llama.cpp multislot scales throughput for batch benchmarks

Multislotting is a throughput game-changer for batch tasks, even if it doesn't solve the "snappiness" problem for single-stream chat.

• –Activating multiple slots (--parallel > 1) can significantly reduce the total time for massive benchmark runs by saturating hardware that a single stream leaves idle.
• –llama.cpp remains the better choice for mixed-memory setups, enabling high-quality Q6 quants that vLLM often forces down to INT4 or INT8 due to limited offloading.
• –The performance gap (170tps vs 400tps) highlights vLLM's architectural optimization for continuous batching, which llama.cpp only partially emulates through static slot partitioning.
• –For users prioritizing accuracy over raw speed, llama.cpp’s ability to run larger models in GGUF format remains its primary competitive moat.