0087: repeat GPU inference proving and add a classifier GPU candidate¶
Date: 2026-10-28. Status: owner-requested bounded experiment; results pending.
The owner requests a fresh GPU proving measurement and explicitly includes classifier GPU implementation. Preserve the published Mac classifier C1 and historical LLM L1 measurements, including their original source identities.
First repeat the unchanged generated-constraint TinyLlama experiment from decision 0079. Hypothesis: a fresh build and device run still produce the same 23 valid segment proofs with no CUDA constraint fallback. Baseline is L1's 53.941293719-second summed session timer and 155.888844427-second complete commands; neither value is a promised outcome. Freeze model, quantization, prompt/answer, witness archive, proof parameters, evaluators and CPU-then-CUDA order. Use the existing controller unchanged at main b7b8376. Mutable paths are this decision, the new numeric evidence and reporting documents; every run uses a fresh ignored directory. Preparation is reused and must be labelled.
The smallest offline controller suite passes before paid work; the worker must pass its real CUDA fixture before the complete pair. Acceptance requires all 23 segments, grid/token bindings, 69 byte-identical CPU/GPU artifact files, positive device constraint calls, zero CPU constraint fallback, fresh Linux and Mac proof-only honest/mutation checks, retained hashes/logs/telemetry and completed temporary-bucket cleanup. Missing checks or failed jobs stay failed. Report every timing boundary, build and transfer separately; nested stages must not be added to enclosing timers. One repeat remains descriptive and is not a latency distribution or a fresh inference-preparation measurement.
Standing provider approval and the owner's current request authorize bounded foreground work. The authenticated HF billing page shows USD 29.95 before launch; no active Jobs exist. Live prices are CPU Upgrade USD 0.0005/min and H200 USD 0.083333/min. Keep the unchanged USD 5 single-run cap and USD 4.66 timeout/startup/storage bound, one concurrent job, three jobs, no retries, top-ups, schedulers or private inputs. Build 3,600 s, H200 2,700 s, Linux verifier 1,800 s, each with 600 s additional creation/cancellation allowance. Retain the existing 200-GiB device-worker sampled RSS stop and resource guards.
Classifier GPU work is a separate candidate, not a relabelled CPU run on a GPU instance. First review compatible backend source/licence, then preserve the exact classifier relation and verify candidate proofs with the unchanged STWO 2.2.0 CPU verifier. Freeze a separate device plan and budget before its launch. Any unavoidable protocol/parameter change requires a separate identity and may not be presented as matched acceleration of C1. Keep the Mac default and its published measurements. No hiding, hardware or qualification claim.