Skip to content

ADR 034 — Candle / QLoRA stack upgrades

ADR 034 — Candle / QLoRA stack upgrades (deferred batch)

Section titled “ADR 034 — Candle / QLoRA stack upgrades (deferred batch)”
  • Mens / Populi training paths depend on Candle 0.9.x, qlora-rs (vendored patch), peft-rs, and transitive stacks (zip, CUDA kernels).
  • The workspace dependency audit shows duplicate majors (e.g. zip) that cannot be collapsed without coordinated Candle + HF ecosystem bumps.
  • GPU builds (mens-candle-cuda, NVCC, MSVC toolchain) require explicit CI coverage.
  • No ad-hoc Candle major bump inside manifest-normalization PRs.
  • Track one upgrade initiative with: MSRV check, cargo vox-cuda-release smoke, MENS eval matrix slice, and lockfile diff review for zip / rand / half transitive shifts.
  • Keep using workspace pins + patches documented in root Cargo.toml until the upgrade PR lands.

Proposed — execution gated on GPU CI sign-off and a green cargo check --workspace + targeted training smoke.