Profile
We measure your models on your GPUs with your traffic, before we change anything.
Tune
We tune each model to your GPUs: precision, attention, memory use and kernels.
Verify
We compare each change with the original model's output on your hardware. A change that fails does not ship.
Reassess
We measure again when your traffic, models or GPUs change, and the loop starts again.
Our auto-tuning inference stack on your GPUs.
We run our stack on your own GPUs, tune it to your diffusion models and verify every change against the original output.
Where it runs and what happens to your data
- The stack runs on your GPUs, in your environment.
- Your models, your prompts and your outputs stay on your hardware.
- A signed agreement with you sets the scope of the work, what we may access and how long we keep any records.
Who it is for
- Teams that run open-source video, image or audio models in production.
- Your margin depends on what inference costs.
- You are past the prototype, and your GPU bill is large enough that a lower cost per output pays for an agreement.
- You want to keep your models, your data and your traffic on your own hardware.
What you get
- Our inference stack, installed on your GPUs and tuned to your models.
- More outputs from the same GPUs, at a lower cost per output.
- A record for every change: what changed, how it compares with the original output and how fast it runs.
- Ongoing tuning as your traffic, models and hardware change.
What we need from you to start
- The models you run, and the outputs they make.
- The GPUs you run them on.
- A sample of your traffic, or a description of it.
- Your quality limits and your latency target.
- A contact on your team who can give us access under the agreement.
A record for every change.
Each change we make gets one record. You see what changed and how it compares with the original output before it serves your traffic.
- model
- The model the change applies to.
- change
- What we changed.
- reference_check
- How the output compares with the original model's output.
- quality_metric
- The score we use for output quality.
- speed_metric
- How fast the changed model runs on your GPUs.
- decision
- Whether the change ships.
How the stack works.
See the whole stackRun our stack on your own GPUs.
We tune it to your diffusion models and verify every change against the original output.