Kache 1.2.1 caches CUDA builds that ran uncached before: every nvcc compile on Windows, and the compile lines CMake generates for CUDA on any platform.
cargo install kache --locked --version 1.2.1Fixed
nvcccompiles on Windows are cached. On Windows,nvcc -Mends continued lines with CRLF, and Kache read the leftover\as the drive root, so everynvcccompile ran uncached. Thanks to Matt Van Horn (@mvanhorn) for the fix, and to @cometkim for the report. #1559- CMake's CUDA compile lines are cached. Kache now models the flags CMake adds to every
nvcccompile:-forward-unknown-to-host-compiler,--generate-codeand, when a host compiler is set,-ccbin/--compiler-bindir. The host compiler is part of the key, and dependency tracking resolves headers with it. Thanks to @cometkim for the report. #1560
Other nvcc flags outside Kache's table, such as -lineinfo or --use_fast_math, still run uncached unless you allow-list them.
Packaging
- The
kache-serviceHelm chart is published only after the release succeeds, so a release tag that fails CI no longer leaves a chart behind. #1561
Upgrade notes
- If you added
-forward-unknown-to-host-compiler,--generate-codeor-ccbintoKACHE_CC_EXTRA_ALLOWLIST_FLAGSfor CMake, you can remove them.
Contributors
Full changelog: v1.2.0...v1.2.1