| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | r"""Benchmark online serving throughput. |
| 4 | |
| 5 | On the server side, run one of the following commands |
| 6 | to launch the vLLM OpenAI API server: |
| 7 | vllm serve <your_model> <engine arguments> |
| 8 | |
| 9 | On the client side, run: |
| 10 | vllm bench serve \ |
| 11 | --backend <backend or endpoint type. Default 'openai'> \ |
| 12 | --label <benchmark result label. Default using backend> \ |
| 13 | --model <your_model. Optional, defaults to first model from server> \ |
| 14 | --dataset-name <dataset_name. Default 'random'> \ |
| 15 | --input-len <general input length. Optional, maps to dataset-specific args> \ |
| 16 | --output-len <general output length. Optional, maps to dataset-specific args> \ |
| 17 | --request-rate <request_rate. Default inf> \ |
| 18 | --num-prompts <num_prompts. Default 1000> |
| 19 | """ |
| 20 | |
| 21 | import argparse |
| 22 | import asyncio |
| 23 | import contextlib |
| 24 | import importlib.util |
| 25 | import json |
| 26 | import math |
| 27 | import os |
| 28 | import random |
| 29 | import shutil |
| 30 | import ssl |
| 31 | import time |
| 32 | import uuid |
| 33 | import warnings |
| 34 | from collections.abc import AsyncGenerator, Iterable, Iterator |
| 35 | from dataclasses import dataclass, replace |
| 36 | from datetime import datetime |
| 37 | from enum import Enum |
| 38 | from pathlib import Path |
| 39 | from typing import Any, Literal |
| 40 | |
| 41 | import aiohttp |
| 42 | import numpy as np |
| 43 | from tqdm.asyncio import tqdm |
| 44 | |
| 45 | from vllm.benchmarks.datasets import SampleRequest, add_dataset_parser, get_samples |
| 46 | from vllm.benchmarks.lib.endpoint_request_func import ( |
| 47 | ASYNC_REQUEST_FUNCS, |
| 48 | OPENAI_COMPATIBLE_BACKENDS, |
| 49 | POOLING_BACKENDS, |
| 50 | RequestFuncInput, |
| 51 | RequestFuncOutput, |
| 52 | ) |
| 53 | from vllm.benchmarks.lib.ready_checker import wait_for_endpoint |
| 54 | from vllm.benchmarks.lib.utils import ( |
| 55 | convert_to_pytorch_benchmark_format, |
| 56 | redact_sensitive_namespace, |
| 57 | write_to_json, |
| 58 | ) |
| 59 | from vllm.tokenizers import TokenizerLike, get_tokenizer |
| 60 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 61 | from vllm.utils.gc_utils import freeze_gc_heap |
| 62 | from vllm.utils.network_utils import join_host_port |
| 63 | |
| 64 | MILLISECONDS_TO_SECONDS_CONVERSION = 1000 |
| 65 | _SENSITIVE_ARG_FIELDS = ("header",) |
| 66 | |
| 67 | |
| 68 | def _merge_overrides(base: dict | None, override: dict | None) -> dict | None: |
| 69 | """Shallow merge; per-request wins. Returns None if both are empty.""" |
| 70 | if not base and not override: |
| 71 | return None |
| 72 | return {**(base or {}), **(override or {})} |
| 73 | |
| 74 | |
| 75 | TERM_PLOTLIB_AVAILABLE = (importlib.util.find_spec("termplotlib") is not None) and ( |
| 76 | shutil.which("gnuplot") is not None |
| 77 | ) |
| 78 | |
| 79 | |
| 80 | async def _align_prompts_to_server_tokenizer( |
| 81 | base_url: str, |
| 82 | model_id: str, |
| 83 | input_requests: list[SampleRequest], |
| 84 | ssl_context: ssl.SSLContext | bool | None = None, |
| 85 | ) -> list[SampleRequest]: |
| 86 | """Re-align prompts if local/server tokenizers disagree.""" |
| 87 | if not input_requests or not isinstance(input_requests[0].prompt, str): |
| 88 | return input_requests |
| 89 | |
| 90 | tok_url = f"{base_url}/tokenize" |
| 91 | detok_url = f"{base_url}/detokenize" |
| 92 | connector = aiohttp.TCPConnector(ssl=ssl_context) |
| 93 | |
| 94 | async with aiohttp.ClientSession(connector=connector) as session: |
| 95 | sem = asyncio.Semaphore(64) |
| 96 | |
| 97 | async def _tokenize(prompt: str) -> list[int]: |
| 98 | async with ( |
| 99 | sem, |
| 100 | session.post( |
| 101 | tok_url, |
| 102 | json={ |
| 103 | "model": model_id, |
| 104 | "prompt": prompt, |
| 105 | "add_special_tokens": False, |
| 106 | }, |
| 107 | ) as r, |
| 108 | ): |
| 109 | r.raise_for_status() |
| 110 | return (await r.json())["tokens"] |
| 111 | |
| 112 | async def _detokenize(tokens: list[int]) -> str: |
| 113 | async with ( |
| 114 | sem, |
| 115 | session.post( |
| 116 | detok_url, json={"model": model_id, "tokens": tokens} |
| 117 | ) as r, |
| 118 | ): |
| 119 | r.raise_for_status() |
| 120 | return (await r.json())["prompt"] |
| 121 | |
| 122 | try: |
| 123 | first_tokens = await _tokenize(input_requests[0].prompt) |
| 124 | except Exception: |
| 125 | print("WARNING: /tokenize unavailable, skipping alignment.") |
| 126 | return input_requests |
| 127 | |
| 128 | expected = input_requests[0].prompt_len |
| 129 | if len(first_tokens) == expected: |
| 130 | return input_requests |
| 131 | |
| 132 | print( |
| 133 | f"WARNING: tokenizer mismatch " |
| 134 | f"(server={len(first_tokens)}, expected={expected}), " |
| 135 | f"re-aligning prompts." |
| 136 | ) |
| 137 | |
| 138 | async def _fix_one(req: SampleRequest) -> SampleRequest: |
| 139 | assert isinstance(req.prompt, str) |
| 140 | tokens = await _tokenize(req.prompt) |
| 141 | if len(tokens) <= req.prompt_len: |
| 142 | return req |
| 143 | corrected = await _detokenize(tokens[: req.prompt_len]) |
| 144 | return replace(req, prompt=corrected, prompt_len=req.prompt_len) |
| 145 | |
| 146 | results = await asyncio.gather( |
| 147 | *[_fix_one(r) for r in input_requests], return_exceptions=True |
| 148 | ) |
| 149 | return [ |
| 150 | res if not isinstance(res, BaseException) else orig |
| 151 | for orig, res in zip(input_requests, results) |
| 152 | ] |
| 153 | |
| 154 | |
| 155 | async def get_first_model_from_server( |
| 156 | base_url: str, |
| 157 | headers: dict | None = None, |
| 158 | ssl_context: ssl.SSLContext | bool | None = None, |
| 159 | ) -> tuple[str, str]: |
| 160 | """Fetch the first model from the server's /v1/models endpoint.""" |
| 161 | models_url = f"{base_url}/v1/models" |
| 162 | connector = aiohttp.TCPConnector(ssl=ssl_context) |
| 163 | async with aiohttp.ClientSession(connector=connector) as session: |
| 164 | try: |
| 165 | async with session.get(models_url, headers=headers) as response: |
| 166 | response.raise_for_status() |
| 167 | data = await response.json() |
| 168 | if "data" in data and len(data["data"]) > 0: |
| 169 | return data["data"][0]["id"], data["data"][0]["root"] |
| 170 | else: |
| 171 | raise ValueError( |
| 172 | f"No models found on the server at {base_url}. " |
| 173 | "Make sure the server is running and has models loaded." |
| 174 | ) |
| 175 | except (aiohttp.ClientError, json.JSONDecodeError) as e: |
| 176 | raise RuntimeError( |
| 177 | f"Failed to fetch models from server at {models_url}. " |
| 178 | "Check that:\n" |
| 179 | "1. The server is running\n" |
| 180 | "2. The server URL is correct\n" |
| 181 | f"Error: {e}" |
| 182 | ) from e |
| 183 | |
| 184 | |
| 185 | @dataclass |
| 186 | class SpecDecodeMetrics: |
| 187 | """Speculative decoding metrics from the server's Prometheus endpoint.""" |
| 188 | |
| 189 | num_drafts: int |
| 190 | num_draft_tokens: int |
| 191 | num_accepted_tokens: int |
| 192 | accepted_per_pos: dict[int, int] |
| 193 | |
| 194 | |
| 195 | async def fetch_spec_decode_metrics( |
| 196 | base_url: str, session: aiohttp.ClientSession |
| 197 | ) -> SpecDecodeMetrics | None: |
| 198 | """Fetch speculative decoding metrics from the server's Prometheus endpoint. |
| 199 | |
| 200 | Returns None if speculative decoding is not enabled or metrics are not available. |
| 201 | """ |
| 202 | metrics_url = f"{base_url}/metrics" |
| 203 | try: |
| 204 | async with session.get(metrics_url) as response: |
| 205 | if response.status != 200: |
| 206 | return None |
| 207 | text = await response.text() |
| 208 | |
| 209 | num_drafts = 0 |
| 210 | num_draft_tokens = 0 |
| 211 | num_accepted_tokens = 0 |
| 212 | accepted_per_pos: dict[int, int] = {} |
| 213 | found_spec_decode = False |
| 214 | |
| 215 | for line in text.split("\n"): |
| 216 | line = line.strip() |
| 217 | if not line or line.startswith("#"): |
| 218 | continue |
| 219 | |
| 220 | if line.startswith("vllm:spec_decode"): |
| 221 | # Extract metric name (before labels) to avoid matching |
| 222 | # substrings inside label values. |
| 223 | parts = line.split(None, 1) |
| 224 | metric_name = parts[0].split("{")[0] |
| 225 | if not metric_name.endswith("_total"): |
| 226 | continue |
| 227 | found_spec_decode = True |
| 228 | with contextlib.suppress(ValueError): |
| 229 | if "num_drafts" in metric_name: |
| 230 | num_drafts += int(float(parts[-1])) |
| 231 | elif "num_draft_tokens" in metric_name: |
| 232 | num_draft_tokens += int(float(parts[-1])) |
| 233 | elif "num_accepted_tokens_per_pos" in metric_name: |
| 234 | pos_label = 'position="' |
| 235 | if pos_label in line: |
| 236 | start = line.index(pos_label) + len(pos_label) |
| 237 | end = line.index('"', start) |
| 238 | pos = int(line[start:end]) |
| 239 | val = int(float(parts[-1])) |
| 240 | accepted_per_pos[pos] = ( |
| 241 | accepted_per_pos.get(pos, 0) + val |
| 242 | ) |
| 243 | elif "num_accepted_tokens" in metric_name: |
| 244 | num_accepted_tokens += int(float(parts[-1])) |
| 245 | |
| 246 | if not found_spec_decode: |
| 247 | return None |
| 248 | |
| 249 | return SpecDecodeMetrics( |
| 250 | num_drafts=num_drafts, |
| 251 | num_draft_tokens=num_draft_tokens, |
| 252 | num_accepted_tokens=num_accepted_tokens, |
| 253 | accepted_per_pos=accepted_per_pos, |
| 254 | ) |
| 255 | except (aiohttp.ClientError, asyncio.TimeoutError): |
| 256 | return None |
| 257 | |
| 258 | |
| 259 | @dataclass |
| 260 | class DiffusionMetrics: |
| 261 | """Diffusion (dLLM) decoding metrics from the server's Prometheus endpoint.""" |
| 262 | |
| 263 | num_denoising_steps: int |
| 264 | num_canvas_positions: int |
| 265 | num_committed_tokens: int |
| 266 | |
| 267 | |
| 268 | async def fetch_diffusion_metrics( |
| 269 | base_url: str, session: aiohttp.ClientSession |
| 270 | ) -> DiffusionMetrics | None: |
| 271 | """Fetch diffusion decoding metrics from the server's Prometheus endpoint. |
| 272 | |
| 273 | Returns None if the model is not a diffusion model or metrics are not |
| 274 | available. |
| 275 | """ |
| 276 | metrics_url = f"{base_url}/metrics" |
| 277 | try: |
| 278 | async with session.get(metrics_url) as response: |
| 279 | if response.status != 200: |
| 280 | return None |
| 281 | text = await response.text() |
| 282 | |
| 283 | num_denoising_steps = 0 |
| 284 | num_canvas_positions = 0 |
| 285 | num_committed_tokens = 0 |
| 286 | found_diffusion = False |
| 287 | |
| 288 | for line in text.split("\n"): |
| 289 | line = line.strip() |
| 290 | if not line or line.startswith("#"): |
| 291 | continue |
| 292 | |
| 293 | if line.startswith("vllm:diffusion"): |
| 294 | # Extract metric name (before labels) to avoid matching |
| 295 | # substrings inside label values. |
| 296 | parts = line.split(None, 1) |
| 297 | metric_name = parts[0].split("{")[0] |
| 298 | if not metric_name.endswith("_total"): |
| 299 | continue |
| 300 | found_diffusion = True |
| 301 | with contextlib.suppress(ValueError): |
| 302 | if "num_denoising_steps" in metric_name: |
| 303 | num_denoising_steps += int(float(parts[-1])) |
| 304 | elif "num_canvas_positions" in metric_name: |
| 305 | num_canvas_positions += int(float(parts[-1])) |
| 306 | elif "num_committed_tokens" in metric_name: |
| 307 | num_committed_tokens += int(float(parts[-1])) |
| 308 | |
| 309 | if not found_diffusion: |
| 310 | return None |
| 311 | |
| 312 | return DiffusionMetrics( |
| 313 | num_denoising_steps=num_denoising_steps, |
| 314 | num_canvas_positions=num_canvas_positions, |
| 315 | num_committed_tokens=num_committed_tokens, |
| 316 | ) |
| 317 | except (aiohttp.ClientError, asyncio.TimeoutError): |
| 318 | return None |
| 319 | |
| 320 | |
| 321 | class TaskType(Enum): |
| 322 | GENERATION = "generation" |
| 323 | POOLING = "pooling" |
| 324 | |
| 325 | |
| 326 | @dataclass |
| 327 | class BenchmarkMetrics: |
| 328 | completed: int |
| 329 | failed: int |
| 330 | total_input: int |
| 331 | total_output: int |
| 332 | request_throughput: float |
| 333 | request_goodput: float |
| 334 | output_throughput: float |
| 335 | total_token_throughput: float |
| 336 | mean_ttft_ms: float |
| 337 | median_ttft_ms: float |
| 338 | std_ttft_ms: float |
| 339 | percentiles_ttft_ms: list[tuple[float, float]] |
| 340 | mean_tpot_ms: float |
| 341 | median_tpot_ms: float |
| 342 | std_tpot_ms: float |
| 343 | percentiles_tpot_ms: list[tuple[float, float]] |
| 344 | mean_itl_ms: float |
| 345 | median_itl_ms: float |
| 346 | std_itl_ms: float |
| 347 | percentiles_itl_ms: list[tuple[float, float]] |
| 348 | # E2EL stands for end-to-end latency per request. |
| 349 | # It is the time taken on the client side from sending |
| 350 | # a request to receiving a complete response. |
| 351 | mean_e2el_ms: float |
| 352 | median_e2el_ms: float |
| 353 | std_e2el_ms: float |
| 354 | percentiles_e2el_ms: list[tuple[float, float]] |
| 355 | # Max output tokens per second and concurrent requests at that peak |
| 356 | max_output_tokens_per_s: float |
| 357 | max_concurrent_requests: int |
| 358 | rtfx: float = 0.0 # Inverse Real-Time Factor for ASR benchmarks |
| 359 | |
| 360 | |
| 361 | @dataclass |
| 362 | class EmbedBenchmarkMetrics: |
| 363 | completed: int |
| 364 | failed: int |
| 365 | total_input: int |
| 366 | total_input_sequences: int |
| 367 | request_throughput: float |
| 368 | input_sequence_throughput: float |
| 369 | total_token_throughput: float |
| 370 | mean_e2el_ms: float |
| 371 | std_e2el_ms: float |
| 372 | median_e2el_ms: float |
| 373 | percentiles_e2el_ms: list[tuple[float, float]] |
| 374 | |
| 375 | |
| 376 | def _get_current_request_rate( |
| 377 | ramp_up_strategy: Literal["linear", "exponential"] | None, |
| 378 | ramp_up_start_rps: int | None, |
| 379 | ramp_up_end_rps: int | None, |
| 380 | request_index: int, |
| 381 | total_requests: int, |
| 382 | request_rate: float, |
| 383 | ) -> float: |
| 384 | if ( |
| 385 | ramp_up_strategy |
| 386 | and ramp_up_start_rps is not None |
| 387 | and ramp_up_end_rps is not None |
| 388 | ): |
| 389 | progress = request_index / max(total_requests - 1, 1) |
| 390 | if ramp_up_strategy == "linear": |
| 391 | increase = (ramp_up_end_rps - ramp_up_start_rps) * progress |
| 392 | return ramp_up_start_rps + increase |
| 393 | elif ramp_up_strategy == "exponential": |
| 394 | ratio = ramp_up_end_rps / ramp_up_start_rps |
| 395 | return ramp_up_start_rps * (ratio**progress) |
| 396 | else: |
| 397 | raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}") |
| 398 | return request_rate |
| 399 | |
| 400 | |
| 401 | async def get_request( |
| 402 | input_requests: list[SampleRequest], |
| 403 | request_rate: float, |
| 404 | burstiness: float = 1.0, |
| 405 | ramp_up_strategy: Literal["linear", "exponential"] | None = None, |
| 406 | ramp_up_start_rps: int | None = None, |
| 407 | ramp_up_end_rps: int | None = None, |
| 408 | self_timed: bool = False, |
| 409 | ) -> AsyncGenerator[tuple[SampleRequest, float], None]: |
| 410 | """Asynchronously generates requests at a specified rate |
| 411 | with OPTIONAL burstiness and OPTIONAL ramp-up strategy. |
| 412 | |
| 413 | Args: |
| 414 | input_requests: |
| 415 | A list of input requests, each represented as a SampleRequest. |
| 416 | self_timed: |
| 417 | If True, the requests carry their own arrival timing and no |
| 418 | request rate, burstiness or ramp-up is applied. |
| 419 | request_rate: |
| 420 | The rate at which requests are generated (requests/s). |
| 421 | burstiness (optional): |
| 422 | The burstiness factor of the request generation. |
| 423 | Only takes effect when request_rate is not inf. |
| 424 | Default value is 1, which follows a Poisson process. |
| 425 | Otherwise, the request intervals follow a gamma distribution. |
| 426 | A lower burstiness value (0 < burstiness < 1) results |
| 427 | in more bursty requests, while a higher burstiness value |
| 428 | (burstiness > 1) results in a more uniform arrival of requests. |
| 429 | ramp_up_strategy (optional): |
| 430 | The ramp-up strategy. Can be "linear" or "exponential". |
| 431 | If None, uses constant request rate (specified by request_rate). |
| 432 | ramp_up_start_rps (optional): |
| 433 | The starting request rate for ramp-up. |
| 434 | ramp_up_end_rps (optional): |
| 435 | The ending request rate for ramp-up. |
| 436 | |
| 437 | """ |
| 438 | assert burstiness > 0, ( |
| 439 | f"A positive burstiness factor is expected, but given {burstiness}." |
| 440 | ) |
| 441 | # Convert to list to get length for ramp-up calculations |
| 442 | if isinstance(input_requests, Iterable) and not isinstance(input_requests, list): |
| 443 | input_requests = list(input_requests) |
| 444 | |
| 445 | total_requests = len(input_requests) |
| 446 | assert total_requests > 0, "No requests provided." |
| 447 | |
| 448 | # Precompute delays among requests to minimize request send laggings |
| 449 | request_rates: list[float] = [] |
| 450 | delay_ts: list[float] = [] |
| 451 | |
| 452 | # if the traces have timing info then: |
| 453 | if not self_timed: |
| 454 | for request_index, request in enumerate(input_requests): |
| 455 | current_request_rate = _get_current_request_rate( |
| 456 | ramp_up_strategy, |
| 457 | ramp_up_start_rps, |
| 458 | ramp_up_end_rps, |
| 459 | request_index, |
| 460 | total_requests, |
| 461 | request_rate, |
| 462 | ) |
| 463 | assert current_request_rate > 0.0, ( |
| 464 | f"Obtained non-positive request rate {current_request_rate}." |
| 465 | ) |
| 466 | request_rates.append(current_request_rate) |
| 467 | if current_request_rate == float("inf"): |
| 468 | delay_ts.append(0) |
| 469 | elif burstiness == float("inf"): |
| 470 | # when burstiness tends to infinity, the delay time becomes constant |
| 471 | # and tends to the inverse of the request rate |
| 472 | delay_ts.append(1.0 / current_request_rate) |
| 473 | else: |
| 474 | theta = 1.0 / (current_request_rate * burstiness) |
| 475 | |
| 476 | # Sample the request interval from the gamma distribution. |
| 477 | # If burstiness is 1, it follows exponential distribution. |
| 478 | delay_ts.append(np.random.gamma(shape=burstiness, scale=theta)) |
| 479 | |
| 480 | # Calculate the cumulative delay time from the first sent out requests. |
| 481 | for i in range(1, len(delay_ts)): |
| 482 | delay_ts[i] += delay_ts[i - 1] |
| 483 | if ramp_up_strategy is None and delay_ts[-1] != 0: |
| 484 | # When ramp_up_strategy is not set, we assume the request rate is fixed |
| 485 | # and all requests should be sent in target_total_delay_s, the following |
| 486 | # logic would re-scale delay time to ensure the final delay_ts |
| 487 | # align with target_total_delay_s. |
| 488 | # |
| 489 | # NOTE: If we simply accumulate the random delta values |
| 490 | # from the gamma distribution, their sum would have 1-2% gap |
| 491 | # from target_total_delay_s. The purpose of the following logic is to |
| 492 | # close the gap for stabilizing the throughput data |
| 493 | # from different random seeds. |
| 494 | target_total_delay_s = total_requests / request_rate |
| 495 | normalize_factor = target_total_delay_s / delay_ts[-1] |
| 496 | delay_ts = [delay * normalize_factor for delay in delay_ts] |
| 497 | else: |
| 498 | for request_index, request in enumerate(input_requests): |
| 499 | # this is cumulative running ts, from which sleep is calculated later |
| 500 | if request.timestamp is not None: |
| 501 | delay_ts.append(request.timestamp) |
| 502 | else: |
| 503 | delay_ts.append(0.0) |
| 504 | # TODO: there is no notion of RPS here, may be we can calculate |
| 505 | # from the trace. |
| 506 | request_rates.append(0.0) |
| 507 | |
| 508 | start_ts = time.time() |
| 509 | for request_index, request in enumerate(input_requests): |
| 510 | if delay_ts[request_index] > 0: |
| 511 | current_ts = time.time() |
| 512 | sleep_interval_s = start_ts + delay_ts[request_index] - current_ts |
| 513 | if sleep_interval_s > 0: |
| 514 | await asyncio.sleep(sleep_interval_s) |
| 515 | yield request, request_rates[request_index] |
| 516 | |
| 517 | |
| 518 | def calculate_metrics_for_embeddings( |
| 519 | outputs: list[RequestFuncOutput], |
| 520 | dur_s: float, |
| 521 | selected_percentiles: list[float], |
| 522 | ) -> EmbedBenchmarkMetrics: |
| 523 | """Calculate the metrics for the embedding requests. |
| 524 | |
| 525 | Args: |
| 526 | outputs: The outputs of the requests. |
| 527 | dur_s: The duration of the benchmark. |
| 528 | selected_percentiles: The percentiles to select. |
| 529 | |
| 530 | Returns: |
| 531 | The calculated benchmark metrics. |
| 532 | |
| 533 | """ |
| 534 | total_input = 0 |
| 535 | total_input_sequences = 0 |
| 536 | completed = 0 |
| 537 | failed = 0 |
| 538 | e2els: list[float] = [] |
| 539 | for i in range(len(outputs)): |
| 540 | if outputs[i].success: |
| 541 | e2els.append(outputs[i].latency) |
| 542 | completed += 1 |
| 543 | total_input += outputs[i].prompt_len |
| 544 | total_input_sequences += outputs[i].num_input_sequences |
| 545 | else: |
| 546 | failed += 1 |
| 547 | |
| 548 | if completed == 0: |
| 549 | warnings.warn( |
| 550 | "All requests failed. This is likely due to a misconfiguration " |
| 551 | "on the benchmark arguments.", |
| 552 | stacklevel=2, |
| 553 | ) |
| 554 | metrics = EmbedBenchmarkMetrics( |
| 555 | completed=completed, |
| 556 | failed=failed, |
| 557 | total_input=total_input, |
| 558 | total_input_sequences=total_input_sequences, |
| 559 | request_throughput=completed / dur_s, |
| 560 | input_sequence_throughput=total_input_sequences / dur_s, |
| 561 | total_token_throughput=total_input / dur_s, |
| 562 | mean_e2el_ms=np.mean(e2els or 0) * 1000, |
| 563 | std_e2el_ms=np.std(e2els or 0) * 1000, |
| 564 | median_e2el_ms=np.median(e2els or 0) * 1000, |
| 565 | percentiles_e2el_ms=[ |
| 566 | (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles |
| 567 | ], |
| 568 | ) |
| 569 | return metrics |
| 570 | |
| 571 | |
| 572 | def calculate_metrics( |
| 573 | input_requests: list[SampleRequest], |
| 574 | outputs: list[RequestFuncOutput], |
| 575 | dur_s: float, |
| 576 | tokenizer: TokenizerLike | None, |
| 577 | selected_percentiles: list[float], |
| 578 | goodput_config_dict: dict[str, float], |
| 579 | ) -> tuple[BenchmarkMetrics, list[int]]: |
| 580 | """Calculate the metrics for the benchmark. |
| 581 | |
| 582 | Args: |
| 583 | input_requests: The input requests. |
| 584 | outputs: The outputs of the requests. |
| 585 | dur_s: The duration of the benchmark. |
| 586 | tokenizer: The tokenizer to use. |
| 587 | selected_percentiles: The percentiles to select. |
| 588 | goodput_config_dict: The goodput configuration. |
| 589 | |
| 590 | Returns: |
| 591 | A tuple of the benchmark metrics and the actual output lengths. |
| 592 | |
| 593 | """ |
| 594 | actual_output_lens: list[int] = [] |
| 595 | total_input = 0 |
| 596 | completed = 0 |
| 597 | good_completed = 0 |
| 598 | itls: list[float] = [] |
| 599 | tpots: list[float] = [] |
| 600 | all_tpots: list[float] = [] |
| 601 | ttfts: list[float] = [] |
| 602 | e2els: list[float] = [] |
| 603 | input_audio_duration = 0.0 |
| 604 | for i in range(len(outputs)): |
| 605 | if outputs[i].success: |
| 606 | output_len = outputs[i].output_tokens |
| 607 | |
| 608 | if not output_len: |
| 609 | if tokenizer is None: |
| 610 | output_len = 1 |
| 611 | else: |
| 612 | # We use the tokenizer to count the number of output tokens |
| 613 | # for some serving backends instead of looking at |
| 614 | # len(outputs[i].itl) since multiple output tokens may be |
| 615 | # bundled together |
| 616 | # Note : this may inflate the output token count slightly |
| 617 | output_len = len( |
| 618 | tokenizer( |
| 619 | outputs[i].generated_text, add_special_tokens=False |
| 620 | ).input_ids |
| 621 | ) |
| 622 | actual_output_lens.append(output_len) |
| 623 | total_input += outputs[i].prompt_len |
| 624 | tpot = 0.0 |
| 625 | if output_len > 1: |
| 626 | latency_minus_ttft = outputs[i].latency - outputs[i].ttft |
| 627 | tpot = latency_minus_ttft / (output_len - 1) |
| 628 | tpots.append(tpot) |
| 629 | # Note: if output_len <= 1, we regard tpot as 0 for goodput |
| 630 | all_tpots.append(tpot) |
| 631 | itls += outputs[i].itl |
| 632 | ttfts.append(outputs[i].ttft) |
| 633 | e2els.append(outputs[i].latency) |
| 634 | input_audio_duration += outputs[i].input_audio_duration |
| 635 | completed += 1 |
| 636 | else: |
| 637 | actual_output_lens.append(0) |
| 638 | |
| 639 | if goodput_config_dict: |
| 640 | valid_metrics = [] |
| 641 | slo_values = [] |
| 642 | |
| 643 | if "ttft" in goodput_config_dict: |
| 644 | valid_metrics.append(ttfts) |
| 645 | slo_values.append( |
| 646 | goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION |
| 647 | ) |
| 648 | if "tpot" in goodput_config_dict: |
| 649 | valid_metrics.append(all_tpots) |
| 650 | slo_values.append( |
| 651 | goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION |
| 652 | ) |
| 653 | if "e2el" in goodput_config_dict: |
| 654 | valid_metrics.append(e2els) |
| 655 | slo_values.append( |
| 656 | goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION |
| 657 | ) |
| 658 | |
| 659 | for req_metric in zip(*valid_metrics): |
| 660 | is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)]) |
| 661 | if is_good_req: |
| 662 | good_completed += 1 |
| 663 | |
| 664 | if completed == 0: |
| 665 | warnings.warn( |
| 666 | "All requests failed. This is likely due to a misconfiguration " |
| 667 | "on the benchmark arguments.", |
| 668 | stacklevel=2, |
| 669 | ) |
| 670 | |
| 671 | # Calculate max output tokens per second metric |
| 672 | max_output_tokens_per_s = 0.0 |
| 673 | max_concurrent_requests = 0 |
| 674 | |
| 675 | # Find the time range across all successful requests |
| 676 | successful_outputs = [output for output in outputs if output.success] |
| 677 | failed_outputs = [output for output in outputs if not output.success] |
| 678 | |
| 679 | if len(failed_outputs) > 0: |
| 680 | print("Failed requests during benchmark run detected (capping to 10):") |
| 681 | for i, err in enumerate(failed_outputs[:10]): |
| 682 | print(f"Error {i}: {err.error}") |
| 683 | |
| 684 | if successful_outputs: |
| 685 | min_start_time = min(output.start_time for output in successful_outputs) |
| 686 | max_end_time = max( |
| 687 | output.start_time + output.latency for output in successful_outputs |
| 688 | ) |
| 689 | |
| 690 | # Create second buckets (ceiling to ensure we capture all time) |
| 691 | duration_seconds = int(np.ceil(max_end_time - min_start_time)) + 1 |
| 692 | tokens_per_second = np.zeros(duration_seconds) |
| 693 | concurrent_requests_per_second = np.zeros(duration_seconds) |
| 694 | |
| 695 | for i, output in enumerate(successful_outputs): |
| 696 | # Calculate token generation timestamp using |
| 697 | # start_time, ttft, and itl |
| 698 | token_times = [output.start_time + output.ttft] |
| 699 | current_time = token_times[0] |
| 700 | for itl_value in output.itl: |
| 701 | current_time += itl_value |
| 702 | token_times.append(current_time) |
| 703 | |
| 704 | # Add tokens to second buckets |
| 705 | for token_time in token_times: |
| 706 | second_bucket = int(token_time - min_start_time) |
| 707 | if 0 <= second_bucket < duration_seconds: |
| 708 | tokens_per_second[second_bucket] += 1 |
| 709 | |
| 710 | # Track concurrent requests for each second this request was active |
| 711 | request_start_second = int(output.start_time - min_start_time) |
| 712 | request_end_second = int( |
| 713 | (output.start_time + output.latency) - min_start_time |
| 714 | ) |
| 715 | |
| 716 | for second in range(request_start_second, request_end_second + 1): |
| 717 | concurrent_requests_per_second[second] += 1 |
| 718 | |
| 719 | # Find the maximum tokens per second and corresponding |
| 720 | # concurrent requests |
| 721 | if len(tokens_per_second) > 0: |
| 722 | max_output_tokens_per_s = float(np.max(tokens_per_second)) |
| 723 | max_concurrent_requests = int(np.max(concurrent_requests_per_second)) |
| 724 | |
| 725 | if TERM_PLOTLIB_AVAILABLE: |
| 726 | import termplotlib as tpl |
| 727 | |
| 728 | fig = tpl.figure() |
| 729 | fig.plot( |
| 730 | np.arange(len(tokens_per_second)), |
| 731 | tokens_per_second, |
| 732 | title="Output tokens per second", |
| 733 | ) |
| 734 | fig.plot( |
| 735 | np.arange(len(concurrent_requests_per_second)), |
| 736 | concurrent_requests_per_second, |
| 737 | title="Concurrent requests per second", |
| 738 | ) |
| 739 | fig.show() |
| 740 | else: |
| 741 | print("tip: install termplotlib and gnuplot to plot the metrics") |
| 742 | |
| 743 | metrics = BenchmarkMetrics( |
| 744 | completed=completed, |
| 745 | failed=len(failed_outputs), |
| 746 | total_input=total_input, |
| 747 | total_output=sum(actual_output_lens), |
| 748 | request_throughput=completed / dur_s, |
| 749 | request_goodput=good_completed / dur_s, |
| 750 | output_throughput=sum(actual_output_lens) / dur_s, |
| 751 | total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s, |
| 752 | mean_ttft_ms=np.mean(ttfts or 0) |
| 753 | * 1000, # ttfts is empty if streaming is not supported by the endpoint |
| 754 | std_ttft_ms=np.std(ttfts or 0) * 1000, |
| 755 | median_ttft_ms=np.median(ttfts or 0) * 1000, |
| 756 | percentiles_ttft_ms=[ |
| 757 | (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles |
| 758 | ], |
| 759 | mean_tpot_ms=np.mean(tpots or 0) * 1000, |
| 760 | std_tpot_ms=np.std(tpots or 0) * 1000, |
| 761 | median_tpot_ms=np.median(tpots or 0) * 1000, |
| 762 | percentiles_tpot_ms=[ |
| 763 | (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles |
| 764 | ], |
| 765 | mean_itl_ms=np.mean(itls or 0) * 1000, |
| 766 | std_itl_ms=np.std(itls or 0) * 1000, |
| 767 | median_itl_ms=np.median(itls or 0) * 1000, |
| 768 | percentiles_itl_ms=[ |
| 769 | (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles |
| 770 | ], |
| 771 | mean_e2el_ms=np.mean(e2els or 0) * 1000, |
| 772 | std_e2el_ms=np.std(e2els or 0) * 1000, |
| 773 | median_e2el_ms=np.median(e2els or 0) * 1000, |
| 774 | percentiles_e2el_ms=[ |
| 775 | (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles |
| 776 | ], |
| 777 | max_output_tokens_per_s=max_output_tokens_per_s, |
| 778 | max_concurrent_requests=max_concurrent_requests, |
| 779 | rtfx=input_audio_duration / dur_s, |
| 780 | ) |
| 781 | |
| 782 | return metrics, actual_output_lens |
| 783 | |
| 784 | |
| 785 | async def benchmark( |
| 786 | task_type: TaskType, |
| 787 | endpoint_type: str, |
| 788 | api_url: str, |
| 789 | base_url: str, |
| 790 | model_id: str, |
| 791 | model_name: str, |
| 792 | tokenizer: TokenizerLike | None, |
| 793 | input_requests: list[SampleRequest], |
| 794 | logprobs: int | None, |
| 795 | request_rate: float, |
| 796 | burstiness: float, |
| 797 | disable_tqdm: bool, |
| 798 | num_warmups: int, |
| 799 | profile: bool, |
| 800 | selected_percentile_metrics: list[str], |
| 801 | selected_percentiles: list[float], |
| 802 | ignore_eos: bool, |
| 803 | goodput_config_dict: dict[str, float], |
| 804 | max_concurrency: int | None, |
| 805 | lora_modules: Iterable[str] | None, |
| 806 | extra_headers: dict | None, |
| 807 | extra_body: dict | None, |
| 808 | lora_assignment: Literal["random", "round-robin"] = "random", |
| 809 | ramp_up_strategy: Literal["linear", "exponential"] | None = None, |
| 810 | ramp_up_start_rps: int | None = None, |
| 811 | ramp_up_end_rps: int | None = None, |
| 812 | ready_check_timeout_sec: int = 600, |
| 813 | ssl_context: ssl.SSLContext | bool | None = None, |
| 814 | self_timed: bool = False, |
| 815 | probe_request_rate: float = 0.0, |
| 816 | ): |
| 817 | try: |
| 818 | request_func = ASYNC_REQUEST_FUNCS[endpoint_type] |
| 819 | except KeyError: |
| 820 | raise ValueError(f"Unknown backend: {endpoint_type}") from None |
| 821 | |
| 822 | # Reuses connections across requests to reduce TLS handshake overhead. |
| 823 | # Use ssl_context if provided, otherwise default to True for https URLs |
| 824 | ssl_setting = ssl_context if ssl_context is not None else ("https://" in api_url) |
| 825 | connector = aiohttp.TCPConnector( |
| 826 | limit=max_concurrency or 0, |
| 827 | limit_per_host=max_concurrency or 0, |
| 828 | ttl_dns_cache=300, |
| 829 | use_dns_cache=True, |
| 830 | keepalive_timeout=60, |
| 831 | enable_cleanup_closed=True, |
| 832 | force_close=False, |
| 833 | ssl=ssl_setting, |
| 834 | ) |
| 835 | |
| 836 | session = aiohttp.ClientSession( |
| 837 | connector=connector, |
| 838 | trust_env=True, |
| 839 | timeout=aiohttp.ClientTimeout(total=6 * 60 * 60), |
| 840 | ) |
| 841 | |
| 842 | print("Starting initial single prompt test run...") |
| 843 | test_prompt, test_prompt_len, test_output_len, test_mm_content = ( |
| 844 | input_requests[0].prompt, |
| 845 | input_requests[0].prompt_len, |
| 846 | input_requests[0].expected_output_len, |
| 847 | input_requests[0].multi_modal_data, |
| 848 | ) |
| 849 | test_extra_body = _merge_overrides(extra_body, input_requests[0].request_overrides) |
| 850 | test_chat_messages = input_requests[0].chat_messages |
| 851 | |
| 852 | assert ( |
| 853 | test_mm_content is None |
| 854 | or isinstance(test_mm_content, dict) |
| 855 | or ( |
| 856 | isinstance(test_mm_content, list) |
| 857 | and all(isinstance(item, dict) for item in test_mm_content) |
| 858 | ) |
| 859 | ), "multi_modal_data must be a dict or list[dict]" |
| 860 | test_input = RequestFuncInput( |
| 861 | model=model_id, |
| 862 | model_name=model_name, |
| 863 | prompt=test_prompt, |
| 864 | api_url=api_url, |
| 865 | prompt_len=test_prompt_len, |
| 866 | output_len=test_output_len, |
| 867 | logprobs=logprobs, |
| 868 | multi_modal_content=test_mm_content, |
| 869 | ignore_eos=ignore_eos, |
| 870 | extra_headers=extra_headers, |
| 871 | extra_body=test_extra_body, |
| 872 | chat_messages=test_chat_messages, |
| 873 | ) |
| 874 | |
| 875 | if ready_check_timeout_sec > 0: |
| 876 | test_output = await wait_for_endpoint( |
| 877 | request_func, |
| 878 | test_input, |
| 879 | session, |
| 880 | timeout_seconds=ready_check_timeout_sec, |
| 881 | ) |
| 882 | if not test_output.success: |
| 883 | raise ValueError( |
| 884 | "Initial test run failed - Please make sure benchmark " |
| 885 | "arguments are correctly specified. " |
| 886 | f"Error: {test_output.error}" |
| 887 | ) |
| 888 | else: |
| 889 | print("Initial test run completed.") |
| 890 | else: |
| 891 | print("Skipping endpoint ready check.") |
| 892 | |
| 893 | if num_warmups > 0: |
| 894 | print(f"Warming up with {num_warmups} requests...") |
| 895 | warmup_pbar = None if disable_tqdm else tqdm(total=num_warmups) |
| 896 | warmup_semaphore = ( |
| 897 | asyncio.Semaphore(max_concurrency) |
| 898 | if max_concurrency |
| 899 | else contextlib.nullcontext() |
| 900 | ) |
| 901 | warmup_tasks = [] |
| 902 | |
| 903 | async def warmup_limited_request_func(): |
| 904 | async with warmup_semaphore: |
| 905 | return await request_func( |
| 906 | request_func_input=test_input, session=session, pbar=warmup_pbar |
| 907 | ) |
| 908 | |
| 909 | for _ in range(num_warmups): |
| 910 | request_task = asyncio.create_task(warmup_limited_request_func()) |
| 911 | warmup_tasks.append(request_task) |
| 912 | _ = await asyncio.gather(*warmup_tasks) |
| 913 | |
| 914 | if warmup_pbar is not None: |
| 915 | warmup_pbar.close() |
| 916 | print("Warmup run completed.") |
| 917 | |
| 918 | print("Starting main benchmark run...") |
| 919 | |
| 920 | lora_modules_iter: Iterator[str] | None = None |
| 921 | if lora_modules: |
| 922 | lora_modules_list = list(lora_modules) |
| 923 | if lora_assignment == "round-robin": |
| 924 | # Deterministic round-robin assignment across requests. |
| 925 | lora_modules_iter = iter( |
| 926 | [ |
| 927 | lora_modules_list[i % len(lora_modules_list)] |
| 928 | for i in range(len(input_requests)) |
| 929 | ] |
| 930 | ) |
| 931 | else: |
| 932 | # For each input request, choose a LoRA module at random. |
| 933 | lora_modules_iter = iter( |
| 934 | [random.choice(lora_modules_list) for _ in range(len(input_requests))] |
| 935 | ) |
| 936 | |
| 937 | if profile: |
| 938 | print("Starting profiler...") |
| 939 | profile_input = RequestFuncInput( |
| 940 | model=model_id, |
| 941 | model_name=model_name, |
| 942 | prompt=test_prompt, |
| 943 | api_url=base_url + "/start_profile", |
| 944 | prompt_len=test_prompt_len, |
| 945 | output_len=test_output_len, |
| 946 | logprobs=logprobs, |
| 947 | multi_modal_content=test_mm_content, |
| 948 | ignore_eos=ignore_eos, |
| 949 | extra_headers=extra_headers, |
| 950 | extra_body=test_extra_body, |
| 951 | chat_messages=test_chat_messages, |
| 952 | ) |
| 953 | profile_output = await request_func( |
| 954 | request_func_input=profile_input, session=session |
| 955 | ) |
| 956 | if profile_output.success: |
| 957 | print("Profiler started") |
| 958 | |
| 959 | distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution" |
| 960 | if not self_timed: |
| 961 | if ramp_up_strategy is not None: |
| 962 | print(f"Traffic ramp-up strategy: {ramp_up_strategy}.") |
| 963 | print( |
| 964 | f"Will increase RPS from {ramp_up_start_rps} to " |
| 965 | f"{ramp_up_end_rps} RPS over the duration of the benchmark." |
| 966 | ) |
| 967 | else: |
| 968 | print(f"Traffic request rate: {request_rate}") |
| 969 | |
| 970 | print(f"Burstiness factor: {burstiness} ({distribution})") |
| 971 | print(f"Maximum request concurrency: {max_concurrency}") |
| 972 | else: |
| 973 | print("Self timing is set, using the timestamps from the trace file.") |
| 974 | |
| 975 | spec_decode_metrics_before = await fetch_spec_decode_metrics(base_url, session) |
| 976 | diffusion_metrics_before = await fetch_diffusion_metrics(base_url, session) |
| 977 | |
| 978 | pbar = None if disable_tqdm else tqdm(total=len(input_requests)) |
| 979 | |
| 980 | semaphore = ( |
| 981 | asyncio.Semaphore(max_concurrency) |
| 982 | if max_concurrency |
| 983 | else contextlib.nullcontext() |
| 984 | ) |
| 985 | |
| 986 | async def limited_request_func( |
| 987 | request_func_input, session, pbar, request_arrival_time |
| 988 | ): |
| 989 | async with semaphore: |
| 990 | output = await request_func( |
| 991 | request_func_input=request_func_input, session=session, pbar=pbar |
| 992 | ) |
| 993 | # Preserve start_time as the time the request was sent. It is used by |
| 994 | # throughput calculations; record client-side semaphore delay separately. |
| 995 | output.client_queue_time = output.start_time - request_arrival_time |
| 996 | return output |
| 997 | |
| 998 | probe_outputs: list[RequestFuncOutput] = [] |
| 999 | probe_stop = asyncio.Event() |
| 1000 | |
| 1001 | async def probe_loop(): |
| 1002 | probe_input = replace( |
| 1003 | test_input, |
| 1004 | prompt="Hi", |
| 1005 | prompt_len=1, |
| 1006 | output_len=1, |
| 1007 | multi_modal_content=None, |
| 1008 | chat_messages=None, |
| 1009 | ) |
| 1010 | interval = 1 / probe_request_rate |
| 1011 | while not probe_stop.is_set(): |
| 1012 | probe_outputs.append( |
| 1013 | await request_func(request_func_input=probe_input, session=session) |
| 1014 | ) |
| 1015 | await asyncio.sleep(interval) |
| 1016 | |
| 1017 | probe_task: asyncio.Task | None = None |
| 1018 | if probe_request_rate > 0: |
| 1019 | print(f"Probe request rate: {probe_request_rate} req/s") |
| 1020 | probe_task = asyncio.create_task(probe_loop()) |
| 1021 | |
| 1022 | benchmark_start_time = time.perf_counter() |
| 1023 | tasks: list[asyncio.Task] = [] |
| 1024 | |
| 1025 | rps_change_events = [] |
| 1026 | last_int_rps = -1 |
| 1027 | if ramp_up_strategy is not None and ramp_up_start_rps is not None: |
| 1028 | last_int_rps = ramp_up_start_rps |
| 1029 | rps_change_events.append( |
| 1030 | { |
| 1031 | "rps": last_int_rps, |
| 1032 | "timestamp": datetime.now().isoformat(), |
| 1033 | } |
| 1034 | ) |
| 1035 | |
| 1036 | async for request, current_request_rate in get_request( |
| 1037 | input_requests, |
| 1038 | request_rate, |
| 1039 | burstiness, |
| 1040 | ramp_up_strategy, |
| 1041 | ramp_up_start_rps, |
| 1042 | ramp_up_end_rps, |
| 1043 | self_timed, |
| 1044 | ): |
| 1045 | if ramp_up_strategy is not None: |
| 1046 | current_int_rps = int(current_request_rate) |
| 1047 | if current_int_rps > last_int_rps: |
| 1048 | timestamp = datetime.now().isoformat() |
| 1049 | for rps_val in range(last_int_rps + 1, current_int_rps + 1): |
| 1050 | rps_change_events.append({"rps": rps_val, "timestamp": timestamp}) |
| 1051 | last_int_rps = current_int_rps |
| 1052 | prompt, prompt_len, output_len, mm_content, request_id = ( |
| 1053 | request.prompt, |
| 1054 | request.prompt_len, |
| 1055 | request.expected_output_len, |
| 1056 | request.multi_modal_data, |
| 1057 | request.request_id, |
| 1058 | ) |
| 1059 | per_request_extra_body = _merge_overrides(extra_body, request.request_overrides) |
| 1060 | req_model_id, req_model_name = model_id, model_name |
| 1061 | if lora_modules_iter: |
| 1062 | req_lora_module = next(lora_modules_iter) |
| 1063 | req_model_id, req_model_name = req_lora_module, req_lora_module |
| 1064 | |
| 1065 | mm_content_typed: dict[str, Any] | list[dict[str, Any]] | None = None |
| 1066 | if isinstance(mm_content, (dict, list)): |
| 1067 | mm_content_typed = mm_content |
| 1068 | |
| 1069 | request_func_input = RequestFuncInput( |
| 1070 | model=req_model_id, |
| 1071 | model_name=req_model_name, |
| 1072 | prompt=prompt, |
| 1073 | api_url=api_url, |
| 1074 | prompt_len=prompt_len, |
| 1075 | output_len=output_len or 0, |
| 1076 | logprobs=logprobs, |
| 1077 | multi_modal_content=mm_content_typed, |
| 1078 | ignore_eos=ignore_eos, |
| 1079 | extra_headers=extra_headers, |
| 1080 | extra_body=per_request_extra_body, |
| 1081 | request_id=request_id, |
| 1082 | chat_messages=request.chat_messages, |
| 1083 | ) |
| 1084 | request_arrival_time = time.perf_counter() |
| 1085 | tasks.append( |
| 1086 | asyncio.create_task( |
| 1087 | limited_request_func( |
| 1088 | request_func_input, session, pbar, request_arrival_time |
| 1089 | ) |
| 1090 | ) |
| 1091 | ) |
| 1092 | outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks) |
| 1093 | |
| 1094 | if probe_task is not None: |
| 1095 | probe_stop.set() |
| 1096 | await probe_task |
| 1097 | |
| 1098 | if pbar is not None: |
| 1099 | pbar.close() |
| 1100 | |
| 1101 | benchmark_duration = time.perf_counter() - benchmark_start_time |
| 1102 | |
| 1103 | spec_decode_metrics_after = await fetch_spec_decode_metrics(base_url, session) |
| 1104 | spec_decode_stats: dict[str, Any] | None = None |
| 1105 | if spec_decode_metrics_before is not None and spec_decode_metrics_after is not None: |
| 1106 | delta_drafts = ( |
| 1107 | spec_decode_metrics_after.num_drafts - spec_decode_metrics_before.num_drafts |
| 1108 | ) |
| 1109 | delta_draft_tokens = ( |
| 1110 | spec_decode_metrics_after.num_draft_tokens |
| 1111 | - spec_decode_metrics_before.num_draft_tokens |
| 1112 | ) |
| 1113 | delta_accepted = ( |
| 1114 | spec_decode_metrics_after.num_accepted_tokens |
| 1115 | - spec_decode_metrics_before.num_accepted_tokens |
| 1116 | ) |
| 1117 | per_pos_rates: list[float] = [] |
| 1118 | if delta_drafts > 0: |
| 1119 | positions = sorted( |
| 1120 | set(spec_decode_metrics_before.accepted_per_pos.keys()) |
| 1121 | | set(spec_decode_metrics_after.accepted_per_pos.keys()) |
| 1122 | ) |
| 1123 | for pos in positions: |
| 1124 | before_val = spec_decode_metrics_before.accepted_per_pos.get(pos, 0) |
| 1125 | after_val = spec_decode_metrics_after.accepted_per_pos.get( |
| 1126 | pos, before_val |
| 1127 | ) |
| 1128 | delta_pos = after_val - before_val |
| 1129 | per_pos_rates.append(delta_pos / delta_drafts) |
| 1130 | |
| 1131 | if delta_draft_tokens > 0: |
| 1132 | acceptance_rate = (delta_accepted / delta_draft_tokens) * 100 |
| 1133 | acceptance_length = ( |
| 1134 | 1 + delta_accepted / delta_drafts if delta_drafts > 0 else 0.0 |
| 1135 | ) |
| 1136 | spec_decode_stats = { |
| 1137 | "num_drafts": delta_drafts, |
| 1138 | "draft_tokens": delta_draft_tokens, |
| 1139 | "accepted_tokens": delta_accepted, |
| 1140 | "acceptance_rate": acceptance_rate, |
| 1141 | "acceptance_length": acceptance_length, |
| 1142 | "per_position_acceptance_rates": per_pos_rates, |
| 1143 | } |
| 1144 | |
| 1145 | diffusion_metrics_after = await fetch_diffusion_metrics(base_url, session) |
| 1146 | diffusion_stats: dict[str, Any] | None = None |
| 1147 | if diffusion_metrics_before is not None and diffusion_metrics_after is not None: |
| 1148 | delta_steps = ( |
| 1149 | diffusion_metrics_after.num_denoising_steps |
| 1150 | - diffusion_metrics_before.num_denoising_steps |
| 1151 | ) |
| 1152 | delta_positions = ( |
| 1153 | diffusion_metrics_after.num_canvas_positions |
| 1154 | - diffusion_metrics_before.num_canvas_positions |
| 1155 | ) |
| 1156 | delta_committed = ( |
| 1157 | diffusion_metrics_after.num_committed_tokens |
| 1158 | - diffusion_metrics_before.num_committed_tokens |
| 1159 | ) |
| 1160 | if delta_steps > 0 and delta_committed > 0: |
| 1161 | block_size = delta_positions / delta_steps # canvas length (CL) |
| 1162 | num_canvases = delta_committed / block_size # = number of commit steps |
| 1163 | denoising_steps = delta_steps - num_canvases # exclude commit steps |
| 1164 | diffusion_stats = { |
| 1165 | "denoising_steps": denoising_steps, |
| 1166 | "canvas_positions": delta_positions, |
| 1167 | "committed_tokens": delta_committed, |
| 1168 | "committed_throughput": delta_committed / benchmark_duration, |
| 1169 | "steps_per_canvas": denoising_steps / num_canvases, |
| 1170 | "committed_per_step": delta_committed / denoising_steps, |
| 1171 | } |
| 1172 | |
| 1173 | metrics: BenchmarkMetrics | EmbedBenchmarkMetrics |
| 1174 | actual_output_lens: list[int] | int |
| 1175 | if task_type == TaskType.GENERATION: |
| 1176 | metrics, actual_output_lens = calculate_metrics( |
| 1177 | input_requests=input_requests, |
| 1178 | outputs=outputs, |
| 1179 | dur_s=benchmark_duration, |
| 1180 | tokenizer=tokenizer, |
| 1181 | selected_percentiles=selected_percentiles, |
| 1182 | goodput_config_dict=goodput_config_dict, |
| 1183 | ) |
| 1184 | else: |
| 1185 | metrics = calculate_metrics_for_embeddings( |
| 1186 | outputs=outputs, |
| 1187 | dur_s=benchmark_duration, |
| 1188 | selected_percentiles=selected_percentiles, |
| 1189 | ) |
| 1190 | actual_output_lens = 0 |
| 1191 | |
| 1192 | print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="=")) |
| 1193 | print("{:<40} {:<10}".format("Successful requests:", metrics.completed)) |
| 1194 | print("{:<40} {:<10}".format("Failed requests:", metrics.failed)) |
| 1195 | if max_concurrency is not None: |
| 1196 | print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency)) |
| 1197 | if request_rate != float("inf"): |
| 1198 | print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate)) |
| 1199 | print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration)) |
| 1200 | print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input)) |
| 1201 | if isinstance(metrics, BenchmarkMetrics) and tokenizer: |
| 1202 | print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output)) |
| 1203 | print( |
| 1204 | "{:<40} {:<10.2f}".format( |
| 1205 | "Request throughput (req/s):", metrics.request_throughput |
| 1206 | ) |
| 1207 | ) |
| 1208 | if isinstance(metrics, EmbedBenchmarkMetrics): |
| 1209 | print( |
| 1210 | "{:<40} {:<10.2f}".format( |
| 1211 | "Input throughput (inputs/s):", metrics.input_sequence_throughput |
| 1212 | ) |
| 1213 | ) |
| 1214 | if goodput_config_dict and isinstance(metrics, BenchmarkMetrics): |
| 1215 | print( |
| 1216 | "{:<40} {:<10.2f}".format( |
| 1217 | "Request goodput (req/s):", metrics.request_goodput |
| 1218 | ) |
| 1219 | ) |
| 1220 | if isinstance(metrics, BenchmarkMetrics): |
| 1221 | if tokenizer: |
| 1222 | print( |
| 1223 | "{:<40} {:<10.2f}".format( |
| 1224 | "Output token throughput (tok/s):", metrics.output_throughput |
| 1225 | ) |
| 1226 | ) |
| 1227 | print( |
| 1228 | "{:<40} {:<10.2f}".format( |
| 1229 | "Peak output token throughput (tok/s):", |
| 1230 | metrics.max_output_tokens_per_s, |
| 1231 | ) |
| 1232 | ) |
| 1233 | print( |
| 1234 | "{:<40} {:<10.2f}".format( |
| 1235 | "Peak concurrent requests:", metrics.max_concurrent_requests |
| 1236 | ) |
| 1237 | ) |
| 1238 | if metrics.rtfx > 0.0: |
| 1239 | print( |
| 1240 | "{:<40} {:<10.2f}".format( |
| 1241 | "RTFx (Inverse Real-Time Factor):", metrics.rtfx |
| 1242 | ) |
| 1243 | ) |
| 1244 | if tokenizer: |
| 1245 | print( |
| 1246 | "{:<40} {:<10.2f}".format( |
| 1247 | "Total token throughput (tok/s):", metrics.total_token_throughput |
| 1248 | ) |
| 1249 | ) |
| 1250 | |
| 1251 | probe_stats: dict[str, Any] | None = None |
| 1252 | if probe_task is not None: |
| 1253 | probe_lats = [o.latency for o in probe_outputs if o.success] |
| 1254 | if probe_lats: |
| 1255 | probe_stats = { |
| 1256 | "probe_completed": len(probe_lats), |
| 1257 | "probe_failed": len(probe_outputs) - len(probe_lats), |
| 1258 | "probe_median_e2el_ms": float(np.median(probe_lats)) * 1000, |
| 1259 | "probe_p99_e2el_ms": float(np.percentile(probe_lats, 99)) * 1000, |
| 1260 | "probe_max_e2el_ms": float(max(probe_lats)) * 1000, |
| 1261 | } |
| 1262 | print("{s:{c}^{n}}".format(s="Probe Requests", n=50, c="-")) |
| 1263 | print( |
| 1264 | "{:<40} {:<10}".format( |
| 1265 | "Probe requests completed:", probe_stats["probe_completed"] |
| 1266 | ) |
| 1267 | ) |
| 1268 | print( |
| 1269 | "{:<40} {:<10}".format( |
| 1270 | "Probe requests failed:", probe_stats["probe_failed"] |
| 1271 | ) |
| 1272 | ) |
| 1273 | print( |
| 1274 | "{:<40} {:<10.2f}".format( |
| 1275 | "Median probe E2EL (ms):", probe_stats["probe_median_e2el_ms"] |
| 1276 | ) |
| 1277 | ) |
| 1278 | print( |
| 1279 | "{:<40} {:<10.2f}".format( |
| 1280 | "P99 probe E2EL (ms):", probe_stats["probe_p99_e2el_ms"] |
| 1281 | ) |
| 1282 | ) |
| 1283 | print( |
| 1284 | "{:<40} {:<10.2f}".format( |
| 1285 | "Max probe E2EL (ms):", probe_stats["probe_max_e2el_ms"] |
| 1286 | ) |
| 1287 | ) |
| 1288 | |
| 1289 | result: dict[str, Any] |
| 1290 | if isinstance(metrics, BenchmarkMetrics): |
| 1291 | result = { |
| 1292 | "duration": benchmark_duration, |
| 1293 | "completed": metrics.completed, |
| 1294 | "failed": metrics.failed, |
| 1295 | "total_input_tokens": metrics.total_input, |
| 1296 | "total_output_tokens": metrics.total_output, |
| 1297 | "request_throughput": metrics.request_throughput, |
| 1298 | "request_goodput": metrics.request_goodput if goodput_config_dict else None, |
| 1299 | "output_throughput": metrics.output_throughput, |
| 1300 | "total_token_throughput": metrics.total_token_throughput, |
| 1301 | "input_lens": [output.prompt_len for output in outputs], |
| 1302 | "output_lens": actual_output_lens, |
| 1303 | "ttfts": [output.ttft for output in outputs], |
| 1304 | "itls": [output.itl for output in outputs], |
| 1305 | "latencies": [output.latency for output in outputs], |
| 1306 | "start_times": [output.start_time for output in outputs], |
| 1307 | "queue_times": [output.client_queue_time for output in outputs], |
| 1308 | "generated_texts": [output.generated_text for output in outputs], |
| 1309 | "errors": [output.error for output in outputs], |
| 1310 | "max_output_tokens_per_s": metrics.max_output_tokens_per_s, |
| 1311 | "max_concurrent_requests": metrics.max_concurrent_requests, |
| 1312 | "rtfx": metrics.rtfx, |
| 1313 | } |
| 1314 | else: |
| 1315 | result = { |
| 1316 | "duration": benchmark_duration, |
| 1317 | "completed": metrics.completed, |
| 1318 | "total_input_tokens": metrics.total_input, |
| 1319 | "total_input_sequences": metrics.total_input_sequences, |
| 1320 | "request_throughput": metrics.request_throughput, |
| 1321 | "input_sequence_throughput": metrics.input_sequence_throughput, |
| 1322 | "total_token_throughput": metrics.total_token_throughput, |
| 1323 | "input_lens": [output.prompt_len for output in outputs], |
| 1324 | "latencies": [output.latency for output in outputs], |
| 1325 | "queue_times": [output.client_queue_time for output in outputs], |
| 1326 | "errors": [output.error for output in outputs], |
| 1327 | } |
| 1328 | |
| 1329 | queue_times: list[float] | None = None |
| 1330 | e2els_including_queue: list[float] | None = None |
| 1331 | if max_concurrency is not None: |
| 1332 | queue_times = [output.client_queue_time for output in outputs if output.success] |
| 1333 | if not math.isinf(request_rate): |
| 1334 | e2els_including_queue = [ |
| 1335 | output.latency + output.client_queue_time |
| 1336 | for output in outputs |
| 1337 | if output.success |
| 1338 | ] |
| 1339 | |
| 1340 | if probe_stats is not None: |
| 1341 | result.update(probe_stats) |
| 1342 | |
| 1343 | if rps_change_events: |
| 1344 | result["rps_change_events"] = rps_change_events |
| 1345 | |
| 1346 | if spec_decode_stats is not None: |
| 1347 | result["spec_decode_acceptance_rate"] = spec_decode_stats["acceptance_rate"] |
| 1348 | result["spec_decode_acceptance_length"] = spec_decode_stats["acceptance_length"] |
| 1349 | result["spec_decode_num_drafts"] = int(spec_decode_stats["num_drafts"]) |
| 1350 | result["spec_decode_draft_tokens"] = int(spec_decode_stats["draft_tokens"]) |
| 1351 | result["spec_decode_accepted_tokens"] = int( |
| 1352 | spec_decode_stats["accepted_tokens"] |
| 1353 | ) |
| 1354 | result["spec_decode_per_position_acceptance_rates"] = spec_decode_stats.get( |
| 1355 | "per_position_acceptance_rates", [] |
| 1356 | ) |
| 1357 | |
| 1358 | if diffusion_stats is not None: |
| 1359 | result["diffusion_committed_throughput"] = diffusion_stats[ |
| 1360 | "committed_throughput" |
| 1361 | ] |
| 1362 | result["diffusion_steps_per_canvas"] = diffusion_stats["steps_per_canvas"] |
| 1363 | result["diffusion_committed_per_step"] = diffusion_stats["committed_per_step"] |
| 1364 | result["diffusion_committed_tokens"] = int(diffusion_stats["committed_tokens"]) |
| 1365 | result["diffusion_denoising_steps"] = int(diffusion_stats["denoising_steps"]) |
| 1366 | result["diffusion_canvas_positions"] = int(diffusion_stats["canvas_positions"]) |
| 1367 | |
| 1368 | def process_one_metric( |
| 1369 | # E.g., "ttft" |
| 1370 | metric_attribute_name: str, |
| 1371 | # E.g., "TTFT" |
| 1372 | metric_name: str, |
| 1373 | # E.g., "Time to First Token" |
| 1374 | metric_header: str, |
| 1375 | values: list[float] | None = None, |
| 1376 | ): |
| 1377 | # This function prints and adds statistics of the specified |
| 1378 | # metric. |
| 1379 | if metric_attribute_name not in selected_percentile_metrics: |
| 1380 | return |
| 1381 | if values is None: |
| 1382 | mean = getattr(metrics, f"mean_{metric_attribute_name}_ms") |
| 1383 | median = getattr(metrics, f"median_{metric_attribute_name}_ms") |
| 1384 | std = getattr(metrics, f"std_{metric_attribute_name}_ms") |
| 1385 | percentiles = getattr(metrics, f"percentiles_{metric_attribute_name}_ms") |
| 1386 | else: |
| 1387 | mean = np.mean(values or 0) * 1000 |
| 1388 | median = np.median(values or 0) * 1000 |
| 1389 | std = np.std(values or 0) * 1000 |
| 1390 | percentiles = [ |
| 1391 | (p, np.percentile(values or 0, p) * 1000) for p in selected_percentiles |
| 1392 | ] |
| 1393 | print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-")) |
| 1394 | print("{:<40} {:<10.2f}".format(f"Mean {metric_name} (ms):", mean)) |
| 1395 | print("{:<40} {:<10.2f}".format(f"Median {metric_name} (ms):", median)) |
| 1396 | result[f"mean_{metric_attribute_name}_ms"] = mean |
| 1397 | result[f"median_{metric_attribute_name}_ms"] = median |
| 1398 | result[f"std_{metric_attribute_name}_ms"] = std |
| 1399 | for p, value in percentiles: |
| 1400 | p_word = str(int(p)) if int(p) == p else str(p) |
| 1401 | print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value)) |
| 1402 | result[f"p{p_word}_{metric_attribute_name}_ms"] = value |
| 1403 | |
| 1404 | if task_type == TaskType.GENERATION and tokenizer: |
| 1405 | process_one_metric("ttft", "TTFT", "Time to First Token") |
| 1406 | process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)") |
| 1407 | process_one_metric("itl", "ITL", "Inter-token Latency") |
| 1408 | process_one_metric("e2el", "E2EL", "End-to-end Latency") |
| 1409 | if queue_times is not None: |
| 1410 | process_one_metric( |
| 1411 | "client_queue_time", |
| 1412 | "Client Queue Time", |
| 1413 | "Client-side Queueing", |
| 1414 | queue_times, |
| 1415 | ) |
| 1416 | if e2els_including_queue is not None: |
| 1417 | process_one_metric( |
| 1418 | "e2el_including_client_queue", |
| 1419 | "E2EL incl. Client Queue", |
| 1420 | "Queue-inclusive End-to-end Latency", |
| 1421 | e2els_including_queue, |
| 1422 | ) |
| 1423 | |
| 1424 | if diffusion_stats is not None: |
| 1425 | print("{s:{c}^{n}}".format(s="Diffusion Decoding", n=50, c="-")) |
| 1426 | for label, key, value_fmt in ( |
| 1427 | ("Committed throughput (tok/s):", "committed_throughput", "{:<10.2f}"), |
| 1428 | ("Denoising steps per canvas:", "steps_per_canvas", "{:<10.2f}"), |
| 1429 | ("Committed per denoising step:", "committed_per_step", "{:<10.2f}"), |
| 1430 | ("Committed tokens:", "committed_tokens", "{:<10d}"), |
| 1431 | ("Denoising steps:", "denoising_steps", "{:<10d}"), |
| 1432 | ("Canvas positions evaluated:", "canvas_positions", "{:<10d}"), |
| 1433 | ): |
| 1434 | value = diffusion_stats[key] |
| 1435 | if value_fmt.endswith("d}"): |
| 1436 | value = int(value) |
| 1437 | print("{:<40} ".format(label) + value_fmt.format(value)) |
| 1438 | |
| 1439 | if spec_decode_stats is not None and diffusion_stats is None: |
| 1440 | print("{s:{c}^{n}}".format(s="Speculative Decoding", n=50, c="-")) |
| 1441 | print( |
| 1442 | "{:<40} {:<10.2f}".format( |
| 1443 | "Acceptance rate (%):", spec_decode_stats["acceptance_rate"] |
| 1444 | ) |
| 1445 | ) |
| 1446 | print( |
| 1447 | "{:<40} {:<10.2f}".format( |
| 1448 | "Acceptance length:", spec_decode_stats["acceptance_length"] |
| 1449 | ) |
| 1450 | ) |
| 1451 | print("{:<40} {:<10}".format("Drafts:", int(spec_decode_stats["num_drafts"]))) |
| 1452 | print( |
| 1453 | "{:<40} {:<10}".format( |
| 1454 | "Draft tokens:", int(spec_decode_stats["draft_tokens"]) |
| 1455 | ) |
| 1456 | ) |
| 1457 | print( |
| 1458 | "{:<40} {:<10}".format( |
| 1459 | "Accepted tokens:", int(spec_decode_stats["accepted_tokens"]) |
| 1460 | ) |
| 1461 | ) |
| 1462 | per_pos = spec_decode_stats.get("per_position_acceptance_rates", []) |
| 1463 | if per_pos: |
| 1464 | print("Per-position acceptance (%):") |
| 1465 | for i, rate in enumerate(per_pos): |
| 1466 | print("{:<40} {:<10.2f}".format(f" Position {i}:", rate * 100)) |
| 1467 | |
| 1468 | print("=" * 50) |
| 1469 | |
| 1470 | if profile: |
| 1471 | print("Stopping profiler...") |
| 1472 | profile_input = RequestFuncInput( |
| 1473 | model=model_id, |
| 1474 | prompt=test_prompt, |
| 1475 | api_url=base_url + "/stop_profile", |
| 1476 | prompt_len=test_prompt_len, |
| 1477 | output_len=test_output_len, |
| 1478 | logprobs=logprobs, |
| 1479 | ) |
| 1480 | profile_output = await request_func( |
| 1481 | request_func_input=profile_input, session=session |
| 1482 | ) |
| 1483 | if profile_output.success: |
| 1484 | print("Profiler stopped") |
| 1485 | |
| 1486 | await session.close() |
| 1487 | return result |
| 1488 | |
| 1489 | |
| 1490 | def check_goodput_args(args): |
| 1491 | # Check and parse goodput arguments |
| 1492 | goodput_config_dict = {} |
| 1493 | VALID_NAMES = ["ttft", "tpot", "e2el"] |
| 1494 | if args.goodput: |
| 1495 | goodput_config_dict = parse_goodput(args.goodput) |
| 1496 | for slo_name, slo_val in goodput_config_dict.items(): |
| 1497 | if slo_name not in VALID_NAMES: |
| 1498 | raise ValueError( |
| 1499 | f"Invalid metric name found, {slo_name}: {slo_val}. " |
| 1500 | "The service level objective name should be one of " |
| 1501 | f"{str(VALID_NAMES)}. " |
| 1502 | ) |
| 1503 | if slo_val < 0: |
| 1504 | raise ValueError( |
| 1505 | f"Invalid value found, {slo_name}: {slo_val}. " |
| 1506 | "The service level objective value should be " |
| 1507 | "non-negative." |
| 1508 | ) |
| 1509 | return goodput_config_dict |
| 1510 | |
| 1511 | |
| 1512 | def parse_goodput(slo_pairs): |
| 1513 | goodput_config_dict = {} |
| 1514 | try: |
| 1515 | for slo_pair in slo_pairs: |
| 1516 | slo_name, slo_val = slo_pair.split(":") |
| 1517 | goodput_config_dict[slo_name] = float(slo_val) |
| 1518 | except ValueError as err: |
| 1519 | raise argparse.ArgumentTypeError( |
| 1520 | "Invalid format found for service level objectives. " |
| 1521 | 'Specify service level objectives for goodput as "KEY:VALUE" ' |
| 1522 | "pairs, where the key is a metric name, and the value is a " |
| 1523 | "number in milliseconds." |
| 1524 | ) from err |
| 1525 | return goodput_config_dict |
| 1526 | |
| 1527 | |
| 1528 | def save_to_pytorch_benchmark_format( |
| 1529 | args: argparse.Namespace, results: dict[str, Any], file_name: str |
| 1530 | ) -> None: |
| 1531 | metrics = [ |
| 1532 | "median_ttft_ms", |
| 1533 | "mean_ttft_ms", |
| 1534 | "std_ttft_ms", |
| 1535 | "p99_ttft_ms", |
| 1536 | "mean_tpot_ms", |
| 1537 | "median_tpot_ms", |
| 1538 | "std_tpot_ms", |
| 1539 | "p99_tpot_ms", |
| 1540 | "median_itl_ms", |
| 1541 | "mean_itl_ms", |
| 1542 | "std_itl_ms", |
| 1543 | "p99_itl_ms", |
| 1544 | ] |
| 1545 | # These raw data might be useful, but they are rather big. They can be added |
| 1546 | # later if needed |
| 1547 | ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"] |
| 1548 | pt_records = convert_to_pytorch_benchmark_format( |
| 1549 | args=redact_sensitive_namespace(args, _SENSITIVE_ARG_FIELDS), |
| 1550 | metrics={k: [results[k]] for k in metrics if k in results}, |
| 1551 | extra_info={ |
| 1552 | k: results[k] |
| 1553 | for k in results |
| 1554 | if k not in metrics and k not in ignored_metrics |
| 1555 | }, |
| 1556 | ) |
| 1557 | if pt_records: |
| 1558 | # Don't use json suffix here as we don't want CI to pick it up |
| 1559 | pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json" |
| 1560 | write_to_json(pt_file, pt_records) |
| 1561 | |
| 1562 | |
| 1563 | def compute_result_filename( |
| 1564 | args: argparse.Namespace, |
| 1565 | model_id: str, |
| 1566 | label: str, |
| 1567 | current_dt: str, |
| 1568 | ) -> str | None: |
| 1569 | """Compute the result filename based on benchmark configuration. |
| 1570 | |
| 1571 | Args: |
| 1572 | args: Command line arguments containing result configuration |
| 1573 | model_id: The model identifier |
| 1574 | label: The benchmark label |
| 1575 | current_dt: Current datetime string |
| 1576 | |
| 1577 | Returns: |
| 1578 | The computed filename path or None if no result saving is requested |
| 1579 | |
| 1580 | """ |
| 1581 | if not (args.plot_timeline or args.save_result or args.append_result): |
| 1582 | return None |
| 1583 | |
| 1584 | base_model_id = model_id.split("/")[-1] |
| 1585 | max_concurrency_str = ( |
| 1586 | f"-concurrency{args.max_concurrency}" |
| 1587 | if args.max_concurrency is not None |
| 1588 | else "" |
| 1589 | ) |
| 1590 | label = label or args.backend |
| 1591 | |
| 1592 | if args.ramp_up_strategy is not None: |
| 1593 | file_name = f"{label}-ramp-up-{args.ramp_up_strategy}-{args.ramp_up_start_rps}qps-{args.ramp_up_end_rps}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa |
| 1594 | else: |
| 1595 | file_name = f"{label}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json" # noqa |
| 1596 | |
| 1597 | if args.result_filename: |
| 1598 | file_name = args.result_filename |
| 1599 | |
| 1600 | if args.result_dir: |
| 1601 | os.makedirs(args.result_dir, exist_ok=True) |
| 1602 | file_name = os.path.join(args.result_dir, file_name) |
| 1603 | |
| 1604 | return file_name |
| 1605 | |
| 1606 | |
| 1607 | def add_cli_args(parser: FlexibleArgumentParser): |
| 1608 | add_dataset_parser(parser) |
| 1609 | parser.add_argument( |
| 1610 | "--label", |
| 1611 | type=str, |
| 1612 | default=None, |
| 1613 | help="The label (prefix) of the benchmark results. If not specified, " |
| 1614 | "the value of '--backend' will be used as the label.", |
| 1615 | ) |
| 1616 | parser.add_argument( |
| 1617 | "--backend", |
| 1618 | type=str, |
| 1619 | default="openai", |
| 1620 | choices=list(ASYNC_REQUEST_FUNCS.keys()), |
| 1621 | help="The type of backend or endpoint to use for the benchmark.", |
| 1622 | ) |
| 1623 | parser.add_argument( |
| 1624 | "--base-url", |
| 1625 | type=str, |
| 1626 | default=None, |
| 1627 | help="Server or API base url if not using http host and port.", |
| 1628 | ) |
| 1629 | # Use 127.0.0.1 here instead of localhost to force the use of ipv4 |
| 1630 | parser.add_argument("--host", type=str, default="127.0.0.1") |
| 1631 | parser.add_argument("--port", type=int, default=8000) |
| 1632 | parser.add_argument( |
| 1633 | "--endpoint", |
| 1634 | type=str, |
| 1635 | default="/v1/completions", |
| 1636 | help="API endpoint.", |
| 1637 | ) |
| 1638 | parser.add_argument( |
| 1639 | "--header", |
| 1640 | metavar="KEY=VALUE", |
| 1641 | nargs="*", |
| 1642 | help="Key-value pairs (e.g, --header x-additional-info=0.3.3) " |
| 1643 | "for headers to be passed with each request. These headers override " |
| 1644 | "per backend constants and values set via environment variable, and " |
| 1645 | "will be overridden by other arguments (such as request ids).", |
| 1646 | ) |
| 1647 | parser.add_argument( |
| 1648 | "--max-concurrency", |
| 1649 | type=int, |
| 1650 | default=None, |
| 1651 | help="Maximum number of concurrent requests. This can be used " |
| 1652 | "to help simulate an environment where a higher level component " |
| 1653 | "is enforcing a maximum number of concurrent requests. While the " |
| 1654 | "--request-rate argument controls the rate at which requests are " |
| 1655 | "initiated, this argument will control how many are actually allowed " |
| 1656 | "to execute at a time. This means that when used in combination, the " |
| 1657 | "actual request rate may be lower than specified with --request-rate, " |
| 1658 | "if the server is not processing requests fast enough to keep up.", |
| 1659 | ) |
| 1660 | |
| 1661 | parser.add_argument( |
| 1662 | "--model", |
| 1663 | type=str, |
| 1664 | required=False, |
| 1665 | default=None, |
| 1666 | help="Name of the model. If not specified, will fetch the first model " |
| 1667 | "from the server's /v1/models endpoint.", |
| 1668 | ) |
| 1669 | parser.add_argument( |
| 1670 | "--input-len", |
| 1671 | type=int, |
| 1672 | default=None, |
| 1673 | help="General input length for datasets. Maps to dataset-specific " |
| 1674 | "input length arguments (e.g., --random-input-len, --sonnet-input-len). " |
| 1675 | "If not specified, uses dataset defaults.", |
| 1676 | ) |
| 1677 | parser.add_argument( |
| 1678 | "--output-len", |
| 1679 | type=int, |
| 1680 | default=None, |
| 1681 | help="General output length for datasets. Maps to dataset-specific " |
| 1682 | "output length arguments (e.g., --random-output-len, --sonnet-output-len). " |
| 1683 | "If not specified, uses dataset defaults.", |
| 1684 | ) |
| 1685 | parser.add_argument( |
| 1686 | "--tokenizer", |
| 1687 | type=str, |
| 1688 | help="Name or path of the tokenizer, if not using the default tokenizer.", # noqa: E501 |
| 1689 | ) |
| 1690 | parser.add_argument( |
| 1691 | "--tokenizer-mode", |
| 1692 | type=str, |
| 1693 | default="auto", |
| 1694 | help="""Tokenizer mode:\n |
| 1695 | - "auto" will use the tokenizer from `mistral_common` for Mistral models |
| 1696 | if available, otherwise it will use the "hf" tokenizer.\n |
| 1697 | - "hf" will use the fast tokenizer if available.\n |
| 1698 | - "mistral" will always use the tokenizer from `mistral_common`.\n |
| 1699 | - "deepseek_v32" will always use the tokenizer from `deepseek_v32`.\n |
| 1700 | - Other custom values can be supported via plugins.""", |
| 1701 | ) |
| 1702 | parser.add_argument("--use-beam-search", action="store_true") |
| 1703 | parser.add_argument( |
| 1704 | "--logprobs", |
| 1705 | type=int, |
| 1706 | default=None, |
| 1707 | help=( |
| 1708 | "Number of logprobs-per-token to compute & return as part of " |
| 1709 | "the request. If unspecified, then either (1) if beam search " |
| 1710 | "is disabled, no logprobs are computed & a single dummy " |
| 1711 | "logprob is returned for each token; or (2) if beam search " |
| 1712 | "is enabled 1 logprob per token is computed" |
| 1713 | ), |
| 1714 | ) |
| 1715 | parser.add_argument( |
| 1716 | "--request-rate", |
| 1717 | type=float, |
| 1718 | default=float("inf"), |
| 1719 | help="Number of requests per second. If this is inf, " |
| 1720 | "then all the requests are sent at time 0. " |
| 1721 | "Otherwise, we use Poisson process or gamma distribution " |
| 1722 | "to synthesize the request arrival times.", |
| 1723 | ) |
| 1724 | parser.add_argument( |
| 1725 | "--burstiness", |
| 1726 | type=float, |
| 1727 | default=1.0, |
| 1728 | help="Burstiness factor of the request generation. " |
| 1729 | "Only take effect when request_rate is not inf. " |
| 1730 | "Default value is 1, which follows Poisson process. " |
| 1731 | "Otherwise, the request intervals follow a gamma distribution. " |
| 1732 | "A lower burstiness value (0 < burstiness < 1) results in more " |
| 1733 | "bursty requests. A higher burstiness value (burstiness > 1) " |
| 1734 | "results in a more uniform arrival of requests.", |
| 1735 | ) |
| 1736 | parser.add_argument( |
| 1737 | "--probe-request-rate", |
| 1738 | type=float, |
| 1739 | default=0.0, |
| 1740 | help="If positive, send single-token text-only probe requests at " |
| 1741 | "this rate (req/s) alongside the main workload, bypassing " |
| 1742 | "--max-concurrency, and report their latency separately. Useful " |
| 1743 | "for measuring how the main workload stalls unrelated requests.", |
| 1744 | ) |
| 1745 | parser.add_argument( |
| 1746 | "--disable-tqdm", |
| 1747 | action="store_true", |
| 1748 | help="Specify to disable tqdm progress bar.", |
| 1749 | ) |
| 1750 | parser.add_argument( |
| 1751 | "--num-warmups", |
| 1752 | type=int, |
| 1753 | default=0, |
| 1754 | help="Number of warmup requests.", |
| 1755 | ) |
| 1756 | parser.add_argument( |
| 1757 | "--profile", |
| 1758 | action="store_true", |
| 1759 | help="Use vLLM Profiling. --profiler-config must be provided on the server.", |
| 1760 | ) |
| 1761 | parser.add_argument( |
| 1762 | "--save-result", |
| 1763 | action="store_true", |
| 1764 | help="Specify to save benchmark results to a json file", |
| 1765 | ) |
| 1766 | parser.add_argument( |
| 1767 | "--save-detailed", |
| 1768 | action="store_true", |
| 1769 | help="When saving the results, whether to include per request " |
| 1770 | "information such as response, error, ttfts, tpots, etc.", |
| 1771 | ) |
| 1772 | parser.add_argument( |
| 1773 | "--append-result", |
| 1774 | action="store_true", |
| 1775 | help="Append the benchmark result to the existing json file.", |
| 1776 | ) |
| 1777 | parser.add_argument( |
| 1778 | "--metadata", |
| 1779 | metavar="KEY=VALUE", |
| 1780 | nargs="*", |
| 1781 | help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) " |
| 1782 | "for metadata of this run to be saved in the result JSON file " |
| 1783 | "for record keeping purposes.", |
| 1784 | ) |
| 1785 | parser.add_argument( |
| 1786 | "--result-dir", |
| 1787 | type=str, |
| 1788 | default=None, |
| 1789 | help="Specify directory to save benchmark json results." |
| 1790 | "If not specified, results are saved in the current directory.", |
| 1791 | ) |
| 1792 | parser.add_argument( |
| 1793 | "--result-filename", |
| 1794 | type=str, |
| 1795 | default=None, |
| 1796 | help="Specify the filename to save benchmark json results." |
| 1797 | "If not specified, results will be saved in " |
| 1798 | "{label}-{args.request_rate}qps-{base_model_id}-{current_dt}.json" # noqa |
| 1799 | " format.", |
| 1800 | ) |
| 1801 | parser.add_argument( |
| 1802 | "--ignore-eos", |
| 1803 | action="store_true", |
| 1804 | help="Set ignore_eos flag when sending the benchmark request." |
| 1805 | "Warning: ignore_eos is not supported in deepspeed_mii and tgi.", |
| 1806 | ) |
| 1807 | parser.add_argument( |
| 1808 | "--self-timed", |
| 1809 | action=argparse.BooleanOptionalAction, |
| 1810 | default=None, |
| 1811 | help="Use timing information from the traces instead of the configuration. " |
| 1812 | "This is useful when replaying traces faithfully based on their timestamps. " |
| 1813 | "When unset, defaults to False, except for --dataset-name=timed_trace where " |
| 1814 | "it defaults to True. Use --no-self-timed to force off. When off, user " |
| 1815 | "defined generation rates are used and in trace timing info is ignored.", |
| 1816 | ) |
| 1817 | parser.add_argument( |
| 1818 | "--percentile-metrics", |
| 1819 | type=str, |
| 1820 | default=None, |
| 1821 | help="Comma-separated list of selected metrics to report percentiles. " |
| 1822 | "This argument specifies the metrics to report percentiles. " |
| 1823 | 'Allowed metric names are "ttft", "tpot", "itl", "e2el", ' |
| 1824 | '"client_queue_time", "e2el_including_client_queue". ' |
| 1825 | 'If not specified, defaults to "ttft,tpot,itl" for generative models ' |
| 1826 | 'and "e2el" for pooling models.', |
| 1827 | ) |
| 1828 | parser.add_argument( |
| 1829 | "--metric-percentiles", |
| 1830 | type=str, |
| 1831 | default="99", |
| 1832 | help="Comma-separated list of percentiles for selected metrics. " |
| 1833 | 'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". ' |
| 1834 | 'Default value is "99".' |
| 1835 | 'Use "--percentile-metrics" to select metrics.', |
| 1836 | ) |
| 1837 | parser.add_argument( |
| 1838 | "--goodput", |
| 1839 | nargs="+", |
| 1840 | required=False, |
| 1841 | help='Specify service level objectives for goodput as "KEY:VALUE" ' |
| 1842 | "pairs, where the key is a metric name, and the value is in " |
| 1843 | 'milliseconds. Multiple "KEY:VALUE" pairs can be provided, ' |
| 1844 | "separated by spaces. Allowed request level metric names are " |
| 1845 | '"ttft", "tpot", "e2el". For more context on the definition of ' |
| 1846 | "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 " |
| 1847 | "and the blog: https://hao-ai-lab.github.io/blogs/distserve", |
| 1848 | ) |
| 1849 | parser.add_argument( |
| 1850 | "--request-id-prefix", |
| 1851 | type=str, |
| 1852 | required=False, |
| 1853 | default=f"bench-{uuid.uuid4().hex[:8]}-", |
| 1854 | help="Specify the prefix of request id.", |
| 1855 | ) |
| 1856 | |
| 1857 | sampling_group = parser.add_argument_group("sampling parameters") |
| 1858 | sampling_group.add_argument( |
| 1859 | "--top-p", |
| 1860 | type=float, |
| 1861 | default=None, |
| 1862 | help="Top-p sampling parameter. Only has effect on openai-compatible backends.", |
| 1863 | ) |
| 1864 | sampling_group.add_argument( |
| 1865 | "--top-k", |
| 1866 | type=int, |
| 1867 | default=None, |
| 1868 | help="Top-k sampling parameter. Only has effect on openai-compatible backends.", |
| 1869 | ) |
| 1870 | sampling_group.add_argument( |
| 1871 | "--min-p", |
| 1872 | type=float, |
| 1873 | default=None, |
| 1874 | help="Min-p sampling parameter. Only has effect on openai-compatible backends.", |
| 1875 | ) |
| 1876 | sampling_group.add_argument( |
| 1877 | "--temperature", |
| 1878 | type=float, |
| 1879 | default=None, |
| 1880 | help="Temperature sampling parameter. Only has effect on " |
| 1881 | "openai-compatible backends.", |
| 1882 | ) |
| 1883 | sampling_group.add_argument( |
| 1884 | "--frequency-penalty", |
| 1885 | type=float, |
| 1886 | default=None, |
| 1887 | help="Frequency penalty sampling parameter. Only has effect on " |
| 1888 | "openai-compatible backends.", |
| 1889 | ) |
| 1890 | sampling_group.add_argument( |
| 1891 | "--presence-penalty", |
| 1892 | type=float, |
| 1893 | default=None, |
| 1894 | help="Presence penalty sampling parameter. Only has effect on " |
| 1895 | "openai-compatible backends.", |
| 1896 | ) |
| 1897 | sampling_group.add_argument( |
| 1898 | "--repetition-penalty", |
| 1899 | type=float, |
| 1900 | default=None, |
| 1901 | help="Repetition penalty sampling parameter. Only has effect on " |
| 1902 | "openai-compatible backends.", |
| 1903 | ) |
| 1904 | |
| 1905 | parser.add_argument( |
| 1906 | "--served-model-name", |
| 1907 | type=str, |
| 1908 | default=None, |
| 1909 | help="The model name used in the API. " |
| 1910 | "If not specified, the model name will be the " |
| 1911 | "same as the `--model` argument. ", |
| 1912 | ) |
| 1913 | |
| 1914 | parser.add_argument( |
| 1915 | "--lora-modules", |
| 1916 | nargs="+", |
| 1917 | default=None, |
| 1918 | help="A subset of LoRA module names passed in when " |
| 1919 | "launching the server. For each request, the " |
| 1920 | "script chooses a LoRA module at random by default. " |
| 1921 | "Use --lora-assignment to control selection strategy.", |
| 1922 | ) |
| 1923 | |
| 1924 | parser.add_argument( |
| 1925 | "--lora-assignment", |
| 1926 | type=str, |
| 1927 | default="random", |
| 1928 | choices=["random", "round-robin"], |
| 1929 | help="Strategy for assigning LoRA modules to requests. " |
| 1930 | "'random' (default) selects a LoRA at random for each request. " |
| 1931 | "'round-robin' cycles through LoRA modules deterministically.", |
| 1932 | ) |
| 1933 | |
| 1934 | parser.add_argument( |
| 1935 | "--ramp-up-strategy", |
| 1936 | type=str, |
| 1937 | default=None, |
| 1938 | choices=["linear", "exponential"], |
| 1939 | help="The ramp-up strategy. This would be used to " |
| 1940 | "ramp up the request rate from initial RPS to final " |
| 1941 | "RPS rate (specified by --ramp-up-start-rps and " |
| 1942 | "--ramp-up-end-rps.) over the duration of the benchmark.", |
| 1943 | ) |
| 1944 | parser.add_argument( |
| 1945 | "--ramp-up-start-rps", |
| 1946 | type=int, |
| 1947 | default=None, |
| 1948 | help="The starting request rate for ramp-up (RPS). " |
| 1949 | "Needs to be specified when --ramp-up-strategy is used.", |
| 1950 | ) |
| 1951 | parser.add_argument( |
| 1952 | "--ramp-up-end-rps", |
| 1953 | type=int, |
| 1954 | default=None, |
| 1955 | help="The ending request rate for ramp-up (RPS). " |
| 1956 | "Needs to be specified when --ramp-up-strategy is used.", |
| 1957 | ) |
| 1958 | parser.add_argument( |
| 1959 | "--ready-check-timeout-sec", |
| 1960 | type=int, |
| 1961 | default=0, |
| 1962 | help="Maximum time to wait for the endpoint to become ready " |
| 1963 | "in seconds. Ready check will be skipped by default.", |
| 1964 | ) |
| 1965 | |
| 1966 | parser.add_argument( |
| 1967 | "--chat-template-kwargs", |
| 1968 | type=json.loads, |
| 1969 | default=None, |
| 1970 | help="A JSON string of kwargs forwarded to the tokenizer's " |
| 1971 | "apply_chat_template when a dataset renders prompts client-side " |
| 1972 | "(e.g. custom / speed_bench). " |
| 1973 | "Example: '{\"thinking\": true}' to enable reasoning models.", |
| 1974 | ) |
| 1975 | parser.add_argument( |
| 1976 | "--extra-body", |
| 1977 | help="A JSON string representing extra body parameters to include " |
| 1978 | "in each request." |
| 1979 | 'Example: \'{"chat_template_kwargs":{"enable_thinking":false}}\'', |
| 1980 | type=json.loads, |
| 1981 | default=None, |
| 1982 | ) |
| 1983 | parser.add_argument( |
| 1984 | "--skip-tokenizer-init", |
| 1985 | action="store_true", |
| 1986 | default=False, |
| 1987 | help="Skip initialization of tokenizer and detokenizer", |
| 1988 | ) |
| 1989 | |
| 1990 | parser.add_argument( |
| 1991 | "--insecure", |
| 1992 | action="store_true", |
| 1993 | default=False, |
| 1994 | help="Disable SSL certificate verification. Use this option when " |
| 1995 | "connecting to servers with self-signed certificates.", |
| 1996 | ) |
| 1997 | |
| 1998 | parser.add_argument( |
| 1999 | "--plot-timeline", |
| 2000 | action="store_true", |
| 2001 | help="Generate an HTML timeline plot showing request execution. " |
| 2002 | "The plot will be saved alongside the results JSON file.", |
| 2003 | ) |
| 2004 | parser.add_argument( |
| 2005 | "--timeline-itl-thresholds", |
| 2006 | type=str, |
| 2007 | default="25,50", |
| 2008 | help="ITL thresholds in milliseconds for timeline plot coloring. " |
| 2009 | "Specify two comma-separated values to categorize inter-token " |
| 2010 | "latencies into three groups: below first threshold (green), " |
| 2011 | "between thresholds (orange), and above second threshold (red).", |
| 2012 | ) |
| 2013 | parser.add_argument( |
| 2014 | "--plot-dataset-stats", |
| 2015 | action="store_true", |
| 2016 | help="Generate a matplotlib figure with dataset statistics showing " |
| 2017 | "prompt tokens, output tokens, and combined token distributions.", |
| 2018 | ) |
| 2019 | |
| 2020 | |
| 2021 | def main(args: argparse.Namespace) -> dict[str, Any]: |
| 2022 | return asyncio.run(main_async(args)) |
| 2023 | |
| 2024 | |
| 2025 | async def main_async(args: argparse.Namespace) -> dict[str, Any]: |
| 2026 | print(redact_sensitive_namespace(args, _SENSITIVE_ARG_FIELDS)) |
| 2027 | if args.max_concurrency is not None and args.max_concurrency <= 0: |
| 2028 | raise ValueError("--max-concurrency must be greater than 0") |
| 2029 | |
| 2030 | random.seed(args.seed) |
| 2031 | np.random.seed(args.seed) |
| 2032 | |
| 2033 | # Validate timeline ITL thresholds |
| 2034 | if args.plot_timeline: |
| 2035 | try: |
| 2036 | itl_thresholds = [ |
| 2037 | float(t.strip()) for t in args.timeline_itl_thresholds.split(",") |
| 2038 | ] |
| 2039 | if len(itl_thresholds) != 2: |
| 2040 | raise ValueError( |
| 2041 | f"Expected 2 ITL threshold values, got {len(itl_thresholds)}" |
| 2042 | ) |
| 2043 | except ValueError as e: |
| 2044 | raise ValueError(f"Invalid --timeline-itl-thresholds format: {e}") from e |
| 2045 | |
| 2046 | # Validate ramp-up arguments |
| 2047 | if args.ramp_up_strategy is not None: |
| 2048 | if args.request_rate != float("inf"): |
| 2049 | raise ValueError( |
| 2050 | "When using ramp-up, do not specify --request-rate. " |
| 2051 | "The request rate will be controlled by ramp-up parameters. " |
| 2052 | "Please remove the --request-rate argument." |
| 2053 | ) |
| 2054 | if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None: |
| 2055 | raise ValueError( |
| 2056 | "When using --ramp-up-strategy, both --ramp-up-start-rps and " |
| 2057 | "--ramp-up-end-rps must be specified" |
| 2058 | ) |
| 2059 | if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0: |
| 2060 | raise ValueError("Ramp-up start and end RPS must be non-negative") |
| 2061 | if args.ramp_up_start_rps > args.ramp_up_end_rps: |
| 2062 | raise ValueError("Ramp-up start RPS must be less than end RPS") |
| 2063 | if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0: |
| 2064 | raise ValueError("For exponential ramp-up, the start RPS cannot be 0.") |
| 2065 | |
| 2066 | label = args.label |
| 2067 | |
| 2068 | if args.base_url is not None: |
| 2069 | api_url = f"{args.base_url}{args.endpoint}" |
| 2070 | base_url = f"{args.base_url}" |
| 2071 | else: |
| 2072 | host_port = join_host_port(args.host, args.port) |
| 2073 | api_url = f"http://{host_port}{args.endpoint}" |
| 2074 | base_url = f"http://{host_port}" |
| 2075 | |
| 2076 | # Headers |
| 2077 | headers = None |
| 2078 | if args.header: |
| 2079 | headers = {} |
| 2080 | for item in args.header: |
| 2081 | if "=" in item: |
| 2082 | kvstring = item.split("=", 1) |
| 2083 | headers[kvstring[0].strip()] = kvstring[1].strip() |
| 2084 | else: |
| 2085 | raise ValueError("Invalid header format. Please use KEY=VALUE format.") |
| 2086 | |
| 2087 | # SSL context configuration |
| 2088 | ssl_context: ssl.SSLContext | bool | None = None |
| 2089 | if args.insecure: |
| 2090 | # Disable SSL certificate verification |
| 2091 | ssl_context = False |
| 2092 | elif "https://" in base_url: |
| 2093 | # Use default SSL context for HTTPS |
| 2094 | ssl_context = True |
| 2095 | |
| 2096 | # Fetch model from server if not specified |
| 2097 | if args.model is None: |
| 2098 | print("Model not specified, fetching first model from server...") |
| 2099 | model_name, model_id = await get_first_model_from_server( |
| 2100 | base_url, headers, ssl_context |
| 2101 | ) |
| 2102 | print(f"First model name: {model_name}, first model id: {model_id}") |
| 2103 | else: |
| 2104 | model_name = args.served_model_name |
| 2105 | model_id = args.model |
| 2106 | |
| 2107 | if args.skip_tokenizer_init: |
| 2108 | tokenizer_id = None |
| 2109 | tokenizer_mode = None |
| 2110 | tokenizer = None |
| 2111 | else: |
| 2112 | tokenizer_id = args.tokenizer if args.tokenizer is not None else model_id |
| 2113 | tokenizer_mode = args.tokenizer_mode |
| 2114 | tokenizer = get_tokenizer( |
| 2115 | tokenizer_id, |
| 2116 | tokenizer_mode=tokenizer_mode, |
| 2117 | trust_remote_code=args.trust_remote_code, |
| 2118 | ) |
| 2119 | |
| 2120 | # Validate dataset name/path |
| 2121 | if args.dataset_name is None: |
| 2122 | raise ValueError( |
| 2123 | "Please specify '--dataset-name' and the corresponding " |
| 2124 | "'--dataset-path' if required." |
| 2125 | ) |
| 2126 | |
| 2127 | if ( |
| 2128 | args.dataset_name |
| 2129 | in ["random", "random-mm", "random-rerank", "prefix_repetition"] |
| 2130 | and args.dataset_path is not None |
| 2131 | ): |
| 2132 | raise ValueError( |
| 2133 | f"Cannot use '{args.dataset_name}' dataset with --dataset-path. " |
| 2134 | "Please specify the appropriate --dataset-name (e.g., " |
| 2135 | "'sharegpt', 'custom', 'sonnet') for your dataset file: " |
| 2136 | f"{args.dataset_path}" |
| 2137 | ) |
| 2138 | |
| 2139 | # Map general --input-len and --output-len to all dataset-specific arguments |
| 2140 | if args.input_len is not None: |
| 2141 | args.random_input_len = args.input_len |
| 2142 | args.sonnet_input_len = args.input_len |
| 2143 | |
| 2144 | if args.output_len is not None: |
| 2145 | args.random_output_len = args.output_len |
| 2146 | args.sonnet_output_len = args.output_len |
| 2147 | args.sharegpt_output_len = args.output_len |
| 2148 | args.custom_output_len = args.output_len |
| 2149 | args.hf_output_len = args.output_len |
| 2150 | args.spec_bench_output_len = args.output_len |
| 2151 | args.prefix_repetition_output_len = args.output_len |
| 2152 | |
| 2153 | # when using random datasets, default to ignoring EOS |
| 2154 | # so generation runs to the requested length |
| 2155 | if ( |
| 2156 | args.dataset_name in ("random", "random-mm") |
| 2157 | and args.backend in OPENAI_COMPATIBLE_BACKENDS |
| 2158 | ): |
| 2159 | args.ignore_eos = True |
| 2160 | |
| 2161 | if args.dataset_name == "timed_trace": |
| 2162 | if args.backend not in ("vllm", "openai"): |
| 2163 | raise ValueError( |
| 2164 | "timed_trace dataset passes pre-tokenized prompts (list[int])" |
| 2165 | " and requires a completions backend ('vllm' or 'openai')." |
| 2166 | ) |
| 2167 | # timed_trace carries per-request timestamps; |
| 2168 | # ignore EOS so generation runs to the trace's specified output length, |
| 2169 | # and default to using those timestamps for scheduling unless the user |
| 2170 | # opted out. |
| 2171 | args.ignore_eos = True |
| 2172 | if args.self_timed is None: |
| 2173 | args.self_timed = True |
| 2174 | else: |
| 2175 | # if this is set for anything else, it is an error |
| 2176 | if args.self_timed is not None: |
| 2177 | raise ValueError( |
| 2178 | "--self-timed/--no-self-timed is only supported with " |
| 2179 | "--dataset-name=timed_trace" |
| 2180 | ) |
| 2181 | # for any non self-timed trace, this is False |
| 2182 | args.self_timed = False |
| 2183 | |
| 2184 | # Load the dataset. |
| 2185 | input_requests = get_samples(args, tokenizer) |
| 2186 | |
| 2187 | if args.dataset_name in ("random", "prefix_repetition"): |
| 2188 | input_requests = await _align_prompts_to_server_tokenizer( |
| 2189 | base_url, model_id, input_requests, ssl_context |
| 2190 | ) |
| 2191 | |
| 2192 | goodput_config_dict = check_goodput_args(args) |
| 2193 | |
| 2194 | backend = args.backend |
| 2195 | task_type = TaskType.POOLING if backend in POOLING_BACKENDS else TaskType.GENERATION |
| 2196 | |
| 2197 | # Collect the sampling parameters. |
| 2198 | if task_type == TaskType.GENERATION: |
| 2199 | sampling_params = { |
| 2200 | k: v |
| 2201 | for k, v in { |
| 2202 | "top_p": args.top_p, |
| 2203 | "top_k": args.top_k, |
| 2204 | "min_p": args.min_p, |
| 2205 | "temperature": args.temperature, |
| 2206 | "frequency_penalty": args.frequency_penalty, |
| 2207 | "presence_penalty": args.presence_penalty, |
| 2208 | "repetition_penalty": args.repetition_penalty, |
| 2209 | }.items() |
| 2210 | if v is not None |
| 2211 | } |
| 2212 | |
| 2213 | # Sampling parameters are only supported by openai-compatible backend. |
| 2214 | if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS: |
| 2215 | raise ValueError( |
| 2216 | "Sampling parameters are only supported by openai-compatible backends." |
| 2217 | ) |
| 2218 | |
| 2219 | # The Responses API accepts every sampling parameter above except |
| 2220 | # min_p, which it would silently drop as an unknown field. |
| 2221 | if args.backend == "openai-responses" and "min_p" in sampling_params: |
| 2222 | raise ValueError("--min-p is not supported by the Responses API.") |
| 2223 | |
| 2224 | if "temperature" not in sampling_params: |
| 2225 | print( |
| 2226 | "WARNING: vllm bench serve no longer sets temperature==0 (greedy) " |
| 2227 | "in requests by default. The default will be determined on the " |
| 2228 | "server side and can be model/API specific. " |
| 2229 | "For the old behavior, include --temperature=0." |
| 2230 | ) |
| 2231 | |
| 2232 | default_percentile_metrics = "ttft,tpot,itl" |
| 2233 | else: |
| 2234 | sampling_params = {} |
| 2235 | default_percentile_metrics = "e2el" |
| 2236 | |
| 2237 | extra_body = args.extra_body or {} |
| 2238 | extra_body = {**sampling_params, **extra_body} |
| 2239 | |
| 2240 | percentile_metrics: str = args.percentile_metrics or default_percentile_metrics |
| 2241 | |
| 2242 | # Avoid GC processing "static" data - reduce pause times. |
| 2243 | freeze_gc_heap() |
| 2244 | |
| 2245 | benchmark_result = await benchmark( |
| 2246 | task_type=task_type, |
| 2247 | endpoint_type=backend, |
| 2248 | api_url=api_url, |
| 2249 | base_url=base_url, |
| 2250 | model_id=model_id, |
| 2251 | model_name=model_name, |
| 2252 | tokenizer=tokenizer, |
| 2253 | input_requests=input_requests, |
| 2254 | logprobs=args.logprobs, |
| 2255 | request_rate=args.request_rate, |
| 2256 | burstiness=args.burstiness, |
| 2257 | disable_tqdm=args.disable_tqdm, |
| 2258 | num_warmups=args.num_warmups, |
| 2259 | profile=args.profile, |
| 2260 | selected_percentile_metrics=percentile_metrics.split(","), |
| 2261 | selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")], |
| 2262 | ignore_eos=args.ignore_eos, |
| 2263 | goodput_config_dict=goodput_config_dict, |
| 2264 | max_concurrency=args.max_concurrency, |
| 2265 | lora_modules=args.lora_modules, |
| 2266 | lora_assignment=args.lora_assignment, |
| 2267 | extra_headers=headers, |
| 2268 | extra_body=extra_body, |
| 2269 | ramp_up_strategy=args.ramp_up_strategy, |
| 2270 | ramp_up_start_rps=args.ramp_up_start_rps, |
| 2271 | ramp_up_end_rps=args.ramp_up_end_rps, |
| 2272 | ready_check_timeout_sec=args.ready_check_timeout_sec, |
| 2273 | ssl_context=ssl_context, |
| 2274 | self_timed=args.self_timed, |
| 2275 | probe_request_rate=args.probe_request_rate, |
| 2276 | ) |
| 2277 | |
| 2278 | # Save config and results to json |
| 2279 | result_json: dict[str, Any] = {} |
| 2280 | |
| 2281 | # Setup |
| 2282 | current_dt = datetime.now().strftime("%Y%m%d-%H%M%S") |
| 2283 | result_json["date"] = current_dt |
| 2284 | result_json["endpoint_type"] = args.backend # for backward compatibility |
| 2285 | result_json["backend"] = args.backend |
| 2286 | result_json["label"] = label |
| 2287 | result_json["model_id"] = model_id |
| 2288 | result_json["tokenizer_id"] = tokenizer_id |
| 2289 | result_json["num_prompts"] = args.num_prompts |
| 2290 | |
| 2291 | # Metadata |
| 2292 | if args.metadata: |
| 2293 | for item in args.metadata: |
| 2294 | if "=" in item: |
| 2295 | kvstring = item.split("=", 1) |
| 2296 | result_json[kvstring[0].strip()] = kvstring[1].strip() |
| 2297 | else: |
| 2298 | raise ValueError( |
| 2299 | "Invalid metadata format. Please use KEY=VALUE format." |
| 2300 | ) |
| 2301 | |
| 2302 | # Traffic |
| 2303 | result_json["request_rate"] = ( |
| 2304 | args.request_rate if args.request_rate < float("inf") else "inf" |
| 2305 | ) |
| 2306 | result_json["burstiness"] = args.burstiness |
| 2307 | result_json["max_concurrency"] = args.max_concurrency |
| 2308 | |
| 2309 | if args.ramp_up_strategy is not None: |
| 2310 | result_json["ramp_up_strategy"] = args.ramp_up_strategy |
| 2311 | result_json["ramp_up_start_rps"] = args.ramp_up_start_rps |
| 2312 | result_json["ramp_up_end_rps"] = args.ramp_up_end_rps |
| 2313 | |
| 2314 | # Merge with benchmark result |
| 2315 | result_json = {**result_json, **benchmark_result} |
| 2316 | |
| 2317 | # Compute file_name once before using it for plots or saving results |
| 2318 | file_name = compute_result_filename(args, model_id, label, current_dt) |
| 2319 | |
| 2320 | # Generate timeline plot if requested |
| 2321 | if args.plot_timeline: |
| 2322 | assert file_name is not None, ( |
| 2323 | "file_name must be set when plot_timeline is enabled" |
| 2324 | ) |
| 2325 | try: |
| 2326 | from vllm.benchmarks.plot import generate_timeline_plot |
| 2327 | |
| 2328 | # Prepare per-request data for timeline |
| 2329 | per_request_data = [] |
| 2330 | start_times = benchmark_result.get("start_times", []) |
| 2331 | ttfts = benchmark_result.get("ttfts", []) |
| 2332 | itls = benchmark_result.get("itls", []) |
| 2333 | input_lens = benchmark_result.get("input_lens", []) |
| 2334 | output_lens = benchmark_result.get("output_lens", []) |
| 2335 | |
| 2336 | if start_times and ttfts and itls: |
| 2337 | for i in range(len(start_times)): |
| 2338 | # Calculate latency as ttft + sum of all itls |
| 2339 | latency = ttfts[i] + sum(itls[i]) if itls[i] else ttfts[i] |
| 2340 | |
| 2341 | per_request_data.append( |
| 2342 | { |
| 2343 | "start_time": start_times[i], |
| 2344 | "ttft": ttfts[i], |
| 2345 | "itl": itls[i], |
| 2346 | "latency": latency, |
| 2347 | "prompt_len": input_lens[i], |
| 2348 | "output_tokens": output_lens[i], |
| 2349 | } |
| 2350 | ) |
| 2351 | |
| 2352 | timeline_path = Path(file_name).with_suffix(".timeline.html") |
| 2353 | # Convert thresholds from milliseconds to seconds |
| 2354 | itl_thresholds_sec = [ |
| 2355 | float(t) / 1000.0 for t in args.timeline_itl_thresholds.split(",") |
| 2356 | ] |
| 2357 | generate_timeline_plot( |
| 2358 | per_request_data, timeline_path, itl_thresholds=itl_thresholds_sec |
| 2359 | ) |
| 2360 | else: |
| 2361 | warnings.warn( |
| 2362 | "Timeline plot requires detailed metrics. " |
| 2363 | "Ensure the benchmark completed successfully.", |
| 2364 | stacklevel=2, |
| 2365 | ) |
| 2366 | except Exception as e: |
| 2367 | warnings.warn(f"Failed to generate timeline plot: {e}", stacklevel=2) |
| 2368 | |
| 2369 | # Generate dataset statistics plot if requested |
| 2370 | if args.plot_dataset_stats: |
| 2371 | assert file_name is not None, ( |
| 2372 | "file_name must be set when plot_dataset_stats is enabled" |
| 2373 | ) |
| 2374 | try: |
| 2375 | from vllm.benchmarks.plot import generate_dataset_stats_plot |
| 2376 | |
| 2377 | # Prepare per-request data for dataset stats |
| 2378 | per_request_data = [] |
| 2379 | input_lens = benchmark_result.get("input_lens", []) |
| 2380 | output_lens = benchmark_result.get("output_lens", []) |
| 2381 | |
| 2382 | if input_lens and output_lens: |
| 2383 | for req_input_len, req_output_len in zip(input_lens, output_lens): |
| 2384 | per_request_data.append( |
| 2385 | { |
| 2386 | "prompt_len": req_input_len, |
| 2387 | "output_tokens": req_output_len, |
| 2388 | } |
| 2389 | ) |
| 2390 | |
| 2391 | stats_path = Path(file_name).with_suffix(".dataset_stats.png") |
| 2392 | generate_dataset_stats_plot(per_request_data, stats_path) |
| 2393 | else: |
| 2394 | warnings.warn( |
| 2395 | "Dataset statistics plot requires input and " |
| 2396 | "output length data. Ensure the benchmark completed " |
| 2397 | "successfully.", |
| 2398 | stacklevel=2, |
| 2399 | ) |
| 2400 | except Exception as e: |
| 2401 | warnings.warn( |
| 2402 | f"Failed to generate dataset statistics plot: {e}", stacklevel=2 |
| 2403 | ) |
| 2404 | |
| 2405 | if not args.save_detailed: |
| 2406 | # Remove fields with too many data points |
| 2407 | for field in [ |
| 2408 | "input_lens", |
| 2409 | "output_lens", |
| 2410 | "start_times", |
| 2411 | "ttfts", |
| 2412 | "itls", |
| 2413 | "generated_texts", |
| 2414 | "errors", |
| 2415 | ]: |
| 2416 | if field in result_json: |
| 2417 | del result_json[field] |
| 2418 | if field in benchmark_result: |
| 2419 | del benchmark_result[field] |
| 2420 | |
| 2421 | # Save to file |
| 2422 | if args.save_result or args.append_result: |
| 2423 | assert file_name is not None, ( |
| 2424 | "file_name must be set when save_result or append_result is enabled" |
| 2425 | ) |
| 2426 | with open( |
| 2427 | file_name, mode="a+" if args.append_result else "w", encoding="utf-8" |
| 2428 | ) as outfile: |
| 2429 | # Append a newline. |
| 2430 | if args.append_result and outfile.tell() != 0: |
| 2431 | outfile.write("\n") |
| 2432 | json.dump(result_json, outfile) |
| 2433 | save_to_pytorch_benchmark_format(args, result_json, file_name) |
| 2434 | |
| 2435 | return result_json |