InstinctFlash
A serving runtime for robotics vision-language-action models that runs custom fused kernels on NVIDIA Jetson Thor and desktop GPUs for faster real-time prediction.
InstinctFlash serves eight robotics model families, including LingBot-VA, LingBot-VLA and pi05, behind Python and WebSocket APIs, using purpose-built engines with fused Triton kernels tuned for FP8 precision and reduced sampling steps. The project reports large median prediction-latency reductions on Jetson Thor compared with a stock PyTorch baseline in real-robot tests, with no observed loss of task performance, and it also runs on RTX 4090 and RTX 5090 desktop cards for development off the robot.
Reported speedups vary by model and configuration: FP8 precision and fewer sampling steps are optional settings, and the project notes when a given benchmark was measured against an earlier version of its own kernels rather than the current release.
It is written in Python and licensed under AGPL-3.0.