aituner/runs/frontier-phase-factorial-v0/run_qwen30_prefill_real_config.sh

#!/usr/bin/env bash

set -euo pipefail

OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT is required}"
TP="${TP:?TP is required}"
MNS="${MNS:?MNS is required}"
RATES="${RATES:-4 8 16 32 64}"
INPUT_TOKENS="${INPUT_TOKENS:-2048}"
OUTPUT_TOKENS="${OUTPUT_TOKENS:-1}"
TPOT_SLO_MS="${TPOT_SLO_MS:-150}"
WARMUP_SECONDS="${WARMUP_SECONDS:-2}"
SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
SERVED_MODEL="qwen3-30b-prefill-only"
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}"

echo "QWEN30_FIXED_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 rates=${RATES// /,} rounds=2 requests=64 shape=ISL${INPUT_TOKENS}_OSL${OUTPUT_TOKENS} ttft_slo=1000+1000*ISL/8000ms tpot_slo=${TPOT_SLO_MS}ms warmup_seconds=${WARMUP_SECONDS} arrivals=uniform prefix=off cuda_graph=runtime_default isolation=fresh_server_per_anchor output=${OUTPUT_ROOT}"
date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
sha256sum run_qwen30_prefill_real_config.sh qwen30_prefill_client.py \
  > "${OUTPUT_ROOT}/provenance/source.sha256"
"${VENV_ROOT}/bin/python" - "${TP}" "${MNS}" "${RATES}" "${INPUT_TOKENS}" "${OUTPUT_TOKENS}" "${TPOT_SLO_MS}" "${WARMUP_SECONDS}" \
  > "${OUTPUT_ROOT}/provenance/contract.json" <<'PY'
import importlib.metadata as metadata
import json
import platform
import sys

tp, mns, rates, input_tokens, output_tokens, tpot_slo_ms, warmup_seconds = sys.argv[1:]
input_tokens = int(input_tokens)
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": 8192},
    "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": f"min(32, max(4, ceil(rate * {warmup_seconds})))",
    "input_tokens": input_tokens,
    "output_tokens": int(output_tokens),
    "ttft_slo_ms": 1000.0 + 1000.0 * input_tokens / 8000.0,
    "tpot_slo_ms": float(tpot_slo_ms) if int(output_tokens) > 1 else None,
    "target_pass_rate": 0.95,
}, 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 TORCH_CUDA_ARCH_LIST=9.0
export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1

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}" --gpu-memory-utilization 0.92 \
      --max-model-len 40960 --max-num-batched-tokens 8192 --max-num-seqs "${MNS}" \
      --no-enable-prefix-caching --enable-chunked-prefill --no-enable-log-requests \
      > "${SERVER_LOG}" 2>&1 &
    SERVER_PID=$!
    READY=0
    for _ in $(seq 1 120); do
      if curl -fsS --max-time 2 "http://127.0.0.1:${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 3
    done
    if [[ "${READY}" -ne 1 ]]; then
      tail -200 "${SERVER_LOG}"
      exit 1
    fi

    WARMUP_REQUESTS="$("${VENV_ROOT}/bin/python" - "${RATE}" "${WARMUP_SECONDS}" <<'PY'
import math
import sys
print(min(32, max(4, math.ceil(float(sys.argv[1]) * float(sys.argv[2])))))
PY
)"
    "${VENV_ROOT}/bin/python" qwen30_prefill_client.py --port "${SERVER_PORT}" \
      --served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
      --requests "${WARMUP_REQUESTS}" --input-tokens "${INPUT_TOKENS}" \
      --output-tokens "${OUTPUT_TOKENS}" --tpot-slo-ms "${TPOT_SLO_MS}" \
      --output "${ROUND_ROOT}/results/warmup_${KEY}.json"
    "${VENV_ROOT}/bin/python" qwen30_prefill_client.py --port "${SERVER_PORT}" \
      --served-model "${SERVED_MODEL}" --model-path "${MODEL_ROOT}" --rate "${RATE}" \
      --requests 64 --input-tokens "${INPUT_TOKENS}" \
      --output-tokens "${OUTPUT_TOKENS}" --tpot-slo-ms "${TPOT_SLO_MS}" \
      --output "${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 "QWEN30_PREFILL_REAL_CONFIG_COMPLETE"