smolvm v1.7.7
What's Changed
Fixes
- Rebuild the macOS libkrun so guests reach userspace instead of exiting at boot by @BinSquare in #911
- Let a golden be forked more than once by retaining its RAM checkpoint by @BinSquare in #888
- Set TERM for interactive exec sessions so a guest shell's line editor can redraw correctly by @BinSquare in #887
- Retry transient KVM failures for fork bases by @BinSquare in #883
- Coordinate forked rollout workers by @BinSquare in #879
CUDA
- Harden CUDA clone recovery by @BinSquare in #891
- Harden post-training CUDA pools by @BinSquare in #886
- Make CUDA clone recovery transactional, preserve CUDA in embedded runtimes, and support the CUDA 13 runtime ABI by @BinSquare
Kubernetes
- Add the k3s deployment for the smolvm Kubernetes runtime, without clobbering existing containerd config by @BinSquare in #876
Docs and tests
- Document the Smolfile as a top-level section and show how to snapshot a configured machine into a reusable image by @BinSquare in #897
- Document the local archive, stdin and rootfs-directory forms of --image in the CLI help by @NickyHeC in #882
- Increment the test harness counters with arithmetic expansion so a suite that does not use the
|| trueidiom survives its first passing test by @NickyHeC in #874 - Bump the workspace to 1.7.5 by @BinSquare in #910
Note on 1.7.6
1.7.6 was withdrawn: its macOS build could not start a machine at all. Anyone on 1.7.6 should move to 1.7.7. Linux was unaffected.
Full Changelog: v1.7.5...v1.7.7