Prefill
The first phase: the model reads the whole prompt at once (in parallel), builds the KV cache, and produces the first token.
The first phase: the model reads the whole prompt at once (in parallel), builds the KV cache, and produces the first token. Compute-bound; sets TTFT.