Three files conflicted, all on the gfx906 (MI50 / Radeon VII) work that landed
on main next to this branch's comment pass and torch 2.11 default rollout.
install.sh: kept main's _is_gfx906_bnb_skip / _gfx906_bnb_snapshot / _prune
helpers and the _gfx906_bnb_snapshot call in the migrated path, with this
branch's condensed comments. The gfx906 reroute keeps its literal
TORCH_CONSTRAINT="torch>=2.4,<2.11.0": the rocm6.3 index it routes to tops out
at torch 2.9.x, so it is now a deliberate narrower cap rather than a restatement
of the default, and the comment says so.
studio/install_python_stack.py: kept main's _runtime_is_gfx906 detection, the
rocm6.3 legacy override and the bnb skip/prune branch, again with the condensed
comments. _ROCM_TORCH_PKG_SPECS["_default"] stays at <2.11.0, which is what the
gfx906 path installs from.
tests: the two sides asserted opposite things about hard-coded torch ranges in
install.sh. This branch composes the default from _TORCH_CEILING and asserted
zero literals; main asserted the literal exists and only on a TORCH_CONSTRAINT=
assignment. Merged into the rule both wanted: a literal range is allowed only on
a curated per-index TORCH_CONSTRAINT= override, never on a pip/uv install line,
plus a separate check that the gfx906 reroute keeps its sub-2.11 cap. The same
resolution is applied to tests/sh/test_torch_constraint.sh, which asserted the
zero-literal count and would otherwise have gone red on main's new assignment.
ac8bc5f9a3