Skip to content

0078: native CUDA build and a complete measurement window

Date: 2026-10-26. Status: bounded follow-up experiment; no backend adoption.

The initial GPU measurement verified both small fixture digests, the complete SIMD case and 22 CUDA transformer segments. Its 1,950-second worker deadline killed the output-head command before it produced any head proof. Preserve that run as failed/incomplete, not a complete LLM proof or an accepted full-case speedup. No verifier/evaluator change follows from it.

The first CUDA fixture took 218.745 seconds. Binary inspection found sm_89 PTX despite a requested sm_90 target; this observation alone does not identify all native cubins. The corrected builder uses an isolated Cargo target directory, a plain CMake architecture setting, and records cuobjdump ELF/PTX inventories. It requires native sm_90 and rejects sm_89 before admitting the new CUDA binary. This separates it from the SIMD build's cached CUDA library.

Archive compression uses level one: raw proof/statement/claim bytes must remain identical, while compression time and archive size may change. This changes packaging, not the inference relation or proof configuration. Preserve the old level-nine archives and timings. Explicitly retain the initial fixture-only shape-admission failure as well; the corrected fixture check never relaxes the complete LLM enrollment check.

Freeze a new run: one same-host pair, unchanged inputs/model/proof parameters, four Rayon threads, H200 provider timeout 2,700 seconds, worker deadline 2,550 seconds, per-command deadline 600 seconds and sampled RSS stop 200 GiB. CPU build and proof-only verification remain at 1,800 seconds each. The new agent-selected single-run ceiling is $5 under the existing standing provider/credit authority; the earlier $4-capped attempts remain separate immutable records. No top-up or automatic retry is permitted. Recheck prepaid credits/rates before launching. The inspected full timeout/startup/storage bound is $4.64 with a fresh build.

The profile is cuda-native-sm90-host-generation. Generation/AIR evaluation remain on the host, with no generated-kernel install. Acceptance still requires all 23 segments, full grid/token verification, byte-identical CPU/CUDA artifacts, fresh CPU proof-only mutations and the existing native Mac verifier. No partial proof or failed job counts as a passing full measurement.

After the passing measurement, controller preflight moves Apple Silicon/Rust prerequisites before SDK initialization and paid work. Four rejection tests cover unsupported architectures and missing Rust. The measured source remains commit a81c10a324fe7e849a40cf1f0caa9338e1a0cc3d, matching all 26 frozen source hashes. The prover/worker, evaluator, inputs and numeric records are unchanged; this admission fix requires no additional paid proof campaign.