aituner/runs/frontier-fidelity-envelope-v1/run_qwen30_fixed_runtime_preflight.sh

#!/usr/bin/env bash

# Capture vLLM's resolved runtime state for Fixed-P/PD before their Frontier
# surfaces are frozen.  No benchmark request is issued in this preflight.

set -euo pipefail

OUT="${1:?output root is required}"
VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
FLASHINFER_WORKSPACE_BASE="${FLASHINFER_WORKSPACE_BASE:-${OUT}/flashinfer-shared-workspace}"
PORT=8600

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

echo "FIXED_RUNTIME_PREFLIGHT_LAUNCH_ECHO host=$(hostname) model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 configs=TP{1,2,4}xMNS{8,16,32,64} MBT=8192 prefix=false chunked_prefill=true requests=0 resolved_state=graph_capture_and_kv_blocks output=${OUT} expected_wall=8-20m expected_gpu_cap=2_H20h"
date -u +START_UTC=%Y-%m-%dT%H:%M:%SZ
sha256sum "${BASH_SOURCE[0]}" "${MODEL_ROOT}/config.json" > "${OUT}/provenance/input.sha256"
"${VENV_ROOT}/bin/vllm" --version > "${OUT}/provenance/vllm.version"

if ! nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | awk '$1 != 0 {exit 1}'; then
  echo 'ERROR: a GPU is not idle before runtime preflight' >&2
  exit 1
fi

declare -a PIDS=()

cleanup_pid() {
  local pid="$1"
  if kill -0 "${pid}" 2>/dev/null; then
    kill -TERM -- "-${pid}" 2>/dev/null || true
    for _ in $(seq 1 30); do
      kill -0 "${pid}" 2>/dev/null || break
      sleep 1
    done
    kill -KILL -- "-${pid}" 2>/dev/null || true
  fi
}

cleanup_all() {
  local pid
  for pid in "${PIDS[@]:-}"; do
    cleanup_pid "${pid}"
  done
}
trap cleanup_all EXIT INT TERM

launch_one() {
  local tp="$1" mns="$2" gpus="$3" port="$4"
  local name="tp${tp}_mns${mns}"
  local root="${OUT}/runs/${name}"
  mkdir -p "${root}"
  (
    export HOME=/tmp/wjh
    export XDG_CACHE_HOME=/tmp/wjh/.cache
    export VLLM_CACHE_ROOT=/tmp/wjh/.cache/vllm
    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
    export CUDA_VISIBLE_DEVICES="${gpus}"
    setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \
      --host 127.0.0.1 --port "${port}" --served-model-name qwen3-30b-fixed-preflight \
      --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 \
      > "${root}/server.log" 2>&1 &
    local server_pid=$!
    local ready=0
    for _ in $(seq 1 150); do
      if curl -fsS --max-time 2 "http://127.0.0.1:${port}/v1/models" > "${root}/models.json" 2>/dev/null; then
        ready=1
        break
      fi
      if ! kill -0 "${server_pid}" 2>/dev/null; then
        break
      fi
      sleep 2
    done
    grep -E -i 'cuda graph|cudagraph|capture|gpu blocks|num_gpu_blocks|kv cache' "${root}/server.log" > "${root}/resolved-state.log" || true
    if [[ "${ready}" -ne 1 ]]; then
      echo "status=failed" > "${root}/status"
      cleanup_pid "${server_pid}"
      exit 1
    fi
    echo "status=ready" > "${root}/status"
    cleanup_pid "${server_pid}"
  ) &
  PIDS+=("$!")
}

wait_wave() {
  local failed=0 pid
  for pid in "${PIDS[@]}"; do
    wait "${pid}" || failed=1
  done
  PIDS=()
  if [[ "${failed}" -ne 0 ]]; then
    echo 'ERROR: one or more runtime preflight servers failed' >&2
    exit 1
  fi
  if ! nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | awk '$1 != 0 {exit 1}'; then
    echo 'ERROR: a GPU remains allocated after runtime preflight wave' >&2
    exit 1
  fi
}

for tp in 4 2 1; do
  echo "PREFLIGHT_WAVE_START tp=${tp}"
  case "${tp}" in
    4)
      launch_one 4 8  '0,1,2,3' "${PORT}"; PORT=$((PORT + 1))
      launch_one 4 16 '4,5,6,7' "${PORT}"; PORT=$((PORT + 1))
      wait_wave
      launch_one 4 32 '0,1,2,3' "${PORT}"; PORT=$((PORT + 1))
      launch_one 4 64 '4,5,6,7' "${PORT}"; PORT=$((PORT + 1))
      ;;
    2)
      launch_one 2 8  '0,1' "${PORT}"; PORT=$((PORT + 1))
      launch_one 2 16 '2,3' "${PORT}"; PORT=$((PORT + 1))
      launch_one 2 32 '4,5' "${PORT}"; PORT=$((PORT + 1))
      launch_one 2 64 '6,7' "${PORT}"; PORT=$((PORT + 1))
      ;;
    1)
      launch_one 1 8  '0' "${PORT}"; PORT=$((PORT + 1))
      launch_one 1 16 '1' "${PORT}"; PORT=$((PORT + 1))
      launch_one 1 32 '2' "${PORT}"; PORT=$((PORT + 1))
      launch_one 1 64 '3' "${PORT}"; PORT=$((PORT + 1))
      ;;
  esac
  wait_wave
  echo "PREFLIGHT_WAVE_COMPLETE tp=${tp}"
done

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