stable against torch
>= 2.11, so that half of the matrix is open ended and cannot be written as a
finite set of minors; both checkers now exempt torchcodec >= 0.12 on torch >=
2.11 rather than flagging a combination upstream supports, and torch 2.12+
installs take the open >=0.12.0 floor instead of the 2.11-only 0.11 line.
Also bump the installer step total for the new step and keep torchcodec inside
the security audit set now that it no longer comes from a requirements file.
ef240d5a61