#!/bin/bash
set -euo pipefail
export CUDA_VISIBLE_DEVICES=GPU-5f32faf3-97b9-895e-0a9e-ffa0f51bca3f
export VLLM_NO_USAGE_STATS=1 HF_HUB_OFFLINE=1 TOKENIZERS_PARALLELISM=false OMP_NUM_THREADS=8
export TRITON_CACHE_DIR=/home/lab/bench/ornith/modular/vllm/cache/triton
if [[ "${ORNITH_POWER_LABEL:-}" == 250W* ]]; then
  if ! nvidia-smi -i "$CUDA_VISIBLE_DEVICES" --query-gpu=power.limit --format=csv,noheader,nounits | awk '$1>0 && $1<=250.5 {ok=1} END {exit !ok}'; then
    printf '%s\n' 'This profile was tested at250W and requires GPU2 cap<=250W; refusing startup. Hardware was not changed.' >&2
    exit 78
  fi
fi
ROOT=/home/lab/bench/ornith/modular/vllm
cd "$ROOT"
MODEL=${ORNITH_MODEL:?ORNITH_MODEL must point at the verified checkpoint}
PORT=${ORNITH_PORT:-8140}
KV=${ORNITH_KV:-auto}
CTX=${ORNITH_CONTEXT:-32768}
CONC=${ORNITH_CONCURRENCY:-3}
GRAPH=${ORNITH_GRAPH:-eager}
SPEC=${ORNITH_SPEC:-none}
args=(serve "$MODEL" --served-model-name ornith-vllm --host 127.0.0.1 --port "$PORT"
  --tokenizer "${ORNITH_TOKENIZER:-/home/lab/models/Ornith-1.5-9B-original}"
  --dtype bfloat16 --max-model-len "$CTX" --gpu-memory-utilization "${ORNITH_MEMORY:-0.90}"
  --max-num-seqs "$CONC" --max-num-batched-tokens "${ORNITH_BATCH:-2048}"
  --kv-cache-dtype "$KV" --attention-backend "${ORNITH_ATTN:-TRITON_ATTN}"
  --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3)
if [ "${ORNITH_PREFIX:-off}" = off ]; then args+=(--no-enable-prefix-caching --mamba-cache-mode none); fi
if [ -n "${ORNITH_QUANT:-}" ]; then args+=(--quantization "$ORNITH_QUANT"); fi
if [ -n "${ORNITH_QUANT_CONFIG:-}" ]; then args+=(--quantization-config "$ORNITH_QUANT_CONFIG"); fi
if [ -n "${ORNITH_LOAD_FORMAT:-}" ]; then args+=(--load-format "$ORNITH_LOAD_FORMAT"); fi
if [ -n "${ORNITH_LINEAR:-}" ]; then args+=(--linear-backend "$ORNITH_LINEAR"); fi
if [ -n "${ORNITH_SSM_DTYPE:-}" ]; then args+=(--mamba-ssm-cache-dtype "$ORNITH_SSM_DTYPE"); fi
if [ -n "${ORNITH_PREFILL_CHUNK:-}" ]; then
  args+=(--long-prefill-token-threshold "$ORNITH_PREFILL_CHUNK")
fi
if [ "${ORNITH_MULTIMODAL_CONFIG:-0}" = 1 ]; then args+=(--limit-mm-per-prompt "{\"image\":${ORNITH_IMAGE_LIMIT:-0},\"video\":0}"); fi
if [ -n "${ORNITH_MM_PROCESSOR:-}" ]; then args+=(--mm-processor-kwargs "$ORNITH_MM_PROCESSOR"); fi
if [ "$GRAPH" = eager ]; then args+=(--enforce-eager); else
  args+=(--compilation-config '{"mode":0,"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,3,4,8,16,32,48]}')
fi
if [ "$SPEC" != none ]; then args+=(--speculative-config "$SPEC"); fi
printf 'GPU2 ONLY | model=%s | port=%s | KV=%s | context=%s | concurrency=%s | graphs=%s\n' "$MODEL" "$PORT" "$KV" "$CTX" "$CONC" "$GRAPH"
printf 'Power condition label: %s (external setting; launcher does not alter hardware)\n' "${ORNITH_POWER_LABEL:-original condition; see hardware.csv}"
exec /home/lab/venvs/vllm/bin/vllm "${args[@]}"
