Prefill Stage

The first stage of large language model inference, in which all input tokens are processed in parallel and the KV Cache is generated. This stage is compute-intensive and is well suited for execution on high-throughput GPUs.