mythikal03

#!/usr/bin/env bash
# benchmark_vllm.sh - streaming-aware vLLM throughput suite
# Usage: ./benchmark_vllm.sh <service|base_url> [short|long|multi|spec|all]
#   short  up to 1k single-shot decode
#   long   up to 16k single-shot decode
#   multi  4-turn growing context
#   spec   best-effort speculative decoding acceptance from /metrics
#   all    everything (default)
#
# Service mode reads /etc/systemd/system/<service>.service by default.