Tagged "inference-pipeline-design"

• Prefill Is Compute-Bound, Decode Is Memory-Bound: Optimizing GPU Utilization for LLM Inference 16 April 2026