ADR-0027: Non-conservative image pins with experimental toolchain flags¶

Status: Accepted
Date: 2026-04-17
Deciders: Lusoris, Claude (Anthropic)
Tags: ci, cuda, sycl, build, supply-chain

Context¶

The fork's container base images and toolchain pins had been conservative (N-2 style) by default. The user directed them to track the newest stable — especially CUDA — and to enable "experimental features" where they unlock measurable throughput. A subsequent clarification narrowed "experimental" to feature flags on stable compilers, not preview compilers themselves: "with experimental i meant features not version so it should be >13.2 on cuda".

Decision¶

Container base images and toolchain pins track the latest stable release. Specifically: CUDA base tracks NVIDIA's newest devel-ubuntu24.04 tag (2026-04-17: 13.2.0; bump on every stable release); Ubuntu 24.04 LTS is the floor; FFmpeg uses the current n<major>.<minor> tag; oneAPI uses the current major.minor. Dev and prod Dockerfiles track the same major.minor unless there's a hard blocker. Enable experimental toolchain feature flags where they give measurable throughput or unblock modern C++: nvcc --expt-relaxed-constexpr, --extended-lambda, --expt-extended-lambda; SYCL/DPC++ -fsycl-unnamed-lambda, -fsycl-allow-func-ptr, -fsycl-device-code-split=per_kernel; Blackwell gencode (sm_120). Not preview/beta CUDA branches. Digest pins retained for reproducibility on the pinned release, but bumped aggressively on every stable release rather than N-2.

Alternatives considered¶

Option	Pros	Cons	Why not chosen
Conservative N-2 pins	Maximum stability	Leaves speed on the table; user explicitly rejected	Rejected
Preview CUDA branches	Newest features	Sacrifices reproducibility without matching speed win	Rejected per user clarification
Latest stable + feature flags (chosen)	GPU perf without instability	Must track NVIDIA release cadence	Matches user's clarified scope

Rationale note: the user's two messages combined — "don't make the img pins conservative... use newer versions... especially cuda and even use experimental features (thats where the speed is)" and the clarification about "features not version". So "experimental" = toolchain feature flags on stable CUDA ≥13.2 / stable DPC++.

Consequences¶

Positive: GPU builds get Blackwell gencode and relaxed-constexpr for free; container tracks current CUDA.
Negative: monthly-to-quarterly bump cadence; CI must re-validate on every bump.
Neutral / follow-ups: commit 8a995cb0 tracks the 13.2 bump.

References¶

Source: req (user: "dont make the img pins conservative... use newer versions... especially cuda and even use experimental features" + "with experimental i meant features not version so it should be >13.2 on cuda")
Related ADRs: ADR-0022