Kairo
Constrain LLM inference paths using measured RTX 5090 results and correctness checks.
SITE CAPTUREBased on information published on the service's website. Selection and review policy
WHAT IT DOES
What does it do?
Kairo manages CUDA Graph and eager execution results measured on an RTX 5090 alongside their workload conditions. It selects only exact measured buckets that pass correctness gates as paths in a fail-closed policy. It is not yet a general-purpose inference engine or a replacement for vLLM, CUTLASS, or cuBLAS.
KEY FEATURES
Key features
- Record results by model, quantization, workload, hardware, and software conditions
- Measure CUDA Graph and eager execution paths
- Apply only results that pass correctness gates to policy
- Route only exact measured workload buckets using a fail-closed approach
- Portable contract validation without a GPU
- Reproduce RTX 5090 experiments in a fixed environment