Most LLM inference runtimes ignore physical deadlines. A new CUDA-based runtime for Qwen2.5-Coder-1.5B-Instruct refuses to start if it can't finish within a 33ms robot control cycle, evicts KV cache by meaning (cosine similarity) instead of age, and uses a lock-free double buffer. It avoids VRAM overflow from continuous vision streams, prevents control-loop deadline misses, and handles a 60Hz camera that outpaces reasoning.

Built entirely in hand-written CUDA — no cuBLAS or libtorch — on a single cloud NVIDIA Hopper GPU (sm_90) with an 8GB VRAM ceiling modeled in code, not measured on edge devices like Jetson. The architecture separates perception, admission control, and reasoning: camera feeds into a lock-free buffer, an admission controller checks feasibility against the 33ms deadline plus 2ms safety margin, then passes frames to a hand-written transformer. The Github repo is https://github.com/Anubhab Banerjee/vla-edge-backend1.

The camera runs at 60Hz (16.7ms/frame) and the model must decide and act within 33ms — faster than a human blink. This is an architecture, not a benchmark.