| 1 | #!/bin/bash |
| 2 | # ============================================================================= |
| 3 | # vLLM disaggregated (P/D) launcher (rank-based / SLURM-native). |
| 4 | # ----------------------------------------------------------------------------- |
| 5 | # |
| 6 | # rank 0 .. xP-1 -> prefill (rank 0 is also the orchestrator + proxy) |
| 7 | # rank xP .. xP+yD-1 -> decode |
| 8 | # |
| 9 | # Rank 0 additionally: starts the MoRIIO proxy, health-gates every server, |
| 10 | # runs the post-health workload (RUN_AFTER_HEALTH=bench|accuracy|none), then |
| 11 | # drops a shared-FS completion sentinel so the other ranks shut down and the |
| 12 | # single `srun` returns. Rank 0's exit code is the run's pass/fail. |
| 13 | # |
| 14 | # Explicit per-role invocation is also supported (granular for manual use): |
| 15 | # proxy | prefill | decode | bench | accuracy. |
| 16 | # |
| 17 | # ---- LAUNCH IT (both supported) -------------------------------------------- |
| 18 | # 1) Single srun (SLURM provides the rank) — what run_xPyD_disagg.slurm does: |
| 19 | # srun --nodes=$((xP+yD)) --ntasks-per-node=1 \ |
| 20 | # bash .../vllm_disagg.sh node |
| 21 | # (IPADDRS must be ordered to match ranks: prefill IPs first, then decode, |
| 22 | # in the SAME order srun assigns PROCID — sort nodes alphabetically.) |
| 23 | # |
| 24 | # 2) Manual, one shell per node (rank from NODE_RANK; share LOG_PATH/FS): |
| 25 | # IPADDRS=ipP0,ipD0 xP=1 yD=1 NODE_RANK=0 bash vllm_disagg.sh node |
| 26 | # IPADDRS=ipP0,ipD0 xP=1 yD=1 NODE_RANK=1 bash vllm_disagg.sh node |
| 27 | # |
| 28 | # 3) Manual, fully granular (per-role): |
| 29 | # bash vllm_disagg.sh proxy |
| 30 | # bash vllm_disagg.sh prefill |
| 31 | # NODE_RANK=1 bash vllm_disagg.sh decode |
| 32 | # bash vllm_disagg.sh bench |
| 33 | # |
| 34 | # Cluster config: cluster.sh (sourced). Model flags: models.yaml. |
| 35 | # Parallelism: WIDE_EP_MODE=0 tp (independent TP servers) | 1 ep (DP+EP groups). |
| 36 | # ============================================================================= |
| 37 | set -euo pipefail |
| 38 | |
| 39 | SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)" |
| 40 | |
| 41 | _PROXY_IP_OVERRIDE="${PROXY_IP:-}" |
| 42 | |
| 43 | log() { echo "[vllm_disagg] $*"; } |
| 44 | die() { echo "ERROR: $*" >&2; exit 1; } |
| 45 | |
| 46 | usage() { |
| 47 | echo "Usage: $0 [node|proxy|prefill|decode|bench|accuracy] [--wide-ep-mode 0|1]" >&2 |
| 48 | echo " (no role => 'node': derive role from \$SLURM_PROCID/\$NODE_RANK)" >&2 |
| 49 | exit 1 |
| 50 | } |
| 51 | |
| 52 | # ----------------------------------------------------------------- arg parsing |
| 53 | # Sets globals: ROLE (default 'node'), _WIDE_EP_MODE_OVERRIDE. |
| 54 | parse_args() { |
| 55 | ROLE="${1:-node}" |
| 56 | shift || true |
| 57 | _WIDE_EP_MODE_OVERRIDE="" |
| 58 | while [[ $# -gt 0 ]]; do |
| 59 | case "$1" in |
| 60 | --wide-ep-mode) _WIDE_EP_MODE_OVERRIDE="${2:-}"; shift 2 ;; |
| 61 | --wide-ep-mode=*) _WIDE_EP_MODE_OVERRIDE="${1#*=}"; shift ;; |
| 62 | -h|--help) usage ;; |
| 63 | *) die "Unknown argument: $1" ;; |
| 64 | esac |
| 65 | done |
| 66 | } |
| 67 | |
| 68 | # ----------------------------------------------------------------- load config |
| 69 | # Sources cluster.sh and resolves WIDE_EP_MODE -> PARALLEL_MODE (tp|ep). |
| 70 | load_config() { |
| 71 | CLUSTER_ENV="${CLUSTER_ENV:-${SCRIPT_DIR}/cluster.sh}" |
| 72 | [[ -f "${CLUSTER_ENV}" ]] || die "cluster env file not found: ${CLUSTER_ENV}" |
| 73 | # shellcheck disable=SC1090 |
| 74 | source "${CLUSTER_ENV}" |
| 75 | |
| 76 | # Topology fanout normally arrives from cluster.sh; restate the defaults so |
| 77 | # the rank arithmetic is defined even when it does not. |
| 78 | : "${xP:=1}" "${yD:=1}" |
| 79 | |
| 80 | [[ -n "${_WIDE_EP_MODE_OVERRIDE}" ]] && WIDE_EP_MODE="${_WIDE_EP_MODE_OVERRIDE}" |
| 81 | WIDE_EP_MODE="${WIDE_EP_MODE:-0}" |
| 82 | case "${WIDE_EP_MODE}" in |
| 83 | 0) PARALLEL_MODE="tp" ;; |
| 84 | 1) PARALLEL_MODE="ep" ;; |
| 85 | *) die "WIDE_EP_MODE must be 0 (tp) or 1 (ep); got '${WIDE_EP_MODE}'" ;; |
| 86 | esac |
| 87 | |
| 88 | # node-mode orchestration knobs |
| 89 | RUN_AFTER_HEALTH="${RUN_AFTER_HEALTH:-accuracy}" # bench | accuracy | none |
| 90 | HEALTH_TIMEOUT_S="${HEALTH_TIMEOUT_S:-3600}" |
| 91 | |
| 92 | # MoRIIO KV transfer direction. 0 (default) |
| 93 | MORIIO_READ_MODE="${MORIIO_READ_MODE:-0}" |
| 94 | } |
| 95 | |
| 96 | # Emits the `read_mode` KV-config fragment (leading comma + newline) when read |
| 97 | # mode is enabled, else nothing (leaving MoRIIO's write-mode default). |
| 98 | moriio_read_mode_kv() { |
| 99 | case "${MORIIO_READ_MODE:-0}" in |
| 100 | 1|true|True|TRUE|yes|on) printf ',\n "read_mode": true' ;; |
| 101 | *) : ;; |
| 102 | esac |
| 103 | } |
| 104 | |
| 105 | # ----------------------------------------------------------------- topology |
| 106 | # Resolves MODEL_PATH, MODELS_YAML, the IP array, and the master/proxy addresses. |
| 107 | # IP_ARRAY[0..xP-1] = prefill nodes, [xP..xP+yD-1] = decode nodes. |
| 108 | resolve_topology() { |
| 109 | MODEL_PATH="${MODEL_DIR%/}/${MODEL_NAME}" |
| 110 | MODELS_YAML="${MODELS_YAML:-${SCRIPT_DIR}/models.yaml}" |
| 111 | mkdir -p "${LOG_PATH}" 2>/dev/null || true |
| 112 | |
| 113 | [[ -z "${IPADDRS}" ]] && IPADDRS="${PREFILL_IP},${DECODE_IP}" # 1P1D fallback |
| 114 | IFS=',' read -ra IP_ARRAY <<< "${IPADDRS}" |
| 115 | PREFILL_MASTER_ADDR="${IP_ARRAY[0]}" |
| 116 | DECODE_MASTER_ADDR="${IP_ARRAY[$xP]:-${IP_ARRAY[-1]}}" |
| 117 | # Proxy is co-located on the prefill master (rank 0) unless the CALLER |
| 118 | # explicitly set PROXY_IP. We use the pre-cluster.sh override (captured at the |
| 119 | # top) — not the live PROXY_IP, which cluster.sh always populates with the |
| 120 | # static PREFILL_IP fallback and would otherwise mask the real master. |
| 121 | PROXY_IP="${_PROXY_IP_OVERRIDE:-${PREFILL_MASTER_ADDR}}" |
| 122 | } |
| 123 | |
| 124 | # ============================================================================ |
| 125 | # Non-server roles |
| 126 | # ============================================================================ |
| 127 | |
| 128 | run_proxy() { |
| 129 | [[ -f "${PROXY_SCRIPT}" ]] || die "proxy script not found: ${PROXY_SCRIPT}" |
| 130 | log "proxy: http=:${PROXY_PORT} discovery=:${PROXY_PING_PORT} (${PROXY_SCRIPT})" |
| 131 | exec python3 "${PROXY_SCRIPT}" --port "${PROXY_PORT}" |
| 132 | } |
| 133 | |
| 134 | # Start the proxy in the BACKGROUND (node-mode rank 0). Sets PROXY_PID. |
| 135 | start_proxy_bg() { |
| 136 | [[ -f "${PROXY_SCRIPT}" ]] || die "proxy script not found: ${PROXY_SCRIPT}" |
| 137 | local plog="${LOG_PATH}/proxy_$(date +%Y%m%d_%H%M%S).log" |
| 138 | log "proxy(bg): http=:${PROXY_PORT} discovery=:${PROXY_PING_PORT} log=${plog}" |
| 139 | python3 "${PROXY_SCRIPT}" --port "${PROXY_PORT}" >"${plog}" 2>&1 & |
| 140 | PROXY_PID=$! |
| 141 | } |
| 142 | |
| 143 | run_bench() { |
| 144 | local base_url="http://127.0.0.1:${GATEWAY_PORT:-${PROXY_PORT}}" |
| 145 | local result_dir="${LOG_PATH}/bench_$(date +%Y%m%d_%H%M%S)" |
| 146 | mkdir -p "${result_dir}" |
| 147 | log "bench -> ${base_url} model=${MODEL_PATH} wide_ep_mode=${WIDE_EP_MODE}(${PARALLEL_MODE})" |
| 148 | log "combinations='${BENCHMARK_COMBINATIONS}' concurrency='${BENCHMARK_CON}'" |
| 149 | |
| 150 | local combo isl osl con nump logf |
| 151 | for combo in ${BENCHMARK_COMBINATIONS}; do |
| 152 | isl="${combo%%/*}"; osl="${combo##*/}" |
| 153 | for con in ${BENCHMARK_CON}; do |
| 154 | nump=$(( NUM_PROMPTS_FACTOR * con )) |
| 155 | (( nump < BENCHMARK_MIN_PROMPTS )) && nump="${BENCHMARK_MIN_PROMPTS}" |
| 156 | logf="${result_dir}/bench_${isl}_${osl}_con${con}.log" |
| 157 | log "ISL=${isl} OSL=${osl} CON=${con} num_prompts=${nump} -> ${logf}" |
| 158 | vllm bench serve \ |
| 159 | --backend openai \ |
| 160 | --base-url "${base_url}" \ |
| 161 | --endpoint /v1/completions \ |
| 162 | --model "${MODEL_PATH}" \ |
| 163 | --dataset-name random \ |
| 164 | --random-input-len "${isl}" \ |
| 165 | --random-output-len "${osl}" \ |
| 166 | --num-prompts "${nump}" \ |
| 167 | --max-concurrency "${con}" \ |
| 168 | --percentile-metrics ttft,tpot,itl,e2el \ |
| 169 | --ignore-eos \ |
| 170 | 2>&1 | tee "${logf}" |
| 171 | done |
| 172 | done |
| 173 | log "bench complete. Results in ${result_dir}" |
| 174 | } |
| 175 | |
| 176 | run_accuracy() { |
| 177 | # lm_eval isn't in the stock vLLM image; install on demand (rank 0 only, and |
| 178 | # only when accuracy is actually selected). Override with ACCURACY_PIP_SPEC. |
| 179 | if ! command -v lm_eval >/dev/null 2>&1; then |
| 180 | log "lm_eval not found — installing ${ACCURACY_PIP_SPEC:-lm_eval[api]}" |
| 181 | python3 -m pip install --no-cache-dir "${ACCURACY_PIP_SPEC:-lm_eval[api]}" |
| 182 | fi |
| 183 | local base_url="http://127.0.0.1:${GATEWAY_PORT:-${PROXY_PORT}}/v1/completions" |
| 184 | local ts; ts="$(date +%Y%m%d_%H%M%S)" |
| 185 | local logf="${LOG_PATH}/accuracy_${MODEL_NAME}_${PARALLEL_MODE}_${ts}.log" |
| 186 | local outdir="${LOG_PATH}/lm_eval_${MODEL_NAME}_${PARALLEL_MODE}_${ts}" |
| 187 | log "accuracy -> ${base_url} model=${MODEL_PATH} tasks=${ACCURACY_TASKS} wide_ep_mode=${WIDE_EP_MODE}(${PARALLEL_MODE})" |
| 188 | log "log=${logf} results=${outdir}" |
| 189 | |
| 190 | # --output_path makes lm_eval persist a results_*.json for scraping. |
| 191 | local eval_rc=0 |
| 192 | ( set -o pipefail |
| 193 | # Model weights are already served (loaded offline in a separate process); |
| 194 | # allow the Hub only here so lm_eval can fetch the eval dataset (e.g. gsm8k). |
| 195 | export HF_HUB_OFFLINE=0 HF_DATASETS_OFFLINE=0 |
| 196 | python3 -m lm_eval --model local-completions \ |
| 197 | --tasks "${ACCURACY_TASKS}" \ |
| 198 | --model_args "model=${MODEL_PATH},base_url=${base_url},num_concurrent=${ACCURACY_NUM_CONCURRENT},max_retries=${ACCURACY_MAX_RETRIES},tokenized_requests=False,trust_remote_code=True,timeout=${ACCURACY_TIMEOUT:-3600}" \ |
| 199 | --output_path "${outdir}" \ |
| 200 | 2>&1 | tee "${logf}" |
| 201 | ) || eval_rc=$? |
| 202 | if (( eval_rc != 0 )); then |
| 203 | log "FAIL: lm_eval exited rc=${eval_rc} (no usable score) — Log: ${logf}" |
| 204 | return "${eval_rc}" |
| 205 | fi |
| 206 | |
| 207 | # Gate on the score. Find the results JSON lm_eval just wrote, pull the max |
| 208 | # value of ACCURACY_METRIC across filters/tasks, and compare to the threshold. |
| 209 | local results_json="" |
| 210 | results_json="$(find "${outdir}" -name 'results_*.json' -type f 2>/dev/null | sort | tail -n1)" |
| 211 | if [[ -z "${results_json}" ]]; then |
| 212 | log "FAIL: no results_*.json under ${outdir} — cannot evaluate threshold" |
| 213 | return 1 |
| 214 | fi |
| 215 | |
| 216 | local score="" |
| 217 | score="$(python3 - "${results_json}" "${ACCURACY_METRIC}" <<'PY' |
| 218 | import json, sys |
| 219 | path, metric = sys.argv[1], sys.argv[2] |
| 220 | with open(path) as f: |
| 221 | data = json.load(f) |
| 222 | best = None |
| 223 | for task, metrics in (data.get("results") or {}).items(): |
| 224 | for key, val in metrics.items(): |
| 225 | if isinstance(val, bool) or not isinstance(val, (int, float)): |
| 226 | continue |
| 227 | if key.split(",")[0] == metric: |
| 228 | best = val if best is None else max(best, val) |
| 229 | print("" if best is None else "%.6f" % best) |
| 230 | PY |
| 231 | )" || score="" |
| 232 | |
| 233 | if [[ -z "${score}" ]]; then |
| 234 | log "FAIL: could not parse metric '${ACCURACY_METRIC}' from ${results_json}" |
| 235 | return 1 |
| 236 | fi |
| 237 | |
| 238 | if awk -v s="${score}" -v t="${ACCURACY_THRESHOLD}" 'BEGIN{exit !(s+0 >= t+0)}'; then |
| 239 | log "PASS: ${ACCURACY_METRIC}=${score} >= threshold=${ACCURACY_THRESHOLD} (tasks=${ACCURACY_TASKS}). Log: ${logf}" |
| 240 | return 0 |
| 241 | fi |
| 242 | log "FAIL: ${ACCURACY_METRIC}=${score} < threshold=${ACCURACY_THRESHOLD} (tasks=${ACCURACY_TASKS}). Log: ${logf}" |
| 243 | return 1 |
| 244 | } |
| 245 | |
| 246 | # ============================================================================ |
| 247 | # Server roles (prefill / decode) and their shared helpers |
| 248 | # ============================================================================ |
| 249 | |
| 250 | # Read model-specific flags from models.yaml (mode + role aware) into globals: |
| 251 | # MODEL_BASE_FLAGS / MODEL_ROLE_FLAGS / MODEL_EXPERIMENTAL_FLAGS -> MODEL_CONFIG |
| 252 | # Also exports the model's env: block (only if not already set; caller env wins). |
| 253 | load_model_flags() { |
| 254 | [[ -f "${MODELS_YAML}" ]] || die "models.yaml not found: ${MODELS_YAML}" |
| 255 | export MODELS_YAML MODEL_NAME PARALLEL_MODE ROLE |
| 256 | eval "$(python3 - <<'PY' |
| 257 | import os, shlex, sys, yaml |
| 258 | path = os.environ["MODELS_YAML"]; name = os.environ["MODEL_NAME"] |
| 259 | mode = os.environ["PARALLEL_MODE"]; role = os.environ["ROLE"] |
| 260 | with open(path, "r", encoding="utf-8") as f: |
| 261 | doc = yaml.safe_load(f) or {} |
| 262 | # Preferred form: top-level `models:` list of {model: <name>, ...} entries. |
| 263 | # Back-compat: a top-level mapping keyed by model name. |
| 264 | entries = doc.get("models") if isinstance(doc, dict) else doc |
| 265 | cfg = None |
| 266 | if isinstance(entries, list): |
| 267 | cfg = next((e for e in entries if isinstance(e, dict) and e.get("model") == name), None) |
| 268 | elif isinstance(doc, dict): |
| 269 | cfg = doc.get(name) |
| 270 | if cfg is None: |
| 271 | print(f'echo "ERROR: model {name} not found in {path}" >&2; exit 1'); sys.exit(0) |
| 272 | role_cfg = cfg.get(role, {}) or {} |
| 273 | def q(v): return shlex.quote(str(v if v is not None else "")) |
| 274 | exports = { |
| 275 | "MODEL_BASE_FLAGS": cfg.get("base_flags", "") or "", |
| 276 | "MODEL_ROLE_FLAGS": role_cfg.get(mode, "") or "", |
| 277 | "MODEL_EXPERIMENTAL_FLAGS": cfg.get("experimental_flags", "") or "", |
| 278 | } |
| 279 | for k, v in exports.items(): print(f"{k}={q(v)}") |
| 280 | # Model-specific env: exported only if not already set (caller env wins). |
| 281 | for k, v in (cfg.get("env", {}) or {}).items(): |
| 282 | if k not in os.environ: |
| 283 | print(f"export {k}={q(v)}") |
| 284 | PY |
| 285 | )" |
| 286 | MODEL_CONFIG="${MODEL_BASE_FLAGS} ${MODEL_ROLE_FLAGS} ${MODEL_EXPERIMENTAL_FLAGS}" |
| 287 | } |
| 288 | |
| 289 | # Build the TP-mode (WIDE_EP_MODE=0) serve command into the global CMD array. |
| 290 | # Each node is an independent TP server registered to the proxy. |
| 291 | build_tp_cmd() { |
| 292 | local port="${TP_PORT}" |
| 293 | [[ -n "${HOST_IP}" ]] && CMD+=(--host "${HOST_IP}") |
| 294 | CMD+=(--port "${port}" --tensor-parallel-size "${TP_SIZE}") |
| 295 | local _mc; read -ra _mc <<< "${MODEL_CONFIG}"; CMD+=("${_mc[@]}") |
| 296 | local kv_cfg read_mode_kv |
| 297 | read_mode_kv="$(moriio_read_mode_kv)" |
| 298 | kv_cfg=$(cat <<EOF |
| 299 | { |
| 300 | "kv_connector": "MoRIIOConnector", |
| 301 | "kv_role": "${KV_ROLE}", |
| 302 | "kv_connector_extra_config": { |
| 303 | "proxy_ip": "${PROXY_IP}", |
| 304 | "proxy_ping_port": "${PROXY_PING_PORT}", |
| 305 | "http_port": "${port}", |
| 306 | "handshake_port": "${HANDSHAKE_PORT}", |
| 307 | "notify_port": "${NOTIFY_PORT}"${read_mode_kv} |
| 308 | } |
| 309 | } |
| 310 | EOF |
| 311 | ) |
| 312 | CMD+=(--kv-transfer-config "${kv_cfg}") |
| 313 | log "role=${ROLE} mode=tp host=${HOST_IP:-0.0.0.0}:${port} kv_role=${KV_ROLE} tp=${TP_SIZE} read_mode=${MORIIO_READ_MODE}" |
| 314 | } |
| 315 | |
| 316 | # Build the EP-mode (WIDE_EP_MODE=1) serve command into the global CMD array. |
| 317 | # DP + expert parallel; masters expose the API server + KV transfer, children |
| 318 | # join headless. |
| 319 | build_ep_cmd() { |
| 320 | # prefill -> mori_high_throughput (paired with --enforce-eager in models.yaml) |
| 321 | # decode -> mori_low_latency (paired with CUDA graphs in models.yaml) |
| 322 | local default_backend="mori_low_latency" |
| 323 | [[ "${ROLE}" == "prefill" ]] && default_backend="mori_high_throughput" |
| 324 | local backend="${ALL2ALL_BACKEND:-${default_backend}}" |
| 325 | export VLLM_ALL2ALL_BACKEND="${VLLM_ALL2ALL_BACKEND:-${backend}}" |
| 326 | local port="${SERVE_PORT}" |
| 327 | CMD+=( |
| 328 | -tp 1 |
| 329 | --data-parallel-size "${DP_GROUP_SIZE}" |
| 330 | --data-parallel-size-local "${GPUS_PER_NODE}" |
| 331 | --data-parallel-address "${DP_MASTER_ADDR}" |
| 332 | --data-parallel-rpc-port "${RPC_PORT}" |
| 333 | --enable-expert-parallel |
| 334 | --all2all-backend "${backend}" |
| 335 | --port "${port}" |
| 336 | --no-enable-prefix-caching |
| 337 | --distributed-timeout-seconds "${DISTRIBUTED_TIMEOUT_SECONDS}" |
| 338 | ) |
| 339 | local _mc; read -ra _mc <<< "${MODEL_CONFIG}"; CMD+=("${_mc[@]}") |
| 340 | |
| 341 | if [[ "${IS_MASTER}" == "1" ]]; then |
| 342 | CMD+=(--api-server-count="${GPUS_PER_NODE}") |
| 343 | local kv_cfg read_mode_kv |
| 344 | read_mode_kv="$(moriio_read_mode_kv)" |
| 345 | kv_cfg=$(cat <<EOF |
| 346 | { |
| 347 | "kv_connector": "MoRIIOConnector", |
| 348 | "kv_role": "${KV_ROLE}", |
| 349 | "kv_port": "${KV_PORT}", |
| 350 | "kv_connector_extra_config": { |
| 351 | "proxy_ip": "${PROXY_IP}", |
| 352 | "proxy_port": "${PROXY_PORT}", |
| 353 | "proxy_ping_port": "${PROXY_PING_PORT}", |
| 354 | "http_port": "${port}", |
| 355 | "local_ping_port": "${LOCAL_PING_PORT}", |
| 356 | "handshake_port": "${HANDSHAKE_PORT}", |
| 357 | "notify_port": "${NOTIFY_PORT}"${read_mode_kv} |
| 358 | } |
| 359 | } |
| 360 | EOF |
| 361 | ) |
| 362 | CMD+=(--kv-transfer-config "${kv_cfg}") |
| 363 | log "role=${ROLE} mode=ep MASTER node_rank=${NODE_RANK} dp_size=${DP_GROUP_SIZE} dp_addr=${DP_MASTER_ADDR} kv_role=${KV_ROLE} all2all=${backend} read_mode=${MORIIO_READ_MODE}" |
| 364 | else |
| 365 | CMD+=(--data-parallel-start-rank "${DP_START_RANK}" --headless) |
| 366 | log "role=${ROLE} mode=ep CHILD node_rank=${NODE_RANK} dp_size=${DP_GROUP_SIZE} dp_start_rank=${DP_START_RANK} dp_addr=${DP_MASTER_ADDR} all2all=${backend}" |
| 367 | fi |
| 368 | } |
| 369 | |
| 370 | # Set the prefill topology globals for the current NODE_RANK (no execution). |
| 371 | configure_prefill() { |
| 372 | : "${NODE_RANK:=0}" # manual default: prefill master |
| 373 | KV_ROLE="kv_producer" |
| 374 | DP_GROUP_SIZE=$(( xP * GPUS_PER_NODE )) |
| 375 | DP_MASTER_ADDR="${PREFILL_MASTER_ADDR}" |
| 376 | DP_START_RANK=$(( NODE_RANK * GPUS_PER_NODE )) |
| 377 | [[ "${NODE_RANK}" -eq 0 ]] && IS_MASTER=1 || IS_MASTER=0 |
| 378 | TP_PORT="${PREFILL_PORT}" |
| 379 | } |
| 380 | |
| 381 | # Set the decode topology globals for the current NODE_RANK (no execution). |
| 382 | configure_decode() { |
| 383 | : "${NODE_RANK:=${xP}}" # manual default: decode master |
| 384 | KV_ROLE="kv_consumer" |
| 385 | DP_GROUP_SIZE=$(( yD * GPUS_PER_NODE )) |
| 386 | DP_MASTER_ADDR="${DECODE_MASTER_ADDR}" |
| 387 | DP_START_RANK=$(( (NODE_RANK - xP) * GPUS_PER_NODE )) |
| 388 | [[ "${NODE_RANK}" -eq "${xP}" ]] && IS_MASTER=1 || IS_MASTER=0 |
| 389 | TP_PORT="${DECODE_PORT}" |
| 390 | } |
| 391 | |
| 392 | # Build the serve command (model flags + CMD array + LOGF). No exec. |
| 393 | # Expects role globals set by configure_prefill/configure_decode. |
| 394 | build_server_cmd() { |
| 395 | HOST_IP="${IP_ARRAY[$NODE_RANK]:-}" # own IP for --host binding, if known |
| 396 | load_model_flags |
| 397 | LOGF="${LOG_PATH}/${ROLE}_${PARALLEL_MODE}_node${NODE_RANK}_$(date +%Y%m%d_%H%M%S).log" |
| 398 | CMD=(vllm serve "${MODEL_PATH}") |
| 399 | if [[ "${WIDE_EP_MODE}" == "0" ]]; then build_tp_cmd; else build_ep_cmd; fi |
| 400 | log "model=${MODEL_PATH}" |
| 401 | } |
| 402 | |
| 403 | # Foreground server run (explicit prefill/decode roles — original behavior). |
| 404 | run_server_fg() { |
| 405 | build_server_cmd |
| 406 | log "log=${LOGF}" |
| 407 | "${CMD[@]}" 2>&1 | tee "${LOGF}" |
| 408 | } |
| 409 | |
| 410 | run_prefill() { configure_prefill; run_server_fg; } |
| 411 | run_decode() { configure_decode; run_server_fg; } |
| 412 | |
| 413 | # ============================================================================ |
| 414 | # node mode — SLURM-native rank-based self-select + rank-0 orchestration |
| 415 | # ============================================================================ |
| 416 | |
| 417 | # Endpoints to health-check, one per line as "ip:port", depending on mode. |
| 418 | # tp: every prefill node (:PREFILL_PORT) and every decode node (:DECODE_PORT) |
| 419 | # ep: only the masters expose an API (prefill rank 0, decode rank xP) :SERVE_PORT |
| 420 | health_endpoints() { |
| 421 | local i |
| 422 | if [[ "${WIDE_EP_MODE}" == "0" ]]; then |
| 423 | for (( i=0; i<xP; i++ )); do echo "${IP_ARRAY[$i]}:${PREFILL_PORT}"; done |
| 424 | for (( i=xP; i<xP+yD; i++ )); do echo "${IP_ARRAY[$i]}:${DECODE_PORT}"; done |
| 425 | else |
| 426 | echo "${IP_ARRAY[0]}:${SERVE_PORT}" |
| 427 | echo "${IP_ARRAY[$xP]}:${SERVE_PORT}" |
| 428 | fi |
| 429 | } |
| 430 | |
| 431 | # Block until every endpoint's /health is OK, or timeout, or our local server dies. |
| 432 | wait_all_healthy() { |
| 433 | local deadline=$(( $(date +%s) + HEALTH_TIMEOUT_S )) |
| 434 | local eps; mapfile -t eps < <(health_endpoints) |
| 435 | log "health-gate: waiting on ${#eps[@]} endpoint(s): ${eps[*]}" |
| 436 | local ep |
| 437 | for ep in "${eps[@]}"; do |
| 438 | until /usr/bin/curl -sf "http://${ep}/health" >/dev/null 2>&1; do |
| 439 | (( $(date +%s) >= deadline )) && { log "TIMEOUT waiting for ${ep}"; return 1; } |
| 440 | kill -0 "${SERVER_PID}" 2>/dev/null || { log "local server (pid ${SERVER_PID}) exited while waiting"; return 1; } |
| 441 | sleep 10 |
| 442 | done |
| 443 | log "healthy: ${ep}" |
| 444 | done |
| 445 | return 0 |
| 446 | } |
| 447 | |
| 448 | run_workload() { |
| 449 | case "${RUN_AFTER_HEALTH}" in |
| 450 | bench) run_bench ;; |
| 451 | accuracy) run_accuracy ;; |
| 452 | none|"") log "RUN_AFTER_HEALTH=none — skipping workload" ;; |
| 453 | *) die "unknown RUN_AFTER_HEALTH='${RUN_AFTER_HEALTH}'" ;; |
| 454 | esac |
| 455 | } |
| 456 | |
| 457 | # rank 0: proxy + health-gate + workload, then write the completion sentinel. |
| 458 | orchestrate_master() { |
| 459 | local sentinel="$1" rc=0 |
| 460 | # Front door: the proxy runs in-container (started here); the vllm-router |
| 461 | # runs as a SEPARATE container started by the SLURM job on this (rank-0) node, |
| 462 | # so in that mode we don't start anything here. |
| 463 | PROXY_PID="" |
| 464 | if [[ "${ROUTER_TYPE:-vllm-router}" == "vllm-router" ]]; then |
| 465 | log "ROUTER_TYPE=vllm-router: external router expected on gateway :${GATEWAY_PORT:-${ROUTER_PORT}} (not starting proxy)" |
| 466 | else |
| 467 | start_proxy_bg |
| 468 | fi |
| 469 | if wait_all_healthy; then |
| 470 | # If using the external vllm-router, wait for it to become ready before |
| 471 | # running the workload. The router starts as a separate container on this |
| 472 | # (rank-0) node; give it up to 300s to bind port ROUTER_PORT. |
| 473 | if [[ "${ROUTER_TYPE:-vllm-router}" == "vllm-router" ]]; then |
| 474 | local router_deadline=$(( $(date +%s) + 300 )) |
| 475 | log "waiting for vllm-router on :${ROUTER_PORT} (up to 300s)" |
| 476 | until /usr/bin/curl -sf "http://127.0.0.1:${ROUTER_PORT}/health" >/dev/null 2>&1; do |
| 477 | if (( $(date +%s) >= router_deadline )); then |
| 478 | log "FAIL: vllm-router not ready on :${ROUTER_PORT} after 300s (bring-up)" |
| 479 | rc=1 |
| 480 | break |
| 481 | fi |
| 482 | sleep 5 |
| 483 | done |
| 484 | (( rc == 0 )) && log "vllm-router healthy on :${ROUTER_PORT}" |
| 485 | fi |
| 486 | if (( rc == 0 )); then |
| 487 | set +e |
| 488 | run_workload |
| 489 | rc=$? |
| 490 | set -e |
| 491 | fi |
| 492 | else |
| 493 | log "FAIL: health-gate did not pass — server bring-up failed" |
| 494 | rc=1 |
| 495 | fi |
| 496 | echo "${rc}" > "${sentinel}" 2>/dev/null || true |
| 497 | log "master: workload rc=${rc}; sentinel=${sentinel}; tearing down local proxy+server" |
| 498 | [[ -n "${PROXY_PID:-}" ]] && kill "${PROXY_PID}" 2>/dev/null || true |
| 499 | kill "${SERVER_PID}" 2>/dev/null || true |
| 500 | wait "${SERVER_PID}" 2>/dev/null || true |
| 501 | return "${rc}" |
| 502 | } |
| 503 | |
| 504 | # non-master ranks: keep serving until the master signals completion (sentinel) |
| 505 | # or our local server dies unexpectedly. |
| 506 | watch_until_done() { |
| 507 | local sentinel="$1" rc=0 |
| 508 | log "rank ${NODE_RANK} serving; waiting for completion sentinel (${sentinel})" |
| 509 | while :; do |
| 510 | [[ -f "${sentinel}" ]] && { log "completion sentinel seen; shutting down"; break; } |
| 511 | if ! kill -0 "${SERVER_PID}" 2>/dev/null; then |
| 512 | log "local server (pid ${SERVER_PID}) exited before completion"; rc=1; break |
| 513 | fi |
| 514 | sleep 10 |
| 515 | done |
| 516 | kill "${SERVER_PID}" 2>/dev/null || true |
| 517 | wait "${SERVER_PID}" 2>/dev/null || true |
| 518 | return "${rc}" |
| 519 | } |
| 520 | |
| 521 | run_node() { |
| 522 | NODE_RANK="${NODE_RANK:-${SLURM_PROCID:-0}}" |
| 523 | local total=$(( xP + yD )) |
| 524 | (( NODE_RANK >= 0 && NODE_RANK < total )) \ |
| 525 | || die "NODE_RANK=${NODE_RANK} out of range [0,${total}) (xP=${xP} yD=${yD})" |
| 526 | |
| 527 | if (( NODE_RANK < xP )); then ROLE="prefill"; configure_prefill |
| 528 | else ROLE="decode"; configure_decode |
| 529 | fi |
| 530 | |
| 531 | # Shared-FS completion sentinel (LOG_PATH is shared & per-run/per-job). |
| 532 | local sentinel="${LOG_PATH}/.disagg_done" |
| 533 | |
| 534 | log "node mode: rank=${NODE_RANK}/${total} role=${ROLE} master=${IS_MASTER} wide_ep=${WIDE_EP_MODE}(${PARALLEL_MODE})" |
| 535 | log "topology: xP=${xP} yD=${yD} gpus/node=${GPUS_PER_NODE} IPADDRS=${IPADDRS}" |
| 536 | |
| 537 | # DRY_RUN: resolve+print the plan (role, serve cmd, endpoints) and exit 0. |
| 538 | # Validates wiring without launching vLLM / the proxy / any workload. |
| 539 | if [[ "${DRY_RUN:-0}" == "1" ]]; then |
| 540 | build_server_cmd |
| 541 | log "DRY_RUN: server cmd: ${CMD[*]}" |
| 542 | if (( NODE_RANK == 0 )); then |
| 543 | local eps; mapfile -t eps < <(health_endpoints) |
| 544 | log "DRY_RUN: health endpoints: ${eps[*]}" |
| 545 | log "DRY_RUN: post-health workload: ${RUN_AFTER_HEALTH}" |
| 546 | log "DRY_RUN: sentinel: ${sentinel}" |
| 547 | fi |
| 548 | log "DRY_RUN complete (rank ${NODE_RANK}) — no processes launched" |
| 549 | exit 0 |
| 550 | fi |
| 551 | |
| 552 | # aiter and flydsl keep runtime JIT/kernel caches inside the aiter install |
| 553 | # tree (…/aiter/jit and …/aiter/jit/flydsl_cache), which is read-only for the |
| 554 | # non-root (--user) container. AITER_JIT_DIR / FLYDSL_RUNTIME_CACHE_DIR |
| 555 | # redirect those to node-local /tmp (both are sanctioned overrides: aiter only |
| 556 | # points flydsl at its bundled cache when FLYDSL_RUNTIME_CACHE_DIR is unset). |
| 557 | # Seed each once from the image's bundled copy so we keep the prebuilt |
| 558 | # modules/kernels (fast, no NFS, no from-source rebuild / MLIR recompile). |
| 559 | local _aiter_src |
| 560 | _aiter_src="$(python3 -c 'import os,aiter; print(os.path.join(os.path.dirname(aiter.__file__),"jit"))' 2>/dev/null || true)" |
| 561 | if [[ -n "${AITER_JIT_DIR:-}" && ! -e "${AITER_JIT_DIR}/.seeded" ]]; then |
| 562 | if [[ -n "${_aiter_src}" && -d "${_aiter_src}" ]]; then |
| 563 | mkdir -p "${AITER_JIT_DIR}" |
| 564 | cp -a "${_aiter_src}/." "${AITER_JIT_DIR}/" 2>/dev/null || true |
| 565 | touch "${AITER_JIT_DIR}/.seeded" 2>/dev/null || true |
| 566 | log "seeded AITER_JIT_DIR=${AITER_JIT_DIR} from ${_aiter_src}" |
| 567 | else |
| 568 | log "WARN: could not locate aiter jit dir to seed AITER_JIT_DIR=${AITER_JIT_DIR}; aiter may rebuild from source" |
| 569 | fi |
| 570 | fi |
| 571 | if [[ -n "${FLYDSL_RUNTIME_CACHE_DIR:-}" && ! -e "${FLYDSL_RUNTIME_CACHE_DIR}/.seeded" ]]; then |
| 572 | mkdir -p "${FLYDSL_RUNTIME_CACHE_DIR}" |
| 573 | if [[ -n "${_aiter_src}" && -d "${_aiter_src}/flydsl_cache" ]]; then |
| 574 | cp -a "${_aiter_src}/flydsl_cache/." "${FLYDSL_RUNTIME_CACHE_DIR}/" 2>/dev/null || true |
| 575 | log "seeded FLYDSL_RUNTIME_CACHE_DIR=${FLYDSL_RUNTIME_CACHE_DIR} from ${_aiter_src}/flydsl_cache" |
| 576 | else |
| 577 | log "FLYDSL_RUNTIME_CACHE_DIR=${FLYDSL_RUNTIME_CACHE_DIR} set; no bundled flydsl_cache to seed (will JIT-compile at runtime)" |
| 578 | fi |
| 579 | touch "${FLYDSL_RUNTIME_CACHE_DIR}/.seeded" 2>/dev/null || true |
| 580 | fi |
| 581 | |
| 582 | # mori JIT-compiles its shmem/all2all kernels to ~/.mori/jit/<arch>_<nic>/ on |
| 583 | # first EP use. HOME here is /workspace, bind-mounted to persistent+shared NFS |
| 584 | # (/data/$USER), so a wrong-arch cache from an earlier run/build survives and |
| 585 | # gets reloaded even after MORI_GPU_ARCHS changes -- an existing cache dir |
| 586 | # wins over the arch setting. Reloading a gfx942 shmem_kernels.hsaco on gfx950 |
| 587 | # hardware fails with "device kernel image is invalid" and every EP worker |
| 588 | # dies at init (surfacing async as a torch.tensor HIP error). Guard: on EP |
| 589 | # runs, if the cache has no build for our target arch, purge it so mori |
| 590 | # recompiles for MORI_GPU_ARCHS. A matching arch dir is kept (fast reload). |
| 591 | if [[ "${WIDE_EP_MODE}" == "1" && -n "${MORI_GPU_ARCHS:-}" ]]; then |
| 592 | local _mori_jit="${HOME:-/workspace}/.mori/jit" |
| 593 | if [[ -d "${_mori_jit}" ]] && ! compgen -G "${_mori_jit}/${MORI_GPU_ARCHS}_*" >/dev/null 2>&1; then |
| 594 | log "purging stale mori jit cache ${_mori_jit} (no ${MORI_GPU_ARCHS}_* build present; forcing recompile for ${MORI_GPU_ARCHS})" |
| 595 | rm -rf "${_mori_jit}" 2>/dev/null || true |
| 596 | fi |
| 597 | fi |
| 598 | |
| 599 | # Start this node's server in the background (full server log -> LOGF; this |
| 600 | # shell's stdout stays free for orchestration logs the srun/CI captures). |
| 601 | build_server_cmd |
| 602 | "${CMD[@]}" >"${LOGF}" 2>&1 & |
| 603 | SERVER_PID=$! |
| 604 | log "server started pid=${SERVER_PID} log=${LOGF}" |
| 605 | |
| 606 | local rc=0 |
| 607 | if (( NODE_RANK == 0 )); then |
| 608 | orchestrate_master "${sentinel}" || rc=$? |
| 609 | # Canonical, human-readable end-of-run verdict for the login-node poller |
| 610 | # and CI logs (the .disagg_done sentinel remains the machine authority). |
| 611 | log "VERDICT: $( (( rc == 0 )) && echo PASS || echo FAIL) rc=${rc}" |
| 612 | else |
| 613 | watch_until_done "${sentinel}" || rc=$? |
| 614 | fi |
| 615 | log "node rank=${NODE_RANK} exiting rc=${rc}" |
| 616 | exit "${rc}" |
| 617 | } |
| 618 | |
| 619 | # ============================================================================ |
| 620 | main() { |
| 621 | parse_args "$@" |
| 622 | load_config |
| 623 | resolve_topology |
| 624 | case "${ROLE}" in |
| 625 | node) run_node ;; |
| 626 | proxy) run_proxy ;; |
| 627 | bench) run_bench ;; |
| 628 | accuracy) run_accuracy ;; |
| 629 | prefill) run_prefill ;; |
| 630 | decode) run_decode ;; |
| 631 | *) usage ;; |
| 632 | esac |
| 633 | } |
| 634 | |
| 635 | main "$@" |