We write a megakernel for your model and your GPU, so the whole model runs in one GPU launch, guaranteed at least 1.5× faster than your setup today.
We time your model on your GPU as it runs today.
We write a megakernel for that model and that GPU.
At start it answers a check question and is timed.
One command moves your model onto it.
yougori model optimize ENV --runner FOLDER
At least 1.5× faster on the same GPU, or you do not pay.
Set how fast your model writes today and watch the same answers written again with our megakernel, on the same GPU.
With one chat on each GPU, our megakernel writes at least 1.5 times as many tokens, so two GPUs do the work of three. Rented by the hour, that is about a third off the bill.
Every time your model starts, the megakernel proves itself before it serves a single request.
One GPU program that runs the whole model. Most inference launches a separate GPU program for every step of every layer, and the GPU waits between them. A megakernel runs every step in one launch, so the waits are gone.
On average our megakernel writes 1.68 times as many tokens per second as others on the same GPU for one chat, 1.40 times with two chats at once and 1.35 times with three.
Yes. If our megakernel does not run your model at least 1.5 times as fast as your current setup on the same GPU, for one chat, you do not pay for the megakernel work.
Dense and mixture-of-experts models on NVIDIA GPUs. Each megakernel is written for one model and one GPU family, so tell us yours and we tell you what to expect.
When your model starts, the megakernel is built for the GPU it is on, answers a check question and is timed before it serves. If it cannot start, your model is served the usual way and its status says why.
Inside your own inference, on your computers or on GPUs you rent.
YOUR MODEL. YOUR GPU.