Skip to content

Publish verified Windows CUDA release sidecars - #37

Draft
leehack wants to merge 4 commits into
mainfrom
experiment/windows-cuda-prebuilt
Draft

Publish verified Windows CUDA release sidecars#37
leehack wants to merge 4 commits into
mainfrom
experiment/windows-cuda-prebuilt

Conversation

@leehack

@leehack leehack commented Aug 17, 2026

Copy link
Copy Markdown
Owner

Summary

  • replace the slow Windows x64 source-built CUDA matrix lane with verified upstream llama.cpp CUDA 12.4 and CUDA 13.3 sidecars
  • bind each sidecar to the exact native release tag, llama.cpp commit, and same-run ggml-base.dll SHA-256
  • version backend/runtime filenames so CUDA 12 and CUDA 13 can coexist without mixing dependency families
  • publish the two verified sidecars through the normal release artifact and assets.json provenance flow
  • retain the non-publishing experiment path for fast packaging-contract validation

The dependent Dart consumer is leehack/llamadart#340. Runtime downloading is intentionally out of scope and tracked in leehack/llamadart#339.

Why

The source-built CUDA lane dominated Windows release time even though upstream llama.cpp already publishes matching CUDA binaries. The previous experiment proved that the exact b10453 assets can be repackaged and loader-smoked without a GPU in about four minutes; this change moves that measured path into the owning native release workflow while preserving an exact compatibility boundary with the locally built core.

Validation

  • 18 CUDA contract/package/verification unit tests passed
  • release provenance contracts passed
  • workflow YAML parse passed
  • actionlint -shellcheck= passed
  • git diff --check passed
  • non-publishing exact-head experiment: https://github.com/leehack/llamadart-native/actions/runs/32041342903
    • exact b10453 provenance and asset digests verified
    • CUDA 12.4 and 13.3 fatbins inspected
    • CUDA 13 loader smoke passed without a GPU
    • CUDA 12 correctly reported the missing NVIDIA driver on the GPU-less runner

Publication gates

This PR remains draft. No native release has been published and no physical NVIDIA GPU was available. Before publication, validate representative CUDA 12/13 hardware and real-model inference, including kernel/PTX execution and the open MTP acceptance, CUDA lockup, DFlash concurrency, DSpark VRAM-leak, EAGLE, and failed-state-restore reports.

Tracked by #64.

@leehack leehack changed the title ci: add verified Windows CUDA prebuilt experiment Publish verified Windows CUDA release sidecars Aug 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant