Skip to content

Releases: thedhruvhegde/cudackpt

Release list

cudackpt 0.3.0

cudackpt 0.3.0 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 29 Jun 18:17
v0.3.0
ae1f4d3

cudackpt 0.3.0

Reliability and operability release — restore metrics, config-driven RPC auth, deep health probes, packaging dependencies, and shim socket cleanup.

Install

sudo dpkg -i cudackpt_0.3.0_amd64.deb

Or build from source: git checkout v0.3.0 && make && sudo make install

Highlights

  • Restore failure metrics on preflight, materialize, CRIU, and shim timeout paths
  • rpc_secret in /etc/cudackpt.conf for orchestrator RPC auth
  • Agent GET /health?deep=1 and cudackpt serve health over control socket
  • .deb declares Depends: criu and recommends NVIDIA driver packages
  • Shim unlinks per-PID IPC socket on process exit
  • Graceful shutdown for cudackpt metrics

See CHANGELOG.md for full notes.

Assets 3
Loading

cudackpt 0.2.0

cudackpt 0.2.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 28 Jun 21:34
v0.2.0
184f7dc

cudackpt 0.2.0

Production-hardening release — configurable run directories, authenticated RPC tests, observability, packaging, and CI improvements since v0.1.0.

Install

sudo dpkg -i cudackpt_0.2.0_amd64.deb

Or build from source: git checkout v0.2.0 && make && sudo make install

Highlights

  • Shim honors CUDACKPT_RUN_DIR with group-owned 0660 sockets
  • CLI: watch --until-running, shell completion, cudackpt serve
  • Agent: /health probe + GC error metrics
  • Hermetic integration tests for auth, retry, rollback, and watch
  • golangci-lint + race detector in CI
  • vectoradd checkpoint example under examples/

See CHANGELOG.md for full notes.

Loading

cudackpt 0.1.0

cudackpt 0.1.0

Choose a tag to compare

View all tags
@thedhruvhegde thedhruvhegde released this 25 Jun 17:23
v0.1.0
574517e

cudackpt 0.1.0

First public release — checkpoint and restore single-GPU CUDA processes on Linux without modifying your application.

What it does

cudackpt combines an LD_PRELOAD CUDA Driver API shim with CRIU process checkpointing. The control plane snapshots GPU allocations, streams, and modules into a versioned on-disk image, then restores CPU and GPU state together.

Install

From source (recommended on GPU hosts):

git clone https://github.com/DDVHegde100/cudackpt.git
cd cudackpt
git checkout v0.1.0
make && sudo make install

Debian/Ubuntu package (attached when available):

sudo dpkg -i cudackpt_0.1.0_amd64.deb

Highlights

  • Transparent CUDA interception via LD_PRELOAD
  • Versioned GPU images with CRC32C integrity checks
  • Optional compression, sparse pages, dedup, and delta snapshots
  • Full CLI: checkpoint, restore, rollback, inspect, validate, report
  • Prometheus metrics via cudackpt agent and systemd units
  • Optional RPC shared-secret auth (CUDACKPT_RPC_SECRET)

Requirements

Linux x86_64 · NVIDIA CUDA 12.x driver · CUDA toolkit · Go 1.22+ · CRIU 3.x

Documentation

Known limitations

Single GPU only · no MIG/NCCL/CUDA graphs · validate on your stack before production use

Loading