v0.20.1: CUDA 13.3.1

Highlights

CUDA 13.3.0 to 13.3.1 (nvcc V13.3.33 to V13.3.73), the newest toolkit that exists; there is no 13.4 or 14.x. The release string stays 13.3, so CI's nvcc assertion and the build-cache keys are untouched. Bumped in both Dockerfile stages, both CI images, the ncu profiling image, the offline fallback of the PTX-survey image resolver and the standalone kernels' compile lines.

--set with a key that does not exist is an error, not a warning. A typo silently measured the default instead: the AWQ reproduction harness passed --set gemm.deterministic=true where the key is runtime.deterministic_gemm, so its scoring runs never got the determinism they asked for. With the key fixed the published numbers reproduce unchanged. An unknown key in imp.conf stays a warning, because a config file may outlive the build that understood every key in it.

imp-quantize --calib states where it is validated and where it is not. A model too big to run in BF16 can be calibrated off any quantization of itself, since the statistics are keyed by layer and tensor kind rather than by checkpoint. That produced the first --calib result on Qwen3-14B and it is negative: perplexity 9.9252 uncalibrated against 12.6016 and 12.2853 calibrated, from two independently produced twins. It still helps on Qwen3-0.6B and 1.7B.

Also in here

• A dead docs/… pointer in a code comment is now a gate failure; the 38 comments citing design memos deleted months ago were repointed at what survived

Gate

make verify-fast on the tagged tree, Qwen3-8B-Q8_0 on one RTX 5090: three runs on the new image read 285.65 / 287.95 / 287.95 tok/s decode (the first is cold clocks), median 287.95 against 288.38 on CUDA 13.3.0 and a pinned 287.19 baseline. Measured perf-neutral rather than assumed, so no re-pin.

No breaking changes. Full detail: CHANGELOG · 3 PRs since v0.20.0.