NVIDIA co-designs model attention for faster long-context inference

As agentic and long-context workloads grow, longer contexts push attention to consume a larger share of inference time; NVIDIA's post discusses co-designing model attention for fast, interactive long-context inference.