How we build Kooper AI: serving image models fast, keeping them safe, and running them on infrastructure we operate.https://kooper.ai/https://kooper.ai/blog/gpu-host-failure-tensorrt-driver/https://kooper.ai/blog/gpu-host-failure-tensorrt-driver/After a GPU host fault, our compiled TensorRT engines refused to load: the failure you get when a host's NVIDIA driver branch differs from the one the engines were built on. We recovered by starting a new pod on a host with the right branch (580, CUDA 13), which took a little over an hour because those GPUs were briefly out of stock. The lesson: a TensorRT engine is tied to its driver branch, so version the two together.Tue, 29 Sep 2026 00:00:00 GMTincidentTensorRTNVIDIA driversGPU inference