一个补丁A patch
git apply autooptm.patch python your/entry.py …
标准 diff，不依赖我们。跑法跟以前一模一样。 A plain diff that depends on nothing of ours. You run it exactly as before.
给一个仓库地址，拿回一个补丁。 Give it a repo. Get back a patch.
只有在你自己的程序里真的变快了，才计费 Billed only when it is faster in your own program
git apply autooptm.patch python your/entry.py …
标准 diff，不依赖我们。跑法跟以前一模一样。 A plain diff that depends on nothing of ours. You run it exactly as before.
用你的命令、你的输入，落地前后各测多次取中位数。 Your command, your inputs, sampled before and after, median reported.
04 对比The alternatives
| AutoOptm | 自己用 Claude / 提示词Prompting an LLM yourself | 请人做性能优化Hiring a performance engineer | |
|---|---|---|---|
| 你要花多少时间盯着How much of your time it takes | 零。提交完就可以走开None. Submit and walk away. | 全程盯着 —— 它跑偏了只有你看得出来All of it. You are the only one who notices when it goes wrong. | 对齐、答疑、reviewBriefing, questions, review |
| 输出会不会变Does the output change | 不会 —— 超出容差的改动一律回滚No. Anything outside tolerance is reverted. | 没人检查Nobody checks | 看这个人的习惯Depends on the engineer |
| 谁验证它真的更快Who proves it is faster | 冻结的评测台，两臂多次采样取中位数A frozen harness, both arms, median of repeated samples | 没有人 —— 你自己搭台子测Nobody. You build the harness. | 同一个人既改也测The same person who wrote it |
| 没提速的话If it does not get faster | 不收费You are not charged | token 照付You still paid for the tokens | 工时照付You still paid for the time |
在你实际要跑的那张卡上优化 —— 优化结论不通用， 同一个改动在 A10 上是提速，在 H100 上可能是拖累。 Optimised on the card you actually run — conclusions do not transfer: the same change that wins on an A10 can lose on an H100.
05 谁在用Who this is for
不用走采购，不用改工作流——一个人、一个仓库地址就能开始。 No procurement, no workflow change — one person and a repo URL is all it takes.
业务模型每周都在重训，训练脚本是没人敢动的老代码。单步压一半，同一份预算多跑一倍实验——效果上去了，账单没涨。绩效评审里最硬的一行，是为公司省下的卡时：一个写得出来的数字，晋升答辩自己会说话。 The production model retrains every week on a script nobody dares touch. Halve the step time and the same budget runs twice the experiments — better model, same bill. The hardest line in a performance review is the GPU spend you saved, with a number on it. Promotions follow numbers.
组里的卡永远在排队，想法比机时多。同一张卡步子快一倍，等于凭空多出一张卡：deadline 前多跑一轮消融，论文表格里多出关键一列。缺的从来不是想法，是机时。 The lab queue is always full and ideas outnumber GPU hours. Twice the step speed on the same card is a second card out of thin air: one more ablation before the deadline, one more column in the paper. It was never ideas you were short of — it was machine time.
推理跑在自己租的卡上，延迟和吞吐就是毛利本身。同一张 A10 扛住两倍请求，单位成本当场减半——定价空间和利润率，都是这么挤出来的。 Inference runs on cards you rent yourself, so latency and throughput are the margin. The same A10 carrying twice the requests halves unit cost on the spot — that is where pricing room comes from.
比赛给每个人的是同一个截止时间，不是同一份算力。别人一晚上训 3 个模型，你训 6 个——榜上的名次，就是这么一格一格爬上去的。 A competition hands everyone the same deadline, not the same compute. While others train three models a night, you train six — leaderboard places are climbed one slot at a time.
06 定价Pricing
每次优化真实消耗 GPU 分钟和 token，所以按量计费而不是固定月费。系统先给出预期加速比和 credit 报价，你确认后才开跑——达不到承诺，credits 全额退回账户。Every run burns real GPU minutes and real tokens, so it's metered, not a flat subscription. You get a projected speedup and a credit quote before anything runs. Miss the target and the credits go straight back to your balance.
订阅赠送的 credits 按月清零；充值的 credits 永不过期。完整费率表见定价页。 Subscription credits reset monthly; topped-up credits never expire. Full rate card on the pricing page.
诊断阶段免费。看到加速比和报价之后，你再决定要不要付。Diagnosis is free. You decide whether to pay once you've seen the speedup and the quote.