#!/bin/bash
set -euo pipefail
export CUDA_VISIBLE_DEVICES=GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f
export VLLM_NO_USAGE_STATS=1
export HF_HUB_OFFLINE=1
export TOKENIZERS_PARALLELISM=false
export OMP_NUM_THREADS=8
cd /home/lab/bench/ornith/vllm
test -f /home/lab/bench/ornith/shared/original-revision.txt
exec /home/lab/venvs/vllm/bin/vllm serve /home/lab/models/Ornith-1.5-9B-original \
  --served-model-name ornith-vllm \
  --host 127.0.0.1 --port 8140 \
  --quantization bitsandbytes --load-format bitsandbytes \
  --dtype bfloat16 --max-model-len 8192 \
  --gpu-memory-utilization 0.80 --max-num-seqs 2 \
  --max-num-batched-tokens 2048 --enforce-eager \
  --limit-mm-per-prompt '{"image":0,"video":0}' \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3
