| 1 | #!/usr/bin/env python3 |
| 2 | # SPDX-License-Identifier: Apache-2.0 |
| 3 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 4 | |
| 5 | """Select one measured scheduler recommendation from vLLM sweep results.""" |
| 6 | |
| 7 | from __future__ import annotations |
| 8 | |
| 9 | import argparse |
| 10 | import json |
| 11 | from collections import defaultdict |
| 12 | from pathlib import Path |
| 13 | from statistics import mean, median |
| 14 | from typing import Any |
| 15 | |
| 16 | try: |
| 17 | import yaml |
| 18 | except ImportError as exc: |
| 19 | raise SystemExit("PyYAML is required. Install it with: pip install pyyaml") from exc |
| 20 | |
| 21 | |
| 22 | DEFAULT_CONFIG_PATH: str | None = None |
| 23 | DEFAULT_ENV_PATH: str | None = None |
| 24 | DEFAULT_TTFT_SLA_MS: float | None = None |
| 25 | DEFAULT_TPOT_SLA_MS: float | None = None |
| 26 | DEFAULT_MINIMUM_COMPLIANCE: float = 0.99 |
| 27 | |
| 28 | |
| 29 | def _number(value: object) -> float | None: |
| 30 | if isinstance(value, bool) or not isinstance(value, (int, float)): |
| 31 | return None |
| 32 | return float(value) |
| 33 | |
| 34 | |
| 35 | def _mean(rows: list[dict[str, Any]], key: str) -> float | None: |
| 36 | values = [_number(row.get(key)) for row in rows] |
| 37 | numbers = [value for value in values if value is not None] |
| 38 | return mean(numbers) if numbers else None |
| 39 | |
| 40 | |
| 41 | def _percentile_summary( |
| 42 | rows: list[dict[str, Any]], key: str |
| 43 | ) -> tuple[float | None, float | None, float | None]: |
| 44 | values = [_number(row.get(key)) for row in rows] |
| 45 | numbers = [value for value in values if value is not None] |
| 46 | if not numbers: |
| 47 | return None, None, None |
| 48 | return mean(numbers), median(numbers), max(numbers) |
| 49 | |
| 50 | |
| 51 | def _positive_int(row: dict[str, Any], key: str) -> int | None: |
| 52 | value = row.get(key) |
| 53 | if isinstance(value, bool) or not isinstance(value, int) or value <= 0: |
| 54 | return None |
| 55 | return value |
| 56 | |
| 57 | |
| 58 | def _resolve(script_dir: Path, value: str | None) -> Path | None: |
| 59 | if value is None: |
| 60 | return None |
| 61 | path = Path(value) |
| 62 | return path if path.is_absolute() else script_dir / path |
| 63 | |
| 64 | |
| 65 | def _load_runs(results_dir: Path) -> list[dict[str, Any]]: |
| 66 | rows: list[dict[str, Any]] = [] |
| 67 | for path in sorted(results_dir.rglob("summary.json")): |
| 68 | data = json.loads(path.read_text(encoding="utf-8")) |
| 69 | if not isinstance(data, list): |
| 70 | continue |
| 71 | for row in data: |
| 72 | if isinstance(row, dict): |
| 73 | copied = dict(row) |
| 74 | copied["_summary_path"] = str(path) |
| 75 | rows.append(copied) |
| 76 | return rows |
| 77 | |
| 78 | |
| 79 | def _aggregate_candidates( |
| 80 | rows: list[dict[str, Any]], |
| 81 | *, |
| 82 | use_goodput: bool, |
| 83 | ttft_sla_ms: float | None = None, |
| 84 | tpot_sla_ms: float | None = None, |
| 85 | minimum_compliance: float = DEFAULT_MINIMUM_COMPLIANCE, |
| 86 | ) -> list[dict[str, Any]]: |
| 87 | grouped: dict[tuple[int, int], list[dict[str, Any]]] = defaultdict(list) |
| 88 | |
| 89 | for row in rows: |
| 90 | seqs = _positive_int(row, "max_num_seqs") |
| 91 | batch = _positive_int(row, "max_num_batched_tokens") |
| 92 | if seqs is None or batch is None: |
| 93 | continue |
| 94 | grouped[(seqs, batch)].append(row) |
| 95 | |
| 96 | candidates: list[dict[str, Any]] = [] |
| 97 | for (seqs, batch), runs in sorted(grouped.items()): |
| 98 | failed_requests = sum(int(_number(run.get("failed")) or 0) for run in runs) |
| 99 | output_throughput = _mean(runs, "output_throughput") |
| 100 | request_throughput = _mean(runs, "request_throughput") |
| 101 | request_goodput = _mean(runs, "request_goodput") |
| 102 | mean_p99_ttft_ms, median_p99_ttft_ms, worst_p99_ttft_ms = _percentile_summary( |
| 103 | runs, "p99_ttft_ms" |
| 104 | ) |
| 105 | mean_p99_tpot_ms, median_p99_tpot_ms, worst_p99_tpot_ms = _percentile_summary( |
| 106 | runs, "p99_tpot_ms" |
| 107 | ) |
| 108 | |
| 109 | completed_requests = sum( |
| 110 | int(_number(run.get("completed")) or 0) for run in runs |
| 111 | ) |
| 112 | estimated_compliant_requests = sum( |
| 113 | (_number(run.get("request_goodput")) or 0.0) |
| 114 | * (_number(run.get("duration")) or 0.0) |
| 115 | for run in runs |
| 116 | ) |
| 117 | combined_compliance_ratio = ( |
| 118 | estimated_compliant_requests / completed_requests |
| 119 | if completed_requests > 0 |
| 120 | else None |
| 121 | ) |
| 122 | |
| 123 | valid = failed_requests == 0 and output_throughput is not None |
| 124 | reason = None |
| 125 | if failed_requests: |
| 126 | valid = False |
| 127 | reason = f"{failed_requests} failed request(s) across repeated runs" |
| 128 | elif use_goodput and request_goodput is None: |
| 129 | valid = False |
| 130 | reason = "request_goodput is missing; rerun sweep with SLO goodput enabled" |
| 131 | |
| 132 | p99_sla_eligible = valid |
| 133 | if ttft_sla_ms is not None: |
| 134 | p99_sla_eligible = bool( |
| 135 | p99_sla_eligible |
| 136 | and median_p99_ttft_ms is not None |
| 137 | and median_p99_ttft_ms <= ttft_sla_ms |
| 138 | ) |
| 139 | if tpot_sla_ms is not None: |
| 140 | p99_sla_eligible = bool( |
| 141 | p99_sla_eligible |
| 142 | and median_p99_tpot_ms is not None |
| 143 | and median_p99_tpot_ms <= tpot_sla_ms |
| 144 | ) |
| 145 | compliance_eligible = bool( |
| 146 | valid |
| 147 | and combined_compliance_ratio is not None |
| 148 | and combined_compliance_ratio >= minimum_compliance |
| 149 | ) |
| 150 | |
| 151 | candidates.append( |
| 152 | { |
| 153 | "max_num_seqs": seqs, |
| 154 | "max_num_batched_tokens": batch, |
| 155 | "run_count": len(runs), |
| 156 | "failed_requests": failed_requests, |
| 157 | "mean_request_goodput": request_goodput, |
| 158 | "mean_request_throughput": request_throughput, |
| 159 | "mean_output_throughput": output_throughput, |
| 160 | "mean_p99_ttft_ms": mean_p99_ttft_ms, |
| 161 | "median_p99_ttft_ms": median_p99_ttft_ms, |
| 162 | "worst_p99_ttft_ms": worst_p99_ttft_ms, |
| 163 | "mean_p99_tpot_ms": mean_p99_tpot_ms, |
| 164 | "median_p99_tpot_ms": median_p99_tpot_ms, |
| 165 | "worst_p99_tpot_ms": worst_p99_tpot_ms, |
| 166 | "estimated_compliant_requests": estimated_compliant_requests, |
| 167 | "completed_requests": completed_requests, |
| 168 | "combined_compliance_ratio": combined_compliance_ratio, |
| 169 | "combined_compliance_percent": ( |
| 170 | combined_compliance_ratio * 100 |
| 171 | if combined_compliance_ratio is not None |
| 172 | else None |
| 173 | ), |
| 174 | "benchmark_valid": valid, |
| 175 | "p99_sla_eligible": p99_sla_eligible, |
| 176 | "compliance_eligible": compliance_eligible, |
| 177 | "sla_eligible": p99_sla_eligible and compliance_eligible, |
| 178 | "valid": valid, |
| 179 | "invalid_reason": reason, |
| 180 | "summary_files": sorted({str(run["_summary_path"]) for run in runs}), |
| 181 | } |
| 182 | ) |
| 183 | |
| 184 | return candidates |
| 185 | |
| 186 | |
| 187 | def _select_candidate( |
| 188 | candidates: list[dict[str, Any]], |
| 189 | *, |
| 190 | use_goodput: bool, |
| 191 | ) -> tuple[dict[str, Any] | None, dict[str, Any], str]: |
| 192 | valid = [candidate for candidate in candidates if candidate["valid"]] |
| 193 | if not valid: |
| 194 | raise ValueError( |
| 195 | "No valid sweep configuration was found. Inspect the sweep results " |
| 196 | "for failed requests or missing metrics." |
| 197 | ) |
| 198 | |
| 199 | if use_goodput: |
| 200 | objective = "p99_constrained_throughput" |
| 201 | best_effort = max( |
| 202 | valid, |
| 203 | key=lambda candidate: ( |
| 204 | candidate["mean_request_goodput"], |
| 205 | candidate["combined_compliance_ratio"] or 0.0, |
| 206 | candidate["mean_output_throughput"], |
| 207 | -candidate["max_num_batched_tokens"], |
| 208 | -candidate["max_num_seqs"], |
| 209 | ), |
| 210 | ) |
| 211 | if best_effort["mean_request_goodput"] <= 0: |
| 212 | raise ValueError( |
| 213 | "No sweep configuration produced non-zero request goodput for " |
| 214 | "the supplied TTFT/TPOT objectives." |
| 215 | ) |
| 216 | eligible = [candidate for candidate in valid if candidate["sla_eligible"]] |
| 217 | winner = ( |
| 218 | max( |
| 219 | eligible, |
| 220 | key=lambda candidate: ( |
| 221 | candidate["mean_output_throughput"], |
| 222 | candidate["combined_compliance_ratio"], |
| 223 | candidate["mean_request_goodput"], |
| 224 | -candidate["max_num_batched_tokens"], |
| 225 | -candidate["max_num_seqs"], |
| 226 | ), |
| 227 | ) |
| 228 | if eligible |
| 229 | else None |
| 230 | ) |
| 231 | else: |
| 232 | objective = "highest_mean_output_throughput" |
| 233 | winner = max( |
| 234 | valid, |
| 235 | key=lambda candidate: ( |
| 236 | candidate["mean_output_throughput"], |
| 237 | -candidate["max_num_batched_tokens"], |
| 238 | -candidate["max_num_seqs"], |
| 239 | ), |
| 240 | ) |
| 241 | best_effort = winner |
| 242 | |
| 243 | return winner, best_effort, objective |
| 244 | |
| 245 | |
| 246 | def _load_config(path: Path) -> dict[str, Any]: |
| 247 | data = yaml.safe_load(path.read_text(encoding="utf-8")) |
| 248 | if not isinstance(data, dict): |
| 249 | raise ValueError(f"{path} does not contain a YAML configuration object.") |
| 250 | return data |
| 251 | |
| 252 | |
| 253 | def _write_config( |
| 254 | path: Path, |
| 255 | *, |
| 256 | source_path: Path, |
| 257 | config: dict[str, Any], |
| 258 | objective: str, |
| 259 | ) -> None: |
| 260 | body = yaml.safe_dump( |
| 261 | config, |
| 262 | sort_keys=False, |
| 263 | default_flow_style=False, |
| 264 | allow_unicode=True, |
| 265 | ) |
| 266 | header = ( |
| 267 | "# Generated from the initial recipe config after a vLLM parameter sweep.\n" |
| 268 | f"# Initial config: {source_path}\n" |
| 269 | f"# Selection objective: {objective}\n" |
| 270 | ) |
| 271 | path.write_text(header + body, encoding="utf-8") |
| 272 | |
| 273 | |
| 274 | def parse_args() -> argparse.Namespace: |
| 275 | parser = argparse.ArgumentParser( |
| 276 | description=( |
| 277 | "Select one recommended max-num-seqs/max-num-batched-tokens pair " |
| 278 | "from generated vLLM sweep results." |
| 279 | ) |
| 280 | ) |
| 281 | parser.add_argument( |
| 282 | "--results-dir", |
| 283 | default="results/runtime-tuning", |
| 284 | help="Sweep experiment directory (default: results/runtime-tuning).", |
| 285 | ) |
| 286 | parser.add_argument( |
| 287 | "--config", |
| 288 | default=DEFAULT_CONFIG_PATH, |
| 289 | help="Initial config.yml used by the sweep.", |
| 290 | ) |
| 291 | parser.add_argument( |
| 292 | "--env", |
| 293 | default=DEFAULT_ENV_PATH, |
| 294 | help="Environment file to show in the deployment instructions.", |
| 295 | ) |
| 296 | parser.add_argument( |
| 297 | "--ttft-sla-ms", |
| 298 | type=float, |
| 299 | default=DEFAULT_TTFT_SLA_MS, |
| 300 | help="TTFT objective used to generate request goodput.", |
| 301 | ) |
| 302 | parser.add_argument( |
| 303 | "--tpot-sla-ms", |
| 304 | type=float, |
| 305 | default=DEFAULT_TPOT_SLA_MS, |
| 306 | help="TPOT objective used to generate request goodput.", |
| 307 | ) |
| 308 | parser.add_argument( |
| 309 | "--minimum-compliance", |
| 310 | type=float, |
| 311 | default=DEFAULT_MINIMUM_COMPLIANCE, |
| 312 | help=( |
| 313 | "Minimum duration-weighted fraction of completed requests that must " |
| 314 | "meet all supplied objectives (default: 0.99)." |
| 315 | ), |
| 316 | ) |
| 317 | parser.add_argument( |
| 318 | "--output-config", |
| 319 | default="recommended-config.yml", |
| 320 | help="Recommended config output path.", |
| 321 | ) |
| 322 | parser.add_argument( |
| 323 | "--output-json", |
| 324 | default="recommendation.json", |
| 325 | help="Recommendation evidence output path.", |
| 326 | ) |
| 327 | return parser.parse_args() |
| 328 | |
| 329 | |
| 330 | def main() -> int: |
| 331 | args = parse_args() |
| 332 | if not 0.0 < args.minimum_compliance <= 1.0: |
| 333 | raise ValueError("--minimum-compliance must be greater than 0 and at most 1.") |
| 334 | script_dir = Path(__file__).resolve().parent |
| 335 | |
| 336 | config_path = _resolve(script_dir, args.config) |
| 337 | env_path = _resolve(script_dir, args.env) |
| 338 | results_dir = _resolve(script_dir, args.results_dir) |
| 339 | output_config = _resolve(script_dir, args.output_config) |
| 340 | output_json = _resolve(script_dir, args.output_json) |
| 341 | |
| 342 | if config_path is None: |
| 343 | raise ValueError("--config is required.") |
| 344 | assert results_dir is not None |
| 345 | assert output_config is not None |
| 346 | assert output_json is not None |
| 347 | |
| 348 | if not results_dir.exists(): |
| 349 | raise ValueError( |
| 350 | f"Sweep results were not found at {results_dir}. Run the sweep first." |
| 351 | ) |
| 352 | |
| 353 | rows = _load_runs(results_dir) |
| 354 | if not rows: |
| 355 | raise ValueError(f"No summary.json sweep results found under {results_dir}.") |
| 356 | |
| 357 | use_goodput = args.ttft_sla_ms is not None or args.tpot_sla_ms is not None |
| 358 | candidates = _aggregate_candidates( |
| 359 | rows, |
| 360 | use_goodput=use_goodput, |
| 361 | ttft_sla_ms=args.ttft_sla_ms, |
| 362 | tpot_sla_ms=args.tpot_sla_ms, |
| 363 | minimum_compliance=args.minimum_compliance, |
| 364 | ) |
| 365 | winner, best_effort, objective = _select_candidate( |
| 366 | candidates, use_goodput=use_goodput |
| 367 | ) |
| 368 | |
| 369 | initial_config = _load_config(config_path) |
| 370 | if winner is not None: |
| 371 | recommended_config = dict(initial_config) |
| 372 | recommended_config["max-num-seqs"] = winner["max_num_seqs"] |
| 373 | recommended_config["max-num-batched-tokens"] = winner["max_num_batched_tokens"] |
| 374 | _write_config( |
| 375 | output_config, |
| 376 | source_path=config_path, |
| 377 | config=recommended_config, |
| 378 | objective=objective, |
| 379 | ) |
| 380 | elif output_config.exists(): |
| 381 | output_config.unlink() |
| 382 | |
| 383 | measured = None |
| 384 | recommended = None |
| 385 | if winner is not None: |
| 386 | recommended = { |
| 387 | "max_num_seqs": winner["max_num_seqs"], |
| 388 | "max_num_batched_tokens": winner["max_num_batched_tokens"], |
| 389 | } |
| 390 | measured = { |
| 391 | key: winner[key] |
| 392 | for key in ( |
| 393 | "run_count", |
| 394 | "mean_request_throughput", |
| 395 | "mean_request_goodput", |
| 396 | "mean_output_throughput", |
| 397 | "combined_compliance_ratio", |
| 398 | "combined_compliance_percent", |
| 399 | "estimated_compliant_requests", |
| 400 | "completed_requests", |
| 401 | "mean_p99_ttft_ms", |
| 402 | "median_p99_ttft_ms", |
| 403 | "worst_p99_ttft_ms", |
| 404 | "mean_p99_tpot_ms", |
| 405 | "median_p99_tpot_ms", |
| 406 | "worst_p99_tpot_ms", |
| 407 | ) |
| 408 | } |
| 409 | |
| 410 | recommendation = { |
| 411 | "status": ( |
| 412 | "sla_feasible" if winner is not None else "no_sla_feasible_configuration" |
| 413 | ), |
| 414 | "selection_objective": objective, |
| 415 | "minimum_compliance_ratio": args.minimum_compliance, |
| 416 | "slo": { |
| 417 | "ttft_ms": args.ttft_sla_ms, |
| 418 | "tpot_ms": args.tpot_sla_ms, |
| 419 | }, |
| 420 | "initial": { |
| 421 | "max_num_seqs": initial_config.get("max-num-seqs"), |
| 422 | "max_num_batched_tokens": initial_config.get("max-num-batched-tokens"), |
| 423 | }, |
| 424 | "recommended": recommended, |
| 425 | "measured": measured, |
| 426 | "best_effort": { |
| 427 | "max_num_seqs": best_effort["max_num_seqs"], |
| 428 | "max_num_batched_tokens": best_effort["max_num_batched_tokens"], |
| 429 | "mean_request_goodput": best_effort["mean_request_goodput"], |
| 430 | "combined_compliance_ratio": best_effort["combined_compliance_ratio"], |
| 431 | "p99_sla_eligible": best_effort["p99_sla_eligible"], |
| 432 | }, |
| 433 | "candidates": candidates, |
| 434 | } |
| 435 | output_json.write_text( |
| 436 | json.dumps(recommendation, indent=2) + "\n", |
| 437 | encoding="utf-8", |
| 438 | ) |
| 439 | |
| 440 | if winner is None: |
| 441 | print("No SLA-feasible runtime configuration was found.") |
| 442 | print(f"Wrote diagnostic evidence to {output_json}") |
| 443 | print("No recommended config was written.") |
| 444 | return 2 |
| 445 | |
| 446 | print("Recommended runtime configuration") |
| 447 | print() |
| 448 | print(f" max-num-seqs: {winner['max_num_seqs']}") |
| 449 | print(f" max-num-batched-tokens: {winner['max_num_batched_tokens']}") |
| 450 | print() |
| 451 | print(f"Selection objective: {objective}") |
| 452 | if use_goodput: |
| 453 | print(f" mean request goodput: {winner['mean_request_goodput']:.2f} req/s") |
| 454 | print(f" combined compliance: {winner['combined_compliance_percent']:.2f}%") |
| 455 | print(f" mean output throughput: {winner['mean_output_throughput']:.2f} tok/s") |
| 456 | if winner["median_p99_ttft_ms"] is not None: |
| 457 | print(f" median P99 TTFT: {winner['median_p99_ttft_ms']:.2f} ms") |
| 458 | if winner["median_p99_tpot_ms"] is not None: |
| 459 | print(f" median P99 TPOT: {winner['median_p99_tpot_ms']:.2f} ms") |
| 460 | print() |
| 461 | print(f"Wrote {output_config}") |
| 462 | print(f"Wrote {output_json}") |
| 463 | print() |
| 464 | print("Deploy:") |
| 465 | if env_path is not None: |
| 466 | print(f" source {env_path}") |
| 467 | print(f" vllm serve --config {output_config}") |
| 468 | return 0 |
| 469 | |
| 470 | |
| 471 | if __name__ == "__main__": |
| 472 | raise SystemExit(main()) |