#!/bin/bash
# Read-only baseline requests. Never changes llama configuration or processes.
set -uo pipefail
root=/home/lab/bench/ornith/modular/shared
export CUDA_VISIBLE_DEVICES=''
/home/lab/venvs/vllm/bin/python "$root/bench.py" \
  --base http://127.0.0.1:8120/v1 --model ornith-local \
  --out "$root/llama-throughput-512" --lengths 4096 --concurrency 1,3 \
  --repeats 5 --task throughput --max-output 512 --seed 85009 \
  --gpu GPU-56b9b8f1-048e-1abc-752c-729769819c5b
result=$?
if [ "$result" -ne 0 ]; then exit "$result"; fi
python3 "$root/quality.py" --base http://127.0.0.1:8120/v1 \
  --model ornith-local --out "$root/quality-llama-final.json"
