Ultra-low-latency inference on Groq's LPU hardware — when tokens-per-second is the product requirement.
