Add Qwen235 vLLM 0.20 real surface runner

runs/frontier-fidelity-envelope-v1/run_qwen235_v020_real_surface.sh

@@ -0,0 +1,119 @@
+#!/usr/bin/env bash
+
+# Three fresh-server trials for the four-cell Qwen235 vLLM 0.20 surface.
+
+set -euo pipefail
+
+OUT="${OUTPUT_ROOT:?OUTPUT_ROOT is required}"
+TRACE_ROOT="${TRACE_ROOT:?TRACE_ROOT is required}"
+CASE_NAME="${CASE_NAME:?CASE_NAME is required}"
+PREFIX_CACHING="${PREFIX_CACHING:?PREFIX_CACHING is required}"
+RUNNER_DIR="${RUNNER_DIR:-$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)}"
+RUNNER="${RUNNER_DIR}/run_qwen30_exact_trace_real_anchor.sh"
+CLIENT="${RUNNER_DIR}/qwen30_exact_trace_client.py"
+VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
+MODEL_ROOT="${MODEL_ROOT:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-235B-A22B-FP8}"
+SERVER_READY_ATTEMPTS="${SERVER_READY_ATTEMPTS:-600}"
+IDLE_GPU_MEMORY_TOLERANCE_MIB="${IDLE_GPU_MEMORY_TOLERANCE_MIB:-16}"
+RESUME_VALID_CELLS="${RESUME_VALID_CELLS:-true}"
+PORT="${BASE_PORT:-9300}"
+REQUEST_COUNT="$(wc -l < "${TRACE_ROOT}/tp4/private/real_requests.jsonl")"
+
+[[ "$(wc -l < "${TRACE_ROOT}/tp8/private/real_requests.jsonl")" == "${REQUEST_COUNT}" ]] || {
+  echo 'ERROR: TP-specific request counts differ' >&2
+  exit 1
+}
+case "${CASE_NAME}" in fixed-pd|fixed-po|trace-pd|trace-po) ;; *) exit 2 ;; esac
+case "${PREFIX_CACHING}" in true|false) ;; *) exit 2 ;; esac
+
+mkdir -p "${OUT}/provenance"
+declare -a WAVE_PIDS=()
+
+has_valid_result() {
+  local result="$1"
+  [[ -s "${result}" ]] || return 1
+  "${VENV_ROOT}/bin/python" - "${result}" "${REQUEST_COUNT}" <<'PY'
+import json, sys
+s = json.load(open(sys.argv[1])).get("summary", {})
+raise SystemExit(0 if s.get("completed") == int(sys.argv[2]) and s.get("failed") == 0 else 1)
+PY
+}
+
+preflight_gpus() {
+  nvidia-smi --query-gpu=index,name,memory.used,utilization.gpu --format=csv,noheader \
+    | tee -a "${OUT}/controller.log"
+  nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits \
+    | awk -v tolerance="${IDLE_GPU_MEMORY_TOLERANCE_MIB}" '$1 > tolerance {exit 1}'
+}
+
+assert_no_server() {
+  ! pgrep -fa 'qwen3-235b-v020-eval' > "${OUT}/provenance/unexpected_server_processes.txt"
+}
+
+wait_for_wave() {
+  local failed=0 pid
+  for pid in "${WAVE_PIDS[@]}"; do wait "${pid}" || failed=1; done
+  WAVE_PIDS=()
+  [[ "${failed}" -eq 0 ]]
+}
+
+launch_config() {
+  local trial="$1" tp="$2" mns="$3" gpus="$4" ep=false
+  [[ "${tp}" == 8 ]] && ep=true
+  local config="tp${tp}_ep${ep}_mns${mns}" run_out="${OUT}/real/${config}/trial${trial}"
+  local requests="${TRACE_ROOT}/tp${tp}/private/real_requests.jsonl" port="${PORT}"
+  PORT=$((PORT + 1))
+  if [[ "${RESUME_VALID_CELLS}" == true ]] && has_valid_result "${run_out}/results/result.json"; then
+    printf 'CONFIG_REUSED_VALID case=%s trial=%s config=%s\n' "${CASE_NAME}" "${trial}" "${config}"
+    return 0
+  fi
+  mkdir -p "${run_out}"
+  (
+    cd "${RUNNER_DIR}"
+    set +e
+    env HOME=/tmp/wjh XDG_CACHE_HOME=/tmp/wjh/.cache \
+      VLLM_CACHE_ROOT=/tmp/wjh/.cache/vllm CUDA_VISIBLE_DEVICES="${gpus}" \
+      TP="${tp}" MNS="${mns}" TRACE_LABEL="${CASE_NAME}/tp${tp}-normalized" \
+      PREFIX_CACHING="${PREFIX_CACHING}" ENABLE_EXPERT_PARALLEL="${ep}" \
+      MODEL_QUANTIZATION=fp8 DISABLE_CUSTOM_ALL_REDUCE=true GPU_MEMORY_UTILIZATION=0.80 \
+      SERVED_MODEL=qwen3-235b-v020-eval SERVER_PORT="${port}" OUTPUT_ROOT="${run_out}" \
+      REQUESTS_FILE="${requests}" VENV_ROOT="${VENV_ROOT}" MODEL_ROOT="${MODEL_ROOT}" \
+      SERVER_READY_ATTEMPTS="${SERVER_READY_ATTEMPTS}" CLIENT_TIMEOUT_SECONDS=6000 \
+      EXACT_TRACE_CLIENT="${CLIENT}" \
+      timeout --signal=TERM --kill-after=60s 7800 bash "${RUNNER}" \
+      > "${run_out}/launcher.stdout.log" 2> "${run_out}/launcher.stderr.log"
+    status="$?"
+    printf '%s\n' "${status}" > "${run_out}/launcher.exit_code"
+    exit "${status}"
+  ) &
+  WAVE_PIDS+=("$!")
+}
+
+run_trial() {
+  local trial="$1" order="$2"
+  IFS=',' read -r -a mnss <<< "${order}"
+  preflight_gpus && assert_no_server
+  printf 'WAVE_START case=%s trial=%s tp=4 mns=%s,%s\n' \
+    "${CASE_NAME}" "${trial}" "${mnss[0]}" "${mnss[1]}" | tee -a "${OUT}/controller.log"
+  launch_config "${trial}" 4 "${mnss[0]}" '0,1,2,3'
+  launch_config "${trial}" 4 "${mnss[1]}" '4,5,6,7'
+  wait_for_wave && preflight_gpus && assert_no_server
+  for mns in "${mnss[@]}"; do
+    printf 'WAVE_START case=%s trial=%s tp=8 mns=%s\n' \
+      "${CASE_NAME}" "${trial}" "${mns}" | tee -a "${OUT}/controller.log"
+    launch_config "${trial}" 8 "${mns}" '0,1,2,3,4,5,6,7'
+    wait_for_wave && preflight_gpus && assert_no_server
+  done
+}
+
+{
+  printf '%s\n' "Q235_REAL_LAUNCH_ECHO host=dash0 model=Qwen3-235B-A22B-FP8 engine=vLLM-0.20.0+cu129 checkpoint=FP8 cases=${CASE_NAME} trace=${TRACE_ROOT}/tp{4,8} requests=${REQUEST_COUNT} transform=t_prime=t/TP prefix=${PREFIX_CACHING} surface={TP4/EP1,TP8/EP8}xMNS{64,128} MBT=8192 trials=3 fresh_server=true metrics=mean,p90(TTFT,TPOT-if-PD,E2E) SLO=not_scored output=${OUT}/real"
+  date -u +START_UTC=%Y-%m-%dT%H:%M:%SZ
+  sha256sum "${BASH_SOURCE[0]}" "${RUNNER}" "${CLIENT}" "${MODEL_ROOT}/config.json" \
+    > "${OUT}/provenance/real-input.sha256"
+  run_trial 1 '64,128'
+  run_trial 2 '128,64'
+  run_trial 3 '64,128'
+  date -u +END_UTC=%Y-%m-%dT%H:%M:%SZ
+  printf '%s\n' 'Q235_REAL_SURFACE_COMPLETE'
+} 2>&1 | tee -a "${OUT}/controller.log"

runs/frontier-fidelity-envelope-v1/run_qwen30_exact_trace_real_anchor.sh

@@ -9,11 +9,16 @@ MNS="${MNS:?MNS is required}"
 TRACE_LABEL="${TRACE_LABEL:?TRACE_LABEL is required}"
 SERVER_PORT="${SERVER_PORT:?SERVER_PORT is required}"
 PREFIX_CACHING="${PREFIX_CACHING:-true}"
+ENABLE_EXPERT_PARALLEL="${ENABLE_EXPERT_PARALLEL:-false}"
+MODEL_QUANTIZATION="${MODEL_QUANTIZATION:-}"
+DISABLE_CUSTOM_ALL_REDUCE="${DISABLE_CUSTOM_ALL_REDUCE:-false}"
+GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.92}"
 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="${FLASHINFER_WORKSPACE_BASE:-/tmp/wjh/flashinfer-workspace-vllm020-profiler-v1}"
 SERVER_READY_ATTEMPTS="${SERVER_READY_ATTEMPTS:-120}"
-SERVED_MODEL="qwen3-30b-exact-trace"
+CLIENT_TIMEOUT_SECONDS="${CLIENT_TIMEOUT_SECONDS:-1800}"
+SERVED_MODEL="${SERVED_MODEL:-qwen3-30b-exact-trace}"
 EXACT_TRACE_CLIENT="${EXACT_TRACE_CLIENT:-qwen30_exact_trace_client.py}"
 SERVER_PID=""
 
@@ -53,8 +58,23 @@ case "${PREFIX_CACHING}" in
     ;;
 esac
 
+case "${ENABLE_EXPERT_PARALLEL}" in
+  true) EP_FLAG=(--enable-expert-parallel) ;;
+  false) EP_FLAG=() ;;
+  *) echo "ERROR: ENABLE_EXPERT_PARALLEL must be true or false" >&2; exit 1 ;;
+esac
+case "${DISABLE_CUSTOM_ALL_REDUCE}" in
+  true) CUSTOM_AR_FLAG=(--disable-custom-all-reduce) ;;
+  false) CUSTOM_AR_FLAG=() ;;
+  *) echo "ERROR: DISABLE_CUSTOM_ALL_REDUCE must be true or false" >&2; exit 1 ;;
+esac
+QUANT_FLAG=()
+if [[ -n "${MODEL_QUANTIZATION}" ]]; then
+  QUANT_FLAG=(--quantization "${MODEL_QUANTIZATION}")
+fi
+
 REQUEST_COUNT="$(wc -l < "${REQUESTS_FILE}")"
-echo "QWEN30_EXACT_TRACE_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 config=TP${TP}_MNS${MNS}_MBT8192 trace=${TRACE_LABEL} requests=${REQUEST_COUNT} source=${REQUESTS_FILE} arrivals=manifest prefix=${PREFIX_CACHING} block=16 metrics=TTFT,TPOT-if-OSL-gt-1,E2E flashinfer_workspace=${FLASHINFER_WORKSPACE_BASE} output=${OUTPUT_ROOT} ready_budget_s=$((SERVER_READY_ATTEMPTS * 3)) hard_wall=3600s"
+echo "EXACT_TRACE_REAL_LAUNCH_ECHO host=$(hostname) gpus=${CUDA_VISIBLE_DEVICES} model=${MODEL_ROOT} runtime=vLLM-0.20.0+cu129 dtype=BF16 quantization=${MODEL_QUANTIZATION:-none} config=TP${TP}_EP${ENABLE_EXPERT_PARALLEL}_MNS${MNS}_MBT8192 trace=${TRACE_LABEL} requests=${REQUEST_COUNT} source=${REQUESTS_FILE} arrivals=manifest prefix=${PREFIX_CACHING} block=16 metrics=TTFT,TPOT-if-OSL-gt-1,E2E flashinfer_workspace=${FLASHINFER_WORKSPACE_BASE} output=${OUTPUT_ROOT} ready_budget_s=$((SERVER_READY_ATTEMPTS * 3)) client_timeout_s=${CLIENT_TIMEOUT_SECONDS}"
 date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
 sha256sum qwen30_exact_trace_client.py run_qwen30_exact_trace_real_anchor.sh \
   ../frontier-phase-factorial-v0/qwen30_prefill_client.py \
@@ -74,9 +94,10 @@ ulimit -n 65536
 
 setsid "${VENV_ROOT}/bin/vllm" serve "${MODEL_ROOT}" \
   --host 127.0.0.1 --port "${SERVER_PORT}" --served-model-name "${SERVED_MODEL}" \
-  --tensor-parallel-size "${TP}" --gpu-memory-utilization 0.92 \
+  --tensor-parallel-size "${TP}" --gpu-memory-utilization "${GPU_MEMORY_UTILIZATION}" \
   --max-model-len 40960 --max-num-batched-tokens 8192 --max-num-seqs "${MNS}" \
   "${PREFIX_CACHING_FLAG}" --enable-chunked-prefill --no-enable-log-requests \
+  "${CUSTOM_AR_FLAG[@]}" "${QUANT_FLAG[@]}" "${EP_FLAG[@]}" \
   > "${OUTPUT_ROOT}/logs/server.log" 2>&1 &
 SERVER_PID=$!
 READY=0
@@ -110,7 +131,7 @@ fi
   --port "${SERVER_PORT}" --requests-file "${REQUESTS_FILE}" \
   --served-model "${SERVED_MODEL}" \
   --output "${OUTPUT_ROOT}/results/result.json" --tpot-slo-ms 150 \
-  --timeout-seconds 1800
+  --timeout-seconds "${CLIENT_TIMEOUT_SECONDS}"
 
 cleanup
 find "${OUTPUT_ROOT}" -type f ! -path '*/provenance/artifacts.sha256' -print0 \