#!/usr/bin/env bash

set -euo pipefail

OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT is required}"
TP="${TP:?TP is required}"
MNS="${MNS:?MNS is required}"
MBT="${MBT:?MBT is required}"
RATES="${RATES:?RATES is required}"
SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
VENV_ROOT="${VENV_ROOT:-/tmp/wjh-frontier-vllm0102-smoke/.venv}"
MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-235B-A22B-FP8}"
SERVED_MODEL="qwen3-235b-t0-surface"
SERVER_PID=""

mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance"
exec > >(tee -a "${OUTPUT_ROOT}/logs/controller.log") 2>&1

cleanup() {
  if [[ -n "${SERVER_PID}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
    kill -TERM -- "-${SERVER_PID}" 2>/dev/null || true
    for _ in $(seq 1 30); do
      kill -0 "${SERVER_PID}" 2>/dev/null || break
      sleep 1
    done
    kill -KILL -- "-${SERVER_PID}" 2>/dev/null || true
  fi
  SERVER_PID=""
}
trap cleanup EXIT INT TERM

IFS=',' read -r -a GPU_IDS <<< "${CUDA_VISIBLE_DEVICES:-}"
if [[ "${#GPU_IDS[@]}" -ne "${TP}" ]]; then
  echo "ERROR: expected ${TP} allocated GPUs, got ${CUDA_VISIBLE_DEVICES:-unset}" >&2
  exit 1
fi
read -r -a RATE_ARRAY <<< "${RATES}"
if [[ "${#RATE_ARRAY[@]}" -lt 1 ]]; then
  echo "ERROR: at least one frozen rate is required" >&2
  exit 1
fi

echo "REAL_CONFIG_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} config=TP${TP}_MNS${MNS}_MBT${MBT} rates=${RATES// /,} repeats=2 requests_per_anchor=64 isolation=fresh_server_per_anchor target_warmup=min32_max4_ceil_rate_x20 trace=fixed_ISL2048_OSL128 prefix=off runtime=community_vllm_0.10.2 execution=eager kv=BF16 output=${OUTPUT_ROOT}"
date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
sha256sum run_t0_real_config.sh t0_rate_client.py > "${OUTPUT_ROOT}/provenance/source.sha256"
"${VENV_ROOT}/bin/python" - "${TP}" "${MNS}" "${MBT}" "${RATES}" \
  > "${OUTPUT_ROOT}/provenance/contract.json" <<'PY'
import importlib.metadata as metadata
import json
import platform
import sys

tp, mns, mbt, rates = sys.argv[1:]
print(json.dumps({
    "python": platform.python_version(),
    "torch": metadata.version("torch"),
    "transformers": metadata.version("transformers"),
    "vllm": metadata.version("vllm"),
    "config": {"tp": int(tp), "mns": int(mns), "mbt": int(mbt)},
    "rates": [float(value) for value in rates.split()],
    "rounds": 2,
    "requests_per_anchor": 64,
    "anchor_isolation": "fresh_server_per_rate_per_round",
    "target_rate_warmup_requests": "min(32, max(4, ceil(rate * 20)))",
    "ttft_slo_ms": 1256.0,
    "tpot_slos_ms": [40.0, 120.0, 150.0, 180.0],
}, indent=2, sort_keys=True))
PY
nvidia-smi --query-gpu=index,name,uuid,driver_version --format=csv,noheader \
  > "${OUTPUT_ROOT}/provenance/gpus.csv"
sha256sum "${MODEL_ROOT}/config.json" > "${OUTPUT_ROOT}/provenance/model.sha256"

export TOKENIZERS_PARALLELISM=false
export VLLM_USE_V1=1
export VLLM_ATTENTION_BACKEND=FLASHINFER
export TORCH_CUDA_ARCH_LIST=9.0

EXTRA_FLAGS=()
NUM_BLOCKS=26101
if [[ "${TP}" -eq 8 ]]; then
  EXTRA_FLAGS+=(--enable-expert-parallel)
  NUM_BLOCKS=62351
fi

for ROUND in 1 2; do
  ROUND_ROOT="${OUTPUT_ROOT}/round${ROUND}"
  mkdir -p "${ROUND_ROOT}/logs" "${ROUND_ROOT}/results"
  ORDERED_RATES=("${RATE_ARRAY[@]}")
  if [[ "${ROUND}" -eq 2 ]]; then
    ORDERED_RATES=()
    for ((index=${#RATE_ARRAY[@]}-1; index>=0; index--)); do ORDERED_RATES+=("${RATE_ARRAY[index]}"); done
  fi
  for RATE in "${ORDERED_RATES[@]}"; do
    KEY="$(printf 'r%.2f' "${RATE}" | tr '.' 'p')"
    SERVER_LOG="${ROUND_ROOT}/logs/server_${KEY}.log"
    setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \
      --host 127.0.0.1 --port "${SERVER_PORT}" --served-model-name "${SERVED_MODEL}" \
      --tensor-parallel-size "${TP}" --disable-custom-all-reduce --quantization fp8 \
      --gpu-memory-utilization 0.80 --num-gpu-blocks-override "${NUM_BLOCKS}" \
      --kv-cache-dtype auto --max-model-len 40960 --max-num-batched-tokens "${MBT}" \
      --max-num-seqs "${MNS}" --no-enable-prefix-caching --enable-chunked-prefill \
      --enforce-eager --disable-log-requests "${EXTRA_FLAGS[@]}" \
      > "${SERVER_LOG}" 2>&1 &
    SERVER_PID=$!
    READY=0
    for _ in $(seq 1 180); do
      if curl -fsS --max-time 2 "http://127.0.0.1:${SERVER_PORT}/v1/models" \
        > "${ROUND_ROOT}/results/models_${KEY}.json" 2>/dev/null; then
        READY=1
        break
      fi
      if ! kill -0 "${SERVER_PID}" 2>/dev/null; then tail -200 "${SERVER_LOG}"; exit 1; fi
      sleep 5
    done
    if [[ "${READY}" -ne 1 ]]; then tail -200 "${SERVER_LOG}"; exit 1; fi

    WARMUP_REQUESTS="$("${VENV_ROOT}/bin/python" - "${RATE}" <<'PY'
import math
import sys
print(min(32, max(4, math.ceil(float(sys.argv[1]) * 20.0))))
PY
)"
    "${VENV_ROOT}/bin/python" t0_rate_client.py --port "${SERVER_PORT}" \
      --served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
      --requests "${WARMUP_REQUESTS}" \
      --output "${ROUND_ROOT}/results/warmup_${KEY}.json"
    "${VENV_ROOT}/bin/python" t0_rate_client.py --port "${SERVER_PORT}" \
      --served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
      --requests 64 --output "${ROUND_ROOT}/results/${KEY}.json"
    cleanup
  done
done

find "${OUTPUT_ROOT}" -type f ! -path '*/provenance/artifacts.sha256' -print0 \
  | sort -z | xargs -0 sha256sum > "${OUTPUT_ROOT}/provenance/artifacts.sha256"
date -u +"END_UTC=%Y-%m-%dT%H:%M:%SZ"
echo "T0_REAL_CONFIG_COMPLETE"