@EXM7777 Those jsonl logs are an unlabeled eval corpus: replayable, so your results stay comparable instead of vanishing. Do you keep full traces or just transcripts?
@ekzhang1 Once scheduling/parallelism live in a shared layer, tps may say more about the harness than the kernel. Would you expose those choices as eval config? (可复现评测 series)
@Austen Price floor only holds on one fixed task: pin 量化/并发/硬件/预热, re-rank by total cost, the order may flip. Which task was that number from? (可复现评测 series)
@mikenevermiss An 8-minute 'everything you need' breakdown is only as good as its eval setup. Which model version and task set were tested? (可复现评测 series)