Address 2 high, 3 medium, and 1 low severity issues identified during
an audit of the ROCm support implementation. Also adds 24 new test cases
(21 Python + 3 shell) to cover edge cases and prevent regressions on
CUDA, CPU, macOS, and Windows pathways.
Bug fixes:
- install.sh: add validation guard for _rocm_tag to reject malformed
output (e.g. "rocm." from garbled amd-smi) that would produce bogus
PyTorch index URLs
- install.sh: strip Debian epoch prefix (e.g. "2:6.2.0") from
dpkg-query output before parsing ROCm version
- install_python_stack.py: add timeout=30 to the subprocess probe that
checks whether torch is already GPU-enabled, preventing hangs
- install_llama_prebuilt.py: log a warning that the ROCm prebuilt is
compiled for ROCm 7.2 and may fall back to source build on other
ROCm versions
- install_llama_prebuilt.py: log a message when Windows HIP prebuilt
is not found before silently falling through to CPU
- tests/sh/test_get_torch_index_url.sh: replace no-op test 18 (could
not actually test /opt/rocm path) with 3 real tests for malformed
amd-smi output (empty version, "N/A", trailing text)
Additional changes:
- Export IS_ROCM from studio/backend/utils/hardware/__init__.py so
downstream code can import it directly from the package
Test results: 23/23 shell, 68/68 Python, 205/206 full suite (1
pre-existing failure unrelated to this change).
be0fd8b822