内测期间:充值额外赠送 30% credits,订阅额外赠送 50%–70%。有任何反馈,欢迎到支持页提交。 Closed beta: top-ups come with 30% extra credits, subscriptions with 50–70% extra. Feedback is welcome on the support page.

提交任务Submit a job

01代码来源Where the code comes from

把整个目录打成 .zip,私有代码走这里。脚本里的入口写包内相对路径。上限 64 MB;数据不要打进包里,请使用下方的数据集入口。 Zip the whole directory — this is the path for private code. In the script, name the entry by its path inside the zip. 64 MB max; data goes through the dataset field below, not in here.

支持 .zip / .tar / .tar.gz,上限 1 GB。运行前解压到此路径,脚本里照常引用(例如 --data ./data)。数据只进入沙箱,不进入调度机、报告或补丁,7 天后自动删除。 .zip / .tar / .tar.gz, 1 GB max. Unpacked at this path before the run, so the script refers to it as usual (e.g. --data ./data). It only ever enters the sandbox — never the dispatcher, the report or the patch — and is deleted after 7 days.

02怎么运行你的程序How your program runs

点击行号指定测速行,再点一次恢复自动Click a line number to pin the measured line; click again for automatic

像在自己机器上一样写:装依赖、下数据集(wget / curl / huggingface-cli)、设环境变量,最后一行是你平时跑的命令。 加速比只对测速行端到端复测(默认是最后一条 python / torchrun / accelerate … 开头的命令); 它之前的部分作为环境准备,只执行一次、不计时,运行时没有显卡。 Write it as you would on your own machine: install dependencies, fetch the data (wget / curl / huggingface-cli), set environment variables, and end with the command you normally run. The speedup is measured end to end against the measured line only (by default the last command starting with python / torchrun / accelerate …); everything before it is setup, run once, not timed, and without a GPU.

高级选项Advanced

留 auto 则在 RTX 4090 上运行,多数模型够用且便宜;放不下再在这里选更大的卡。 加速比相对同一张卡的基线计算,换大卡不会让数字虚高,只会更贵。 程序本身不用显卡(CSV 处理、解码、纯 numpy 流水线)请选「CPU only」:不收机器费,只按 token 计价。 Auto runs on an RTX 4090, which is cheap and fits most models. Pick a bigger GPU here if yours does not fit. If your program never touches a GPU (a CSV tool, a decoder, a pure-numpy pipeline) pick "CPU only": no machine fee, tokens priced as usual. The speedup is measured against that same GPU's own baseline, so a larger GPU does not inflate the number — it only costs more.

同一个仓库可以用两个模型各跑一次做对比,token 计价口径不变。 The same repository can be run once on each to compare; tokens are priced the same way.

03开始方式How to start

几秒钟的静态扫描,不扣费、不执行代码。看到预估加速比之后,再决定是否继续。 A static scan that takes seconds. Nothing is charged and none of your code runs. You decide whether to continue once you have seen the projected speedup.

先免费估算(几秒钟,不执行代码)。点「继续」时扣 2 credits 分析押金;没有加速时退回,失败在我们这边时同样退回(基础设施故障、超时、重新部署中断、环境搭建失败、找不到入口实际运行的代码、程序显存超出我们最大的显卡)。拿补丁时全额抵扣,运行按实际用量计费。 A free estimate first (seconds, none of your code runs). Continuing takes the 2-credit analysis deposit, returned when no speedup is found or when the failure is ours: infrastructure failure, a timeout, a redeploy interrupting the run, the environment failing to build, an entry point that dispatches to code we cannot find, or a program that needs more GPU memory than our biggest GPU — and credited in full against the patch; the run bills on actual usage.

账户Account

用量包 credits 自购买起 24 个月内有效;订阅 credits 在每个周期结束时清零。 Usage-pack credits are valid for 24 months from purchase; subscription credits reset at the end of each period.

充值与订阅Add credits & plans

读取中…Loading…

最近账目Recent activity

读取中…Loading…

邀请Invite

读取中…Loading…

邮件通知Email notifications

读取中…Loading…

完整账单明细 →Full statement → · 全部方案与费率All plans & rates