rocm install sdk defaults to --format wheel, but the free-space preflight added in #171 only covers the tarball path.
install_wheel_runtime delegates the download to uv pip install, which fetches the ROCm wheels plus PyTorch, torchvision and torchaudio — in aggregate the larger of the two install paths. Nothing there checks free space beforehand, and nothing maps an out-of-space failure to a readable message, so a user on a nearly-full disk still gets a raw error partway through. That is the symptom #159 asked to eliminate, on the path most users take.
The tarball path could be handled with a Content-Length probe because it is a single known artifact. The wheel path has no equivalent cheap signal: the set of wheels and their sizes are only known after uv resolves the dependency graph, so a preflight needs either a resolve-only pass to sum the wheel sizes, or a fixed conservative floor for a known-heavy install, or a post-hoc mapping of uv's out-of-space output. Those are different enough from the tarball approach to be worth deciding separately rather than bolting onto #171.
Scoped out of #171 for that reason; #171 is "Refs #159" rather than "Fixes #159" as a result.
rocm install sdkdefaults to--format wheel, but the free-space preflight added in #171 only covers the tarball path.install_wheel_runtimedelegates the download touv pip install, which fetches the ROCm wheels plus PyTorch, torchvision and torchaudio — in aggregate the larger of the two install paths. Nothing there checks free space beforehand, and nothing maps an out-of-space failure to a readable message, so a user on a nearly-full disk still gets a raw error partway through. That is the symptom #159 asked to eliminate, on the path most users take.The tarball path could be handled with a
Content-Lengthprobe because it is a single known artifact. The wheel path has no equivalent cheap signal: the set of wheels and their sizes are only known afteruvresolves the dependency graph, so a preflight needs either a resolve-only pass to sum the wheel sizes, or a fixed conservative floor for a known-heavy install, or a post-hoc mapping ofuv's out-of-space output. Those are different enough from the tarball approach to be worth deciding separately rather than bolting onto #171.Scoped out of #171 for that reason; #171 is "Refs #159" rather than "Fixes #159" as a result.