01 自动性能优化Automated performance engineering
提交一个仓库，几十分钟后拿回改好的代码和一个经过验证的加速比。训练衡量每步墙钟时间，推理衡量单位时间吞吐。 Submit a repo. Get back a mergeable diff and a verified speedup — measured as per-step wall clock for training, throughput for inference.
先报价再执行 · 未达到承诺加速比全额退回 credits Quoted before it runs · Credits refunded in full if the target speedup is missed
02 工作原理How it works
修复手段本身是公开知识，就那十几种。门槛在两端：给定一份性能数据，判断该用哪一种、预期能拿到多少；以及事后证明改动既变快了，也没有改坏。 The fixes themselves are public knowledge — there are maybe fifteen of them. The hard parts are at either end: reading a profile and knowing which one applies and what it's worth, then proving afterwards that the change made it faster without breaking anything.
复现环境、跑通基线、记录 CPU 与 CUDA 两条时间轴。不猜，只测。Reproduce the environment, run the baseline, record both the CPU and CUDA timelines. No guessing — only measurement.
由确定性代码完成，不交给模型判断。输出瓶颈类型、源码位置、每项的预期收益上界。Done in deterministic code, not left to the model. Emits bottleneck class, source location, and an upper bound on the payoff of each.
agent 只拿到结论并执行，一次改一处，改完立刻过门禁。默认不碰数值精度。The agent receives conclusions only and executes them, one change at a time, each gated immediately. Numerical precision is never touched by default.
测量条件锁死，报告分布而不是单值。任一门禁失败，整个改动回滚并写进报告。Measurement conditions frozen; we report a distribution, not a single number. Either gate failing rolls the change back and writes up why.
03 双门禁Two gates
确认改动真的带来提升，并且提升来自诊断指出的那一处。快了但瓶颈没变，通常是噪声而不是成果。测量条件完全锁死，报告分布而不是单值。Confirms the change actually helped, and that the gain came from the place the diagnosis pointed at. Faster with the bottleneck unmoved is usually noise, not a result. Conditions are frozen and we report a distribution, never a single number.
确认行为没有改变：输出一致、训练轨迹一致、资源占用不上升、结果可复现。任何一项不通过，整个改动回滚，并把原因写进报告。Confirms behaviour is unchanged: identical outputs, identical training trajectory, no rise in resource usage, reproducible results. Any failure rolls the whole change back and the reason goes in the report.
门禁的工程量必须大于优化本身。用户不会自己验证，出一次事故的代价远大于多拿几个点的加速。这也是最难被复制的部分——手段可以照抄，验证闭环需要时间和运行量才能磨出来。The gates must be more engineering than the optimisations are. Users won't verify for themselves, and one incident costs far more than a few extra points of speedup. It's also the hardest part to copy: the techniques are public, but the verification loop only comes from time and volume.
04 相关项目The landscape
| 维度Dimension | 现有工具Existing tools | AutoOptm |
|---|---|---|
| 输入Input | 一个 nn.Module 的单次 forwardA single forward pass of one nn.Module | 整个仓库，含依赖、数据、启动脚本The whole repo — deps, data, launch scripts |
| 输出Output | kernel 文件A kernel file | 可直接合并的 diffA mergeable diff |
| Profile | 只看 GPU kernelGPU kernels only | CPU 与 CUDA 时间轴一起看CPU and CUDA timelines together |
| 优化目标Objective | 单算子的 GPU 时间GPU time of one operator | 端到端墙钟或吞吐End-to-end wall clock or throughput |
| 算子覆盖Operator coverage | LLM 类算子，无卷积LLM operators; no convolution | 不限，因为不碰算子Unbounded — we don't touch operators |
Amdahl 定律在它们那里只算 GPU kernel 时间。一个 GPU 利用率只有 40% 的训练循环，把所有算子优化到理论峰值，端到端也快不了多少。而这正是绝大多数真实开源项目的状态。For them, Amdahl's law only counts GPU kernel time. Take a training loop running at 40% GPU utilisation, optimise every operator to theoretical peak, and end-to-end barely moves. That 40% is the state of most real open-source projects.
05 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.