serve.py 65.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
r"""Benchmark online serving throughput.

On the server side, run one of the following commands
to launch the vLLM OpenAI API server:
7
    vllm serve <your_model> <engine arguments>
8
9
10

On the client side, run:
    vllm bench serve \
11
12
        --backend <backend or endpoint type. Default 'openai'> \
        --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
16
        --input-len <general input length. Optional, maps to dataset-specific args> \
        --output-len <general output length. Optional, maps to dataset-specific args> \
17
18
19
        --request-rate <request_rate. Default inf> \
        --num-prompts <num_prompts. Default 1000>
"""
20

21
22
import argparse
import asyncio
23
import contextlib
24
import importlib.util
25
26
27
import json
import os
import random
28
import shutil
29
import ssl
30
import time
31
import uuid
32
import warnings
33
from collections.abc import AsyncGenerator, Iterable
34
35
from dataclasses import dataclass
from datetime import datetime
36
from enum import Enum
37
from typing import Any, Literal
38

39
import aiohttp
40
41
42
import numpy as np
from tqdm.asyncio import tqdm

43
from vllm.benchmarks.datasets import SampleRequest, add_dataset_parser, get_samples
44
from vllm.benchmarks.lib.endpoint_request_func import (
45
46
47
48
49
    ASYNC_REQUEST_FUNCS,
    OPENAI_COMPATIBLE_BACKENDS,
    RequestFuncInput,
    RequestFuncOutput,
)
50
from vllm.benchmarks.lib.ready_checker import wait_for_endpoint
51
from vllm.benchmarks.lib.utils import convert_to_pytorch_benchmark_format, write_to_json
52
from vllm.tokenizers import TokenizerLike, get_tokenizer
53
from vllm.utils.gc_utils import freeze_gc_heap
54
from vllm.utils.network_utils import join_host_port
55
56
57

MILLISECONDS_TO_SECONDS_CONVERSION = 1000

58
59
60
TERM_PLOTLIB_AVAILABLE = (importlib.util.find_spec("termplotlib") is not None) and (
    shutil.which("gnuplot") is not None
)
61

62

63
async def get_first_model_from_server(
64
65
66
    base_url: str,
    headers: dict | None = None,
    ssl_context: ssl.SSLContext | bool | None = None,
67
) -> tuple[str, str]:
68
69
    """Fetch the first model from the server's /v1/models endpoint."""
    models_url = f"{base_url}/v1/models"
70
71
    connector = aiohttp.TCPConnector(ssl=ssl_context)
    async with aiohttp.ClientSession(connector=connector) as session:
72
73
74
75
76
        try:
            async with session.get(models_url, headers=headers) as response:
                response.raise_for_status()
                data = await response.json()
                if "data" in data and len(data["data"]) > 0:
77
                    return data["data"][0]["id"], data["data"][0]["root"]
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
                else:
                    raise ValueError(
                        f"No models found on the server at {base_url}. "
                        "Make sure the server is running and has models loaded."
                    )
        except (aiohttp.ClientError, json.JSONDecodeError) as e:
            raise RuntimeError(
                f"Failed to fetch models from server at {models_url}. "
                "Check that:\n"
                "1. The server is running\n"
                "2. The server URL is correct\n"
                f"Error: {e}"
            ) from e


93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
@dataclass
class SpecDecodeMetrics:
    """Speculative decoding metrics from the server's Prometheus endpoint."""

    num_drafts: int
    num_draft_tokens: int
    num_accepted_tokens: int
    accepted_per_pos: dict[int, int]


async def fetch_spec_decode_metrics(
    base_url: str, session: aiohttp.ClientSession
) -> SpecDecodeMetrics | None:
    """Fetch speculative decoding metrics from the server's Prometheus endpoint.

    Returns None if speculative decoding is not enabled or metrics are not available.
    """
    metrics_url = f"{base_url}/metrics"
    try:
        async with session.get(metrics_url) as response:
            if response.status != 200:
                return None
            text = await response.text()

            num_drafts = 0
            num_draft_tokens = 0
            num_accepted_tokens = 0
            accepted_per_pos: dict[int, int] = {}
            found_spec_decode = False

            for line in text.split("\n"):
                line = line.strip()
                if not line or line.startswith("#"):
                    continue

                if line.startswith("vllm:spec_decode"):
                    found_spec_decode = True
                    parts = line.split()
                    if parts:
                        with contextlib.suppress(ValueError):
                            if "num_drafts" in line:
                                num_drafts += int(float(parts[-1]))
                            elif "num_draft_tokens" in line:
                                num_draft_tokens += int(float(parts[-1]))
                            elif "num_accepted_tokens_per_pos" in line:
                                pos_label = 'position="'
                                if pos_label in line:
                                    start = line.index(pos_label) + len(pos_label)
                                    end = line.index('"', start)
                                    pos = int(line[start:end])
                                    val = int(float(parts[-1]))
                                    accepted_per_pos[pos] = (
                                        accepted_per_pos.get(pos, 0) + val
                                    )
                            elif "num_accepted_tokens" in line:
                                num_accepted_tokens += int(float(parts[-1]))

            if not found_spec_decode:
                return None

            return SpecDecodeMetrics(
                num_drafts=num_drafts,
                num_draft_tokens=num_draft_tokens,
                num_accepted_tokens=num_accepted_tokens,
                accepted_per_pos=accepted_per_pos,
            )
    except (aiohttp.ClientError, asyncio.TimeoutError):
        return None


163
164
class TaskType(Enum):
    GENERATION = "generation"
165
    POOLING = "pooling"
166
167


168
169
170
@dataclass
class BenchmarkMetrics:
    completed: int
171
    failed: int
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
    total_input: int
    total_output: int
    request_throughput: float
    request_goodput: float
    output_throughput: float
    total_token_throughput: float
    mean_ttft_ms: float
    median_ttft_ms: float
    std_ttft_ms: float
    percentiles_ttft_ms: list[tuple[float, float]]
    mean_tpot_ms: float
    median_tpot_ms: float
    std_tpot_ms: float
    percentiles_tpot_ms: list[tuple[float, float]]
    mean_itl_ms: float
    median_itl_ms: float
    std_itl_ms: float
    percentiles_itl_ms: list[tuple[float, float]]
    # E2EL stands for end-to-end latency per request.
    # It is the time taken on the client side from sending
    # a request to receiving a complete response.
    mean_e2el_ms: float
    median_e2el_ms: float
    std_e2el_ms: float
    percentiles_e2el_ms: list[tuple[float, float]]
197
198
199
    # Max output tokens per second and concurrent requests at that peak
    max_output_tokens_per_s: float
    max_concurrent_requests: int
200
    rtfx: float = 0.0  # Inverse Real-Time Factor for ASR benchmarks
201

202

203
204
205
@dataclass
class EmbedBenchmarkMetrics:
    completed: int
206
    failed: int
207
208
    total_input: int
    request_throughput: float
209
    total_token_throughput: float
210
211
212
213
    mean_e2el_ms: float
    std_e2el_ms: float
    median_e2el_ms: float
    percentiles_e2el_ms: float
214

215

216
def _get_current_request_rate(
217
218
219
    ramp_up_strategy: Literal["linear", "exponential"] | None,
    ramp_up_start_rps: int | None,
    ramp_up_end_rps: int | None,
220
221
222
223
    request_index: int,
    total_requests: int,
    request_rate: float,
) -> float:
224
225
226
227
228
    if (
        ramp_up_strategy
        and ramp_up_start_rps is not None
        and ramp_up_end_rps is not None
    ):
229
230
231
232
233
234
235
236
237
238
239
240
        progress = request_index / max(total_requests - 1, 1)
        if ramp_up_strategy == "linear":
            increase = (ramp_up_end_rps - ramp_up_start_rps) * progress
            return ramp_up_start_rps + increase
        elif ramp_up_strategy == "exponential":
            ratio = ramp_up_end_rps / ramp_up_start_rps
            return ramp_up_start_rps * (ratio**progress)
        else:
            raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}")
    return request_rate


241
async def get_request(
242
    input_requests: list[SampleRequest],
243
244
    request_rate: float,
    burstiness: float = 1.0,
245
246
247
    ramp_up_strategy: Literal["linear", "exponential"] | None = None,
    ramp_up_start_rps: int | None = None,
    ramp_up_end_rps: int | None = None,
248
) -> AsyncGenerator[tuple[SampleRequest, float], None]:
249
250
    """
    Asynchronously generates requests at a specified rate
251
    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
252
253
254

    Args:
        input_requests:
255
            A list of input requests, each represented as a SampleRequest.
256
257
258
259
260
261
262
263
264
265
        request_rate:
            The rate at which requests are generated (requests/s).
        burstiness (optional):
            The burstiness factor of the request generation.
            Only takes effect when request_rate is not inf.
            Default value is 1, which follows a Poisson process.
            Otherwise, the request intervals follow a gamma distribution.
            A lower burstiness value (0 < burstiness < 1) results
            in more bursty requests, while a higher burstiness value
            (burstiness > 1) results in a more uniform arrival of requests.
266
        ramp_up_strategy (optional):
267
268
269
270
271
272
            The ramp-up strategy. Can be "linear" or "exponential".
            If None, uses constant request rate (specified by request_rate).
        ramp_up_start_rps (optional):
            The starting request rate for ramp-up.
        ramp_up_end_rps (optional):
            The ending request rate for ramp-up.
273
274
    """
    assert burstiness > 0, (
275
276
        f"A positive burstiness factor is expected, but given {burstiness}."
    )
277
    # Convert to list to get length for ramp-up calculations
278
    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
279
        input_requests = list(input_requests)
280

281
    total_requests = len(input_requests)
282
    assert total_requests > 0, "No requests provided."
283

284
285
286
287
    # Precompute delays among requests to minimize request send laggings
    request_rates = []
    delay_ts = []
    for request_index, request in enumerate(input_requests):
288
        current_request_rate = _get_current_request_rate(
289
290
291
292
293
294
295
            ramp_up_strategy,
            ramp_up_start_rps,
            ramp_up_end_rps,
            request_index,
            total_requests,
            request_rate,
        )
296
297
298
        assert current_request_rate > 0.0, (
            f"Obtained non-positive request rate {current_request_rate}."
        )
299
        request_rates.append(current_request_rate)
300
        if current_request_rate == float("inf"):
301
            delay_ts.append(0)
302
303
304
305
        elif burstiness == float("inf"):
            # when burstiness tends to infinity, the delay time becomes constant
            # and tends to the inverse of the request rate
            delay_ts.append(1.0 / current_request_rate)
306
307
308
309
310
311
        else:
            theta = 1.0 / (current_request_rate * burstiness)

            # Sample the request interval from the gamma distribution.
            # If burstiness is 1, it follows exponential distribution.
            delay_ts.append(np.random.gamma(shape=burstiness, scale=theta))
312

313
314
315
316
317
318
319
320
321
    # Calculate the cumulative delay time from the first sent out requests.
    for i in range(1, len(delay_ts)):
        delay_ts[i] += delay_ts[i - 1]
    if ramp_up_strategy is None and delay_ts[-1] != 0:
        # When ramp_up_strategy is not set, we assume the request rate is fixed
        # and all requests should be sent in target_total_delay_s, the following
        # logic would re-scale delay time to ensure the final delay_ts
        # align with target_total_delay_s.
        #
322
323
        # NOTE: If we simply accumulate the random delta values
        # from the gamma distribution, their sum would have 1-2% gap
324
        # from target_total_delay_s. The purpose of the following logic is to
co63oc's avatar
co63oc committed
325
        # close the gap for stabilizing the throughput data
326
        # from different random seeds.
327
328
329
330
331
332
        target_total_delay_s = total_requests / request_rate
        normalize_factor = target_total_delay_s / delay_ts[-1]
        delay_ts = [delay * normalize_factor for delay in delay_ts]

    start_ts = time.time()
    for request_index, request in enumerate(input_requests):
333
334
335
336
337
        if delay_ts[request_index] > 0:
            current_ts = time.time()
            sleep_interval_s = start_ts + delay_ts[request_index] - current_ts
            if sleep_interval_s > 0:
                await asyncio.sleep(sleep_interval_s)
338
        yield request, request_rates[request_index]
339
340


341
def calculate_metrics_for_embeddings(
342
343
344
    outputs: list[RequestFuncOutput],
    dur_s: float,
    selected_percentiles: list[float],
345
) -> EmbedBenchmarkMetrics:
346
347
348
349
350
351
352
353
354
355
356
357
    """Calculate the metrics for the embedding requests.

    Args:
        outputs: The outputs of the requests.
        dur_s: The duration of the benchmark.
        selected_percentiles: The percentiles to select.

    Returns:
        The calculated benchmark metrics.
    """
    total_input = 0
    completed = 0
358
    failed = 0
359
360
361
362
363
364
    e2els: list[float] = []
    for i in range(len(outputs)):
        if outputs[i].success:
            e2els.append(outputs[i].latency)
            completed += 1
            total_input += outputs[i].prompt_len
365
366
        else:
            failed += 1
367
368
369
370
371

    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
            "on the benchmark arguments.",
372
373
            stacklevel=2,
        )
374
375
    metrics = EmbedBenchmarkMetrics(
        completed=completed,
376
        failed=failed,
377
378
379
380
381
382
        total_input=total_input,
        request_throughput=completed / dur_s,
        total_token_throughput=total_input / dur_s,
        mean_e2el_ms=np.mean(e2els or 0) * 1000,
        std_e2el_ms=np.std(e2els or 0) * 1000,
        median_e2el_ms=np.median(e2els or 0) * 1000,
383
384
385
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
386
387
388
389
    )
    return metrics


390
def calculate_metrics(
391
    input_requests: list[SampleRequest],
392
393
    outputs: list[RequestFuncOutput],
    dur_s: float,
394
    tokenizer: TokenizerLike,
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
    selected_percentiles: list[float],
    goodput_config_dict: dict[str, float],
) -> tuple[BenchmarkMetrics, list[int]]:
    """Calculate the metrics for the benchmark.

    Args:
        input_requests: The input requests.
        outputs: The outputs of the requests.
        dur_s: The duration of the benchmark.
        tokenizer: The tokenizer to use.
        selected_percentiles: The percentiles to select.
        goodput_config_dict: The goodput configuration.

    Returns:
        A tuple of the benchmark metrics and the actual output lengths.
    """
    actual_output_lens: list[int] = []
    total_input = 0
    completed = 0
    good_completed = 0
    itls: list[float] = []
    tpots: list[float] = []
    all_tpots: list[float] = []
    ttfts: list[float] = []
    e2els: list[float] = []
420
    input_audio_duration = 0.0
421
422
423
424
    for i in range(len(outputs)):
        if outputs[i].success:
            output_len = outputs[i].output_tokens

425
            if not output_len:
426
427
428
429
430
431
432
433
434
435
436
437
438
                if tokenizer is None:
                    output_len = 1
                else:
                    # We use the tokenizer to count the number of output tokens
                    # for some serving backends instead of looking at
                    # len(outputs[i].itl) since multiple output tokens may be
                    # bundled together
                    # Note : this may inflate the output token count slightly
                    output_len = len(
                        tokenizer(
                            outputs[i].generated_text, add_special_tokens=False
                        ).input_ids
                    )
439
            actual_output_lens.append(output_len)
440
            total_input += input_requests[i].prompt_len
441
442
443
444
445
446
447
448
449
450
            tpot = 0
            if output_len > 1:
                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
                tpot = latency_minus_ttft / (output_len - 1)
                tpots.append(tpot)
            # Note: if output_len <= 1, we regard tpot as 0 for goodput
            all_tpots.append(tpot)
            itls += outputs[i].itl
            ttfts.append(outputs[i].ttft)
            e2els.append(outputs[i].latency)
451
            input_audio_duration += outputs[i].input_audio_duration
452
453
454
455
456
457
458
459
460
461
            completed += 1
        else:
            actual_output_lens.append(0)

    if goodput_config_dict:
        valid_metrics = []
        slo_values = []

        if "ttft" in goodput_config_dict:
            valid_metrics.append(ttfts)
462
463
464
            slo_values.append(
                goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
465
466
        if "tpot" in goodput_config_dict:
            valid_metrics.append(all_tpots)
467
468
469
            slo_values.append(
                goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
470
471
        if "e2el" in goodput_config_dict:
            valid_metrics.append(e2els)
472
473
474
            slo_values.append(
                goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
475
476
477
478
479
480
481
482
483
484

        for req_metric in zip(*valid_metrics):
            is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)])
            if is_good_req:
                good_completed += 1

    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
            "on the benchmark arguments.",
485
486
            stacklevel=2,
        )
487
488
489
490
491
492
493

    # Calculate max output tokens per second metric
    max_output_tokens_per_s = 0.0
    max_concurrent_requests = 0

    # Find the time range across all successful requests
    successful_outputs = [output for output in outputs if output.success]
494
    failed_outputs = [output for output in outputs if not output.success]
495
496
497
498
499
500

    if len(failed_outputs) > 0:
        print("Failed requests during benchmark run detected (capping to 10):")
        for i, err in enumerate(failed_outputs[:10]):
            print(f"Error {i}: {err.error}")

501
    if successful_outputs:
502
503
504
505
        min_start_time = min(output.start_time for output in successful_outputs)
        max_end_time = max(
            output.start_time + output.latency for output in successful_outputs
        )
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528

        # Create second buckets (ceiling to ensure we capture all time)
        duration_seconds = int(np.ceil(max_end_time - min_start_time)) + 1
        tokens_per_second = np.zeros(duration_seconds)
        concurrent_requests_per_second = np.zeros(duration_seconds)

        for i, output in enumerate(successful_outputs):
            # Calculate token generation timestamp using
            # start_time, ttft, and itl
            token_times = [output.start_time + output.ttft]
            current_time = token_times[0]
            for itl_value in output.itl:
                current_time += itl_value
                token_times.append(current_time)

            # Add tokens to second buckets
            for token_time in token_times:
                second_bucket = int(token_time - min_start_time)
                if 0 <= second_bucket < duration_seconds:
                    tokens_per_second[second_bucket] += 1

            # Track concurrent requests for each second this request was active
            request_start_second = int(output.start_time - min_start_time)
529
530
531
            request_end_second = int(
                (output.start_time + output.latency) - min_start_time
            )
532
533
534
535
536
537
538
539

            for second in range(request_start_second, request_end_second + 1):
                concurrent_requests_per_second[second] += 1

        # Find the maximum tokens per second and corresponding
        # concurrent requests
        if len(tokens_per_second) > 0:
            max_output_tokens_per_s = float(np.max(tokens_per_second))
540
            max_concurrent_requests = int(np.max(concurrent_requests_per_second))
541
542
543

        if TERM_PLOTLIB_AVAILABLE:
            import termplotlib as tpl
544

545
            fig = tpl.figure()
546
547
548
549
550
551
552
553
554
555
            fig.plot(
                np.arange(len(tokens_per_second)),
                tokens_per_second,
                title="Output tokens per second",
            )
            fig.plot(
                np.arange(len(concurrent_requests_per_second)),
                concurrent_requests_per_second,
                title="Concurrent requests per second",
            )
556
557
558
559
            fig.show()
        else:
            print("tip: install termplotlib and gnuplot to plot the metrics")

560
561
    metrics = BenchmarkMetrics(
        completed=completed,
562
        failed=len(failed_outputs),
563
564
565
566
567
568
        total_input=total_input,
        total_output=sum(actual_output_lens),
        request_throughput=completed / dur_s,
        request_goodput=good_completed / dur_s,
        output_throughput=sum(actual_output_lens) / dur_s,
        total_token_throughput=(total_input + sum(actual_output_lens)) / dur_s,
569
570
        mean_ttft_ms=np.mean(ttfts or 0)
        * 1000,  # ttfts is empty if streaming is not supported by the endpoint
571
572
        std_ttft_ms=np.std(ttfts or 0) * 1000,
        median_ttft_ms=np.median(ttfts or 0) * 1000,
573
574
575
        percentiles_ttft_ms=[
            (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles
        ],
576
577
578
        mean_tpot_ms=np.mean(tpots or 0) * 1000,
        std_tpot_ms=np.std(tpots or 0) * 1000,
        median_tpot_ms=np.median(tpots or 0) * 1000,
579
580
581
        percentiles_tpot_ms=[
            (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles
        ],
582
583
584
        mean_itl_ms=np.mean(itls or 0) * 1000,
        std_itl_ms=np.std(itls or 0) * 1000,
        median_itl_ms=np.median(itls or 0) * 1000,
585
586
587
        percentiles_itl_ms=[
            (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles
        ],
588
589
590
        mean_e2el_ms=np.mean(e2els or 0) * 1000,
        std_e2el_ms=np.std(e2els or 0) * 1000,
        median_e2el_ms=np.median(e2els or 0) * 1000,
591
592
593
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
594
595
        max_output_tokens_per_s=max_output_tokens_per_s,
        max_concurrent_requests=max_concurrent_requests,
596
        rtfx=input_audio_duration / dur_s,
597
598
599
600
601
602
    )

    return metrics, actual_output_lens


async def benchmark(
603
    task_type: TaskType,
604
605
606
607
608
    endpoint_type: str,
    api_url: str,
    base_url: str,
    model_id: str,
    model_name: str,
609
    tokenizer: TokenizerLike,
610
    input_requests: list[SampleRequest],
611
    logprobs: int | None,
612
613
614
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
615
    num_warmups: int,
616
617
    profile: bool,
    selected_percentile_metrics: list[str],
618
    selected_percentiles: list[float],
619
620
    ignore_eos: bool,
    goodput_config_dict: dict[str, float],
621
622
623
624
625
626
627
    max_concurrency: int | None,
    lora_modules: Iterable[str] | None,
    extra_headers: dict | None,
    extra_body: dict | None,
    ramp_up_strategy: Literal["linear", "exponential"] | None = None,
    ramp_up_start_rps: int | None = None,
    ramp_up_end_rps: int | None = None,
628
    ready_check_timeout_sec: int = 600,
629
    ssl_context: ssl.SSLContext | bool | None = None,
630
):
631
632
633
634
    try:
        request_func = ASYNC_REQUEST_FUNCS[endpoint_type]
    except KeyError:
        raise ValueError(f"Unknown backend: {endpoint_type}") from None
635

636
    # Reuses connections across requests to reduce TLS handshake overhead.
637
638
    # Use ssl_context if provided, otherwise default to True for https URLs
    ssl_setting = ssl_context if ssl_context is not None else ("https://" in api_url)
639
640
641
642
643
644
645
646
    connector = aiohttp.TCPConnector(
        limit=max_concurrency or 0,
        limit_per_host=max_concurrency or 0,
        ttl_dns_cache=300,
        use_dns_cache=True,
        keepalive_timeout=60,
        enable_cleanup_closed=True,
        force_close=False,
647
        ssl=ssl_setting,
648
649
650
651
652
653
654
655
    )

    session = aiohttp.ClientSession(
        connector=connector,
        trust_env=True,
        timeout=aiohttp.ClientTimeout(total=6 * 60 * 60),
    )

656
657
    print("Starting initial single prompt test run...")
    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
658
659
660
661
662
663
        input_requests[0].prompt,
        input_requests[0].prompt_len,
        input_requests[0].expected_output_len,
        input_requests[0].multi_modal_data,
    )

664
665
666
667
668
669
670
671
    assert (
        test_mm_content is None
        or isinstance(test_mm_content, dict)
        or (
            isinstance(test_mm_content, list)
            and all(isinstance(item, dict) for item in test_mm_content)
        )
    ), "multi_modal_data must be a dict or list[dict]"
672
673
674
675
676
677
678
679
680
681
    test_input = RequestFuncInput(
        model=model_id,
        model_name=model_name,
        prompt=test_prompt,
        api_url=api_url,
        prompt_len=test_prompt_len,
        output_len=test_output_len,
        logprobs=logprobs,
        multi_modal_content=test_mm_content,
        ignore_eos=ignore_eos,
682
        extra_headers=extra_headers,
683
        extra_body=extra_body,
684
685
    )

686
687
688
689
690
691
692
693
694
695
696
    if ready_check_timeout_sec > 0:
        test_output = await wait_for_endpoint(
            request_func,
            test_input,
            session,
            timeout_seconds=ready_check_timeout_sec,
        )
        if not test_output.success:
            raise ValueError(
                "Initial test run failed - Please make sure benchmark "
                "arguments are correctly specified. "
697
698
                f"Error: {test_output.error}"
            )
699
        else:
700
            print("Initial test run completed.")
701
    else:
702
        print("Skipping endpoint ready check.")
703

704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
    if num_warmups > 0:
        print(f"Warming up with {num_warmups} requests...")
        warmup_pbar = None if disable_tqdm else tqdm(total=num_warmups)
        warmup_semaphore = (
            asyncio.Semaphore(max_concurrency)
            if max_concurrency
            else contextlib.nullcontext()
        )
        warmup_tasks = []

        async def warmup_limited_request_func():
            async with warmup_semaphore:
                return await request_func(
                    request_func_input=test_input, session=session, pbar=warmup_pbar
                )

        for _ in range(num_warmups):
            request_task = asyncio.create_task(warmup_limited_request_func())
            warmup_tasks.append(request_task)
        _ = await asyncio.gather(*warmup_tasks)

        if warmup_pbar is not None:
            warmup_pbar.close()
        print("Warmup run completed.")

    print("Starting main benchmark run...")

731
732
733
    if lora_modules:
        # For each input request, choose a LoRA module at random.
        lora_modules = iter(
734
735
            [random.choice(lora_modules) for _ in range(len(input_requests))]
        )
736
737
738

    if profile:
        print("Starting profiler...")
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
        profile_input = RequestFuncInput(
            model=model_id,
            model_name=model_name,
            prompt=test_prompt,
            api_url=base_url + "/start_profile",
            prompt_len=test_prompt_len,
            output_len=test_output_len,
            logprobs=logprobs,
            multi_modal_content=test_mm_content,
            ignore_eos=ignore_eos,
            extra_headers=extra_headers,
            extra_body=extra_body,
        )
        profile_output = await request_func(
            request_func_input=profile_input, session=session
        )
755
756
757
        if profile_output.success:
            print("Profiler started")

758
    distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
759
760
761

    if ramp_up_strategy is not None:
        print(f"Traffic ramp-up strategy: {ramp_up_strategy}.")
762
763
764
765
        print(
            f"Will increase RPS from {ramp_up_start_rps} to "
            f"{ramp_up_end_rps} RPS over the duration of the benchmark."
        )
766
    else:
767
        print(f"Traffic request rate: {request_rate}")
768
769
770
771

    print(f"Burstiness factor: {burstiness} ({distribution})")
    print(f"Maximum request concurrency: {max_concurrency}")

772
773
    spec_decode_metrics_before = await fetch_spec_decode_metrics(base_url, session)

774
775
    pbar = None if disable_tqdm else tqdm(total=len(input_requests))

776
777
778
779
780
    semaphore = (
        asyncio.Semaphore(max_concurrency)
        if max_concurrency
        else contextlib.nullcontext()
    )
781

782
    async def limited_request_func(request_func_input, session, pbar):
783
        async with semaphore:
784
785
786
            return await request_func(
                request_func_input=request_func_input, session=session, pbar=pbar
            )
787
788
789

    benchmark_start_time = time.perf_counter()
    tasks: list[asyncio.Task] = []
790
791
792
793
794

    rps_change_events = []
    last_int_rps = -1
    if ramp_up_strategy is not None and ramp_up_start_rps is not None:
        last_int_rps = ramp_up_start_rps
795
796
797
798
799
800
        rps_change_events.append(
            {
                "rps": last_int_rps,
                "timestamp": datetime.now().isoformat(),
            }
        )
801
802

    async for request, current_request_rate in get_request(
803
804
805
806
807
808
809
        input_requests,
        request_rate,
        burstiness,
        ramp_up_strategy,
        ramp_up_start_rps,
        ramp_up_end_rps,
    ):
810
811
812
813
814
        if ramp_up_strategy is not None:
            current_int_rps = int(current_request_rate)
            if current_int_rps > last_int_rps:
                timestamp = datetime.now().isoformat()
                for rps_val in range(last_int_rps + 1, current_int_rps + 1):
815
                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
816
                last_int_rps = current_int_rps
817
        prompt, prompt_len, output_len, mm_content, request_id = (
818
819
820
821
            request.prompt,
            request.prompt_len,
            request.expected_output_len,
            request.multi_modal_data,
822
            request.request_id,
823
        )
824
825
826
827
828
        req_model_id, req_model_name = model_id, model_name
        if lora_modules:
            req_lora_module = next(lora_modules)
            req_model_id, req_model_name = req_lora_module, req_lora_module

829
830
831
832
833
834
835
836
837
838
839
840
841
842
        request_func_input = RequestFuncInput(
            model=req_model_id,
            model_name=req_model_name,
            prompt=prompt,
            api_url=api_url,
            prompt_len=prompt_len,
            output_len=output_len,
            logprobs=logprobs,
            multi_modal_content=mm_content,
            ignore_eos=ignore_eos,
            extra_headers=extra_headers,
            extra_body=extra_body,
            request_id=request_id,
        )
843
844
        tasks.append(
            asyncio.create_task(
845
846
847
848
849
                limited_request_func(
                    request_func_input=request_func_input, session=session, pbar=pbar
                )
            )
        )
850
851
852
853
854
855
856
    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)

    if pbar is not None:
        pbar.close()

    benchmark_duration = time.perf_counter() - benchmark_start_time

857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
    spec_decode_metrics_after = await fetch_spec_decode_metrics(base_url, session)
    spec_decode_stats: dict[str, Any] | None = None
    if spec_decode_metrics_before is not None and spec_decode_metrics_after is not None:
        delta_drafts = (
            spec_decode_metrics_after.num_drafts - spec_decode_metrics_before.num_drafts
        )
        delta_draft_tokens = (
            spec_decode_metrics_after.num_draft_tokens
            - spec_decode_metrics_before.num_draft_tokens
        )
        delta_accepted = (
            spec_decode_metrics_after.num_accepted_tokens
            - spec_decode_metrics_before.num_accepted_tokens
        )
        per_pos_rates: list[float] = []
        if delta_drafts > 0:
            positions = sorted(
                set(spec_decode_metrics_before.accepted_per_pos.keys())
                | set(spec_decode_metrics_after.accepted_per_pos.keys())
            )
            for pos in positions:
                before_val = spec_decode_metrics_before.accepted_per_pos.get(pos, 0)
                after_val = spec_decode_metrics_after.accepted_per_pos.get(
                    pos, before_val
                )
                delta_pos = after_val - before_val
                per_pos_rates.append(delta_pos / delta_drafts)

        if delta_draft_tokens > 0:
            acceptance_rate = (delta_accepted / delta_draft_tokens) * 100
            acceptance_length = (
                1 + delta_accepted / delta_drafts if delta_drafts > 0 else 0.0
            )
            spec_decode_stats = {
                "num_drafts": delta_drafts,
                "draft_tokens": delta_draft_tokens,
                "accepted_tokens": delta_accepted,
                "acceptance_rate": acceptance_rate,
                "acceptance_length": acceptance_length,
                "per_position_acceptance_rates": per_pos_rates,
            }

899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
    if task_type == TaskType.GENERATION:
        metrics, actual_output_lens = calculate_metrics(
            input_requests=input_requests,
            outputs=outputs,
            dur_s=benchmark_duration,
            tokenizer=tokenizer,
            selected_percentiles=selected_percentiles,
            goodput_config_dict=goodput_config_dict,
        )
    else:
        metrics = calculate_metrics_for_embeddings(
            outputs=outputs,
            dur_s=benchmark_duration,
            selected_percentiles=selected_percentiles,
        )
        actual_output_lens = 0
915

916
    print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
917
    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
918
    print("{:<40} {:<10}".format("Failed requests:", metrics.failed))
919
    if max_concurrency is not None:
920
921
922
923
        print("{:<40} {:<10}".format("Maximum request concurrency:", max_concurrency))
    if request_rate != float("inf"):
        print("{:<40} {:<10.2f}".format("Request rate configured (RPS):", request_rate))
    print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
924
    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
925
    if isinstance(metrics, BenchmarkMetrics) and tokenizer:
926
927
928
929
930
931
        print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
    print(
        "{:<40} {:<10.2f}".format(
            "Request throughput (req/s):", metrics.request_throughput
        )
    )
932
    if goodput_config_dict:
933
934
935
936
937
        print(
            "{:<40} {:<10.2f}".format(
                "Request goodput (req/s):", metrics.request_goodput
            )
        )
938
    if isinstance(metrics, BenchmarkMetrics):
939
940
941
942
943
        if tokenizer:
            print(
                "{:<40} {:<10.2f}".format(
                    "Output token throughput (tok/s):", metrics.output_throughput
                )
944
            )
945
946
947
948
949
            print(
                "{:<40} {:<10.2f}".format(
                    "Peak output token throughput (tok/s):",
                    metrics.max_output_tokens_per_s,
                )
950
951
952
953
954
955
            )
        print(
            "{:<40} {:<10.2f}".format(
                "Peak concurrent requests:", metrics.max_concurrent_requests
            )
        )
956
957
958
959
960
961
        if metrics.rtfx > 0.0:
            print(
                "{:<40} {:<10.2f}".format(
                    "RTFx (Inverse Real-Time Factor):", metrics.rtfx
                )
            )
962
963
964
965
966
    if tokenizer:
        print(
            "{:<40} {:<10.2f}".format(
                "Total token throughput (tok/s):", metrics.total_token_throughput
            )
967
        )
968

969
970
971
972
    if isinstance(metrics, BenchmarkMetrics):
        result = {
            "duration": benchmark_duration,
            "completed": metrics.completed,
973
            "failed": metrics.failed,
974
975
976
            "total_input_tokens": metrics.total_input,
            "total_output_tokens": metrics.total_output,
            "request_throughput": metrics.request_throughput,
977
            "request_goodput": metrics.request_goodput if goodput_config_dict else None,
978
979
980
981
982
983
            "output_throughput": metrics.output_throughput,
            "total_token_throughput": metrics.total_token_throughput,
            "input_lens": [output.prompt_len for output in outputs],
            "output_lens": actual_output_lens,
            "ttfts": [output.ttft for output in outputs],
            "itls": [output.itl for output in outputs],
984
            "start_times": [output.start_time for output in outputs],
985
986
            "generated_texts": [output.generated_text for output in outputs],
            "errors": [output.error for output in outputs],
987
988
            "max_output_tokens_per_s": metrics.max_output_tokens_per_s,
            "max_concurrent_requests": metrics.max_concurrent_requests,
989
            "rtfx": metrics.rtfx,
990
991
992
993
994
995
996
997
998
999
1000
        }
    else:
        result = {
            "duration": benchmark_duration,
            "completed": metrics.completed,
            "total_input_tokens": metrics.total_input,
            "request_throughput": metrics.request_throughput,
            "total_token_throughput": metrics.total_token_throughput,
            "input_lens": [output.prompt_len for output in outputs],
            "errors": [output.error for output in outputs],
        }
1001

1002
1003
1004
    if rps_change_events:
        result["rps_change_events"] = rps_change_events

1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
    if spec_decode_stats is not None:
        result["spec_decode_acceptance_rate"] = spec_decode_stats["acceptance_rate"]
        result["spec_decode_acceptance_length"] = spec_decode_stats["acceptance_length"]
        result["spec_decode_num_drafts"] = int(spec_decode_stats["num_drafts"])
        result["spec_decode_draft_tokens"] = int(spec_decode_stats["draft_tokens"])
        result["spec_decode_accepted_tokens"] = int(
            spec_decode_stats["accepted_tokens"]
        )
        result["spec_decode_per_position_acceptance_rates"] = spec_decode_stats.get(
            "per_position_acceptance_rates", []
        )

1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
    def process_one_metric(
        # E.g., "ttft"
        metric_attribute_name: str,
        # E.g., "TTFT"
        metric_name: str,
        # E.g., "Time to First Token"
        metric_header: str,
    ):
        # This function prints and adds statistics of the specified
        # metric.
        if metric_attribute_name not in selected_percentile_metrics:
            return
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
        print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-"))
        print(
            "{:<40} {:<10.2f}".format(
                f"Mean {metric_name} (ms):",
                getattr(metrics, f"mean_{metric_attribute_name}_ms"),
            )
        )
        print(
            "{:<40} {:<10.2f}".format(
                f"Median {metric_name} (ms):",
                getattr(metrics, f"median_{metric_attribute_name}_ms"),
            )
        )
1042
        result[f"mean_{metric_attribute_name}_ms"] = getattr(
1043
1044
            metrics, f"mean_{metric_attribute_name}_ms"
        )
1045
        result[f"median_{metric_attribute_name}_ms"] = getattr(
1046
1047
            metrics, f"median_{metric_attribute_name}_ms"
        )
1048
        result[f"std_{metric_attribute_name}_ms"] = getattr(
1049
1050
1051
            metrics, f"std_{metric_attribute_name}_ms"
        )
        for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"):
1052
            p_word = str(int(p)) if int(p) == p else str(p)
1053
            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value))
1054
1055
            result[f"p{p_word}_{metric_attribute_name}_ms"] = value

1056
    if task_type == TaskType.GENERATION and tokenizer:
1057
        process_one_metric("ttft", "TTFT", "Time to First Token")
1058
        process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)")
1059
        process_one_metric("itl", "ITL", "Inter-token Latency")
1060
1061
    process_one_metric("e2el", "E2EL", "End-to-end Latency")

1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
    if spec_decode_stats is not None:
        print("{s:{c}^{n}}".format(s="Speculative Decoding", n=50, c="-"))
        print(
            "{:<40} {:<10.2f}".format(
                "Acceptance rate (%):", spec_decode_stats["acceptance_rate"]
            )
        )
        print(
            "{:<40} {:<10.2f}".format(
                "Acceptance length:", spec_decode_stats["acceptance_length"]
            )
        )
        print("{:<40} {:<10}".format("Drafts:", int(spec_decode_stats["num_drafts"])))
        print(
            "{:<40} {:<10}".format(
                "Draft tokens:", int(spec_decode_stats["draft_tokens"])
            )
        )
        print(
            "{:<40} {:<10}".format(
                "Accepted tokens:", int(spec_decode_stats["accepted_tokens"])
            )
        )
        per_pos = spec_decode_stats.get("per_position_acceptance_rates", [])
        if per_pos:
            print("Per-position acceptance (%):")
            for i, rate in enumerate(per_pos):
                print("{:<40} {:<10.2f}".format(f"  Position {i}:", rate * 100))

1091
1092
    print("=" * 50)

1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
    if profile:
        print("Stopping profiler...")
        profile_input = RequestFuncInput(
            model=model_id,
            prompt=test_prompt,
            api_url=base_url + "/stop_profile",
            prompt_len=test_prompt_len,
            output_len=test_output_len,
            logprobs=logprobs,
        )
1103
1104
1105
        profile_output = await request_func(
            request_func_input=profile_input, session=session
        )
1106
1107
        if profile_output.success:
            print("Profiler stopped")
1108
1109

    await session.close()
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
    return result


def check_goodput_args(args):
    # Check and parse goodput arguments
    goodput_config_dict = {}
    VALID_NAMES = ["ttft", "tpot", "e2el"]
    if args.goodput:
        goodput_config_dict = parse_goodput(args.goodput)
        for slo_name, slo_val in goodput_config_dict.items():
            if slo_name not in VALID_NAMES:
                raise ValueError(
                    f"Invalid metric name found, {slo_name}: {slo_val}. "
                    "The service level objective name should be one of "
1124
1125
                    f"{str(VALID_NAMES)}. "
                )
1126
1127
1128
1129
            if slo_val < 0:
                raise ValueError(
                    f"Invalid value found, {slo_name}: {slo_val}. "
                    "The service level objective value should be "
1130
1131
                    "non-negative."
                )
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
    return goodput_config_dict


def parse_goodput(slo_pairs):
    goodput_config_dict = {}
    try:
        for slo_pair in slo_pairs:
            slo_name, slo_val = slo_pair.split(":")
            goodput_config_dict[slo_name] = float(slo_val)
    except ValueError as err:
        raise argparse.ArgumentTypeError(
            "Invalid format found for service level objectives. "
1144
            'Specify service level objectives for goodput as "KEY:VALUE" '
1145
            "pairs, where the key is a metric name, and the value is a "
1146
1147
            "number in milliseconds."
        ) from err
1148
1149
1150
    return goodput_config_dict


1151
1152
1153
def save_to_pytorch_benchmark_format(
    args: argparse.Namespace, results: dict[str, Any], file_name: str
) -> None:
1154
    metrics = [
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
        "median_ttft_ms",
        "mean_ttft_ms",
        "std_ttft_ms",
        "p99_ttft_ms",
        "mean_tpot_ms",
        "median_tpot_ms",
        "std_tpot_ms",
        "p99_tpot_ms",
        "median_itl_ms",
        "mean_itl_ms",
        "std_itl_ms",
        "p99_itl_ms",
1167
1168
1169
1170
1171
1172
    ]
    # These raw data might be useful, but they are rather big. They can be added
    # later if needed
    ignored_metrics = ["ttfts", "itls", "generated_texts", "errors"]
    pt_records = convert_to_pytorch_benchmark_format(
        args=args,
1173
        metrics={k: [results[k]] for k in metrics if k in results},
1174
1175
        extra_info={
            k: results[k]
1176
1177
1178
1179
            for k in results
            if k not in metrics and k not in ignored_metrics
        },
    )
1180
1181
1182
1183
1184
1185
1186
    if pt_records:
        # Don't use json suffix here as we don't want CI to pick it up
        pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json"
        write_to_json(pt_file, pt_records)


def add_cli_args(parser: argparse.ArgumentParser):
1187
    add_dataset_parser(parser)
1188
1189
1190
1191
1192
    parser.add_argument(
        "--label",
        type=str,
        default=None,
        help="The label (prefix) of the benchmark results. If not specified, "
1193
        "the value of '--backend' will be used as the label.",
1194
    )
1195
1196
1197
    parser.add_argument(
        "--backend",
        type=str,
1198
1199
        default="openai",
        choices=list(ASYNC_REQUEST_FUNCS.keys()),
1200
        help="The type of backend or endpoint to use for the benchmark.",
1201
    )
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
    parser.add_argument(
        "--base-url",
        type=str,
        default=None,
        help="Server or API base url if not using http host and port.",
    )
    # Use 127.0.0.1 here instead of localhost to force the use of ipv4
    parser.add_argument("--host", type=str, default="127.0.0.1")
    parser.add_argument("--port", type=int, default=8000)
    parser.add_argument(
        "--endpoint",
        type=str,
        default="/v1/completions",
        help="API endpoint.",
    )
1217
1218
1219
1220
1221
    parser.add_argument(
        "--header",
        metavar="KEY=VALUE",
        nargs="*",
        help="Key-value pairs (e.g, --header x-additional-info=0.3.3) "
1222
1223
        "for headers to be passed with each request. These headers override "
        "per backend constants and values set via environment variable, and "
1224
        "will be overridden by other arguments (such as request ids).",
1225
    )
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
    parser.add_argument(
        "--max-concurrency",
        type=int,
        default=None,
        help="Maximum number of concurrent requests. This can be used "
        "to help simulate an environment where a higher level component "
        "is enforcing a maximum number of concurrent requests. While the "
        "--request-rate argument controls the rate at which requests are "
        "initiated, this argument will control how many are actually allowed "
        "to execute at a time. This means that when used in combination, the "
        "actual request rate may be lower than specified with --request-rate, "
1237
1238
        "if the server is not processing requests fast enough to keep up.",
    )
1239
1240
1241
1242

    parser.add_argument(
        "--model",
        type=str,
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
        required=False,
        default=None,
        help="Name of the model. If not specified, will fetch the first model "
        "from the server's /v1/models endpoint.",
    )
    parser.add_argument(
        "--input-len",
        type=int,
        default=None,
        help="General input length for datasets. Maps to dataset-specific "
        "input length arguments (e.g., --random-input-len, --sonnet-input-len). "
        "If not specified, uses dataset defaults.",
    )
    parser.add_argument(
        "--output-len",
        type=int,
        default=None,
        help="General output length for datasets. Maps to dataset-specific "
        "output length arguments (e.g., --random-output-len, --sonnet-output-len). "
        "If not specified, uses dataset defaults.",
1263
1264
1265
1266
    )
    parser.add_argument(
        "--tokenizer",
        type=str,
1267
        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
1268
    )
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
    parser.add_argument(
        "--tokenizer-mode",
        type=str,
        default="auto",
        help="""Tokenizer mode:\n
        - "auto" will use the tokenizer from `mistral_common` for Mistral models
        if available, otherwise it will use the "hf" tokenizer.\n
        - "hf" will use the fast tokenizer if available.\n
        - "slow" will always use the slow tokenizer.\n
        - "mistral" will always use the tokenizer from `mistral_common`.\n
        - "deepseek_v32" will always use the tokenizer from `deepseek_v32`.\n
        - Other custom values can be supported via plugins.""",
    )
1282
1283
1284
1285
1286
    parser.add_argument("--use-beam-search", action="store_true")
    parser.add_argument(
        "--logprobs",
        type=int,
        default=None,
1287
1288
1289
1290
1291
1292
1293
        help=(
            "Number of logprobs-per-token to compute & return as part of "
            "the request. If unspecified, then either (1) if beam search "
            "is disabled, no logprobs are computed & a single dummy "
            "logprob is returned for each token; or (2) if beam search "
            "is enabled 1 logprob per token is computed"
        ),
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
    )
    parser.add_argument(
        "--request-rate",
        type=float,
        default=float("inf"),
        help="Number of requests per second. If this is inf, "
        "then all the requests are sent at time 0. "
        "Otherwise, we use Poisson process or gamma distribution "
        "to synthesize the request arrival times.",
    )
    parser.add_argument(
        "--burstiness",
        type=float,
        default=1.0,
        help="Burstiness factor of the request generation. "
        "Only take effect when request_rate is not inf. "
        "Default value is 1, which follows Poisson process. "
        "Otherwise, the request intervals follow a gamma distribution. "
        "A lower burstiness value (0 < burstiness < 1) results in more "
        "bursty requests. A higher burstiness value (burstiness > 1) "
        "results in a more uniform arrival of requests.",
    )
    parser.add_argument(
        "--disable-tqdm",
        action="store_true",
        help="Specify to disable tqdm progress bar.",
    )
1321
1322
1323
1324
1325
1326
    parser.add_argument(
        "--num-warmups",
        type=int,
        default=0,
        help="Number of warmup requests.",
    )
1327
1328
1329
    parser.add_argument(
        "--profile",
        action="store_true",
1330
        help="Use vLLM Profiling. --profiler-config must be provided on the server.",
1331
1332
1333
1334
1335
1336
    )
    parser.add_argument(
        "--save-result",
        action="store_true",
        help="Specify to save benchmark results to a json file",
    )
1337
1338
1339
1340
    parser.add_argument(
        "--save-detailed",
        action="store_true",
        help="When saving the results, whether to include per request "
1341
        "information such as response, error, ttfts, tpots, etc.",
1342
1343
1344
1345
1346
1347
    )
    parser.add_argument(
        "--append-result",
        action="store_true",
        help="Append the benchmark result to the existing json file.",
    )
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
    parser.add_argument(
        "--metadata",
        metavar="KEY=VALUE",
        nargs="*",
        help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) "
        "for metadata of this run to be saved in the result JSON file "
        "for record keeping purposes.",
    )
    parser.add_argument(
        "--result-dir",
        type=str,
        default=None,
        help="Specify directory to save benchmark json results."
        "If not specified, results are saved in the current directory.",
    )
    parser.add_argument(
        "--result-filename",
        type=str,
        default=None,
        help="Specify the filename to save benchmark json results."
        "If not specified, results will be saved in "
        "{label}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"  # noqa
        " format.",
    )
    parser.add_argument(
        "--ignore-eos",
        action="store_true",
        help="Set ignore_eos flag when sending the benchmark request."
1376
1377
        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.",
    )
1378
1379
1380
    parser.add_argument(
        "--percentile-metrics",
        type=str,
1381
        default=None,
1382
        help="Comma-separated list of selected metrics to report percentiles. "
1383
        "This argument specifies the metrics to report percentiles. "
1384
1385
1386
        'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
        'If not specified, defaults to "ttft,tpot,itl" for generative models '
        'and "e2el" for pooling models.',
1387
    )
1388
1389
1390
1391
    parser.add_argument(
        "--metric-percentiles",
        type=str,
        default="99",
1392
        help="Comma-separated list of percentiles for selected metrics. "
1393
1394
1395
        'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". '
        'Default value is "99".'
        'Use "--percentile-metrics" to select metrics.',
1396
1397
1398
1399
1400
    )
    parser.add_argument(
        "--goodput",
        nargs="+",
        required=False,
1401
        help='Specify service level objectives for goodput as "KEY:VALUE" '
1402
        "pairs, where the key is a metric name, and the value is in "
1403
        'milliseconds. Multiple "KEY:VALUE" pairs can be provided, '
1404
        "separated by spaces. Allowed request level metric names are "
1405
        '"ttft", "tpot", "e2el". For more context on the definition of '
1406
        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
1407
1408
        "and the blog: https://hao-ai-lab.github.io/blogs/distserve",
    )
1409
1410
1411
1412
    parser.add_argument(
        "--request-id-prefix",
        type=str,
        required=False,
1413
        default=f"bench-{uuid.uuid4().hex[:8]}-",
1414
1415
1416
        help="Specify the prefix of request id.",
    )

1417
1418
1419
1420
1421
    sampling_group = parser.add_argument_group("sampling parameters")
    sampling_group.add_argument(
        "--top-p",
        type=float,
        default=None,
1422
        help="Top-p sampling parameter. Only has effect on openai-compatible backends.",
1423
1424
1425
1426
1427
    )
    sampling_group.add_argument(
        "--top-k",
        type=int,
        default=None,
1428
        help="Top-k sampling parameter. Only has effect on openai-compatible backends.",
1429
1430
1431
1432
1433
    )
    sampling_group.add_argument(
        "--min-p",
        type=float,
        default=None,
1434
        help="Min-p sampling parameter. Only has effect on openai-compatible backends.",
1435
1436
1437
1438
1439
1440
    )
    sampling_group.add_argument(
        "--temperature",
        type=float,
        default=None,
        help="Temperature sampling parameter. Only has effect on "
1441
        "openai-compatible backends.",
1442
    )
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
    sampling_group.add_argument(
        "--frequency-penalty",
        type=float,
        default=None,
        help="Frequency penalty sampling parameter. Only has effect on "
        "openai-compatible backends.",
    )
    sampling_group.add_argument(
        "--presence-penalty",
        type=float,
        default=None,
        help="Presence penalty sampling parameter. Only has effect on "
        "openai-compatible backends.",
    )
    sampling_group.add_argument(
        "--repetition-penalty",
        type=float,
        default=None,
        help="Repetition penalty sampling parameter. Only has effect on "
        "openai-compatible backends.",
    )
1464

1465
1466
1467
1468
1469
1470
    parser.add_argument(
        "--served-model-name",
        type=str,
        default=None,
        help="The model name used in the API. "
        "If not specified, the model name will be the "
1471
        "same as the `--model` argument. ",
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
    )

    parser.add_argument(
        "--lora-modules",
        nargs="+",
        default=None,
        help="A subset of LoRA module names passed in when "
        "launching the server. For each request, the "
        "script chooses a LoRA module at random.",
    )
1482

1483
1484
1485
1486
1487
1488
1489
1490
    parser.add_argument(
        "--ramp-up-strategy",
        type=str,
        default=None,
        choices=["linear", "exponential"],
        help="The ramp-up strategy. This would be used to "
        "ramp up the request rate from initial RPS to final "
        "RPS rate (specified by --ramp-up-start-rps and "
1491
1492
        "--ramp-up-end-rps.) over the duration of the benchmark.",
    )
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
    parser.add_argument(
        "--ramp-up-start-rps",
        type=int,
        default=None,
        help="The starting request rate for ramp-up (RPS). "
        "Needs to be specified when --ramp-up-strategy is used.",
    )
    parser.add_argument(
        "--ramp-up-end-rps",
        type=int,
        default=None,
        help="The ending request rate for ramp-up (RPS). "
        "Needs to be specified when --ramp-up-strategy is used.",
    )
1507
1508
1509
    parser.add_argument(
        "--ready-check-timeout-sec",
        type=int,
1510
        default=0,
1511
        help="Maximum time to wait for the endpoint to become ready "
1512
        "in seconds. Ready check will be skipped by default.",
1513
    )
1514

1515
1516
1517
1518
1519
1520
1521
1522
    parser.add_argument(
        "--extra-body",
        help="A JSON string representing extra body parameters to include "
        "in each request."
        'Example: \'{"chat_template_kwargs":{"enable_thinking":false}}\'',
        type=json.loads,
        default=None,
    )
1523
1524
1525
1526
1527
1528
    parser.add_argument(
        "--skip-tokenizer-init",
        action="store_true",
        default=False,
        help="Skip initialization of tokenizer and detokenizer",
    )
1529

1530
1531
1532
1533
1534
1535
1536
1537
    parser.add_argument(
        "--insecure",
        action="store_true",
        default=False,
        help="Disable SSL certificate verification. Use this option when "
        "connecting to servers with self-signed certificates.",
    )

1538

1539
1540
1541
def main(args: argparse.Namespace) -> dict[str, Any]:
    return asyncio.run(main_async(args))

1542

1543
async def main_async(args: argparse.Namespace) -> dict[str, Any]:
1544
1545
1546
1547
    print(args)
    random.seed(args.seed)
    np.random.seed(args.seed)

1548
1549
1550
1551
1552
1553
    # Validate ramp-up arguments
    if args.ramp_up_strategy is not None:
        if args.request_rate != float("inf"):
            raise ValueError(
                "When using ramp-up, do not specify --request-rate. "
                "The request rate will be controlled by ramp-up parameters. "
1554
1555
                "Please remove the --request-rate argument."
            )
1556
1557
1558
        if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None:
            raise ValueError(
                "When using --ramp-up-strategy, both --ramp-up-start-rps and "
1559
1560
                "--ramp-up-end-rps must be specified"
            )
1561
1562
1563
1564
        if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0:
            raise ValueError("Ramp-up start and end RPS must be non-negative")
        if args.ramp_up_start_rps > args.ramp_up_end_rps:
            raise ValueError("Ramp-up start RPS must be less than end RPS")
1565
1566
        if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0:
            raise ValueError("For exponential ramp-up, the start RPS cannot be 0.")
1567

1568
1569
1570
1571
1572
1573
    label = args.label

    if args.base_url is not None:
        api_url = f"{args.base_url}{args.endpoint}"
        base_url = f"{args.base_url}"
    else:
1574
1575
1576
        host_port = join_host_port(args.host, args.port)
        api_url = f"http://{host_port}{args.endpoint}"
        base_url = f"http://{host_port}"
1577

1578
1579
1580
1581
1582
1583
1584
1585
1586
    # Headers
    headers = None
    if args.header:
        headers = {}
        for item in args.header:
            if "=" in item:
                kvstring = item.split("=", 1)
                headers[kvstring[0].strip()] = kvstring[1].strip()
            else:
1587
                raise ValueError("Invalid header format. Please use KEY=VALUE format.")
1588

1589
1590
1591
1592
1593
1594
1595
1596
1597
    # SSL context configuration
    ssl_context: ssl.SSLContext | bool | None = None
    if args.insecure:
        # Disable SSL certificate verification
        ssl_context = False
    elif "https://" in base_url:
        # Use default SSL context for HTTPS
        ssl_context = True

1598
1599
1600
    # Fetch model from server if not specified
    if args.model is None:
        print("Model not specified, fetching first model from server...")
1601
1602
1603
        model_name, model_id = await get_first_model_from_server(
            base_url, headers, ssl_context
        )
1604
        print(f"First model name: {model_name}, first model id: {model_id}")
1605
    else:
1606
        model_name = args.served_model_name
1607
1608
        model_id = args.model

1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
    if args.skip_tokenizer_init:
        tokenizer_id = None
        tokenizer_mode = None
        tokenizer = None
    else:
        tokenizer_id = args.tokenizer if args.tokenizer is not None else model_id
        tokenizer_mode = args.tokenizer_mode
        tokenizer = get_tokenizer(
            tokenizer_id,
            tokenizer_mode=tokenizer_mode,
            trust_remote_code=args.trust_remote_code,
        )
1621

1622
1623
1624
    if args.dataset_name is None:
        raise ValueError(
            "Please specify '--dataset-name' and the corresponding "
1625
1626
            "'--dataset-path' if required."
        )
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640

    # Map general --input-len and --output-len to all dataset-specific arguments
    if args.input_len is not None:
        args.random_input_len = args.input_len
        args.sonnet_input_len = args.input_len

    if args.output_len is not None:
        args.random_output_len = args.output_len
        args.sonnet_output_len = args.output_len
        args.sharegpt_output_len = args.output_len
        args.custom_output_len = args.output_len
        args.hf_output_len = args.output_len
        args.spec_bench_output_len = args.output_len
        args.prefix_repetition_output_len = args.output_len
1641

1642
1643
1644
1645
1646
1647
1648
1649
    # when using random datasets, default to ignoring EOS
    # so generation runs to the requested length
    if (
        args.dataset_name in ("random", "random-mm")
        and args.backend in OPENAI_COMPATIBLE_BACKENDS
    ):
        args.ignore_eos = True

1650
1651
    # Load the dataset.
    input_requests = get_samples(args, tokenizer)
1652
1653
    goodput_config_dict = check_goodput_args(args)

1654
    backend = args.backend
1655
1656
1657
1658
1659
    task_type = (
        TaskType.POOLING
        if "embeddings" in backend or "rerank" in backend
        else TaskType.GENERATION
    )
1660

1661
    # Collect the sampling parameters.
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
    if task_type == TaskType.GENERATION:
        sampling_params = {
            k: v
            for k, v in {
                "top_p": args.top_p,
                "top_k": args.top_k,
                "min_p": args.min_p,
                "temperature": args.temperature,
                "frequency_penalty": args.frequency_penalty,
                "presence_penalty": args.presence_penalty,
                "repetition_penalty": args.repetition_penalty,
            }.items()
            if v is not None
        }

        # Sampling parameters are only supported by openai-compatible backend.
        if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS:
            raise ValueError(
                "Sampling parameters are only supported by openai-compatible backends."
            )
1682

1683
        if "temperature" not in sampling_params:
1684
1685
1686
1687
1688
1689
            print(
                "WARNING: vllm bench serve no longer sets temperature==0 (greedy) "
                "in requests by default. The default will be determined on the "
                "server side and can be model/API specific. "
                "For the old behavior, include --temperature=0."
            )
1690
1691

        default_percentile_metrics = "ttft,tpot,itl"
1692
1693
    else:
        sampling_params = {}
1694
        default_percentile_metrics = "e2el"
1695

1696
1697
1698
    extra_body = args.extra_body or {}
    extra_body = {**sampling_params, **extra_body}

1699
1700
    percentile_metrics: str = args.percentile_metrics or default_percentile_metrics

1701
    # Avoid GC processing "static" data - reduce pause times.
1702
    freeze_gc_heap()
1703

1704
    benchmark_result = await benchmark(
1705
1706
        task_type=task_type,
        endpoint_type=backend,
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
        api_url=api_url,
        base_url=base_url,
        model_id=model_id,
        model_name=model_name,
        tokenizer=tokenizer,
        input_requests=input_requests,
        logprobs=args.logprobs,
        request_rate=args.request_rate,
        burstiness=args.burstiness,
        disable_tqdm=args.disable_tqdm,
1717
        num_warmups=args.num_warmups,
1718
        profile=args.profile,
1719
        selected_percentile_metrics=percentile_metrics.split(","),
1720
        selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")],
1721
1722
1723
1724
        ignore_eos=args.ignore_eos,
        goodput_config_dict=goodput_config_dict,
        max_concurrency=args.max_concurrency,
        lora_modules=args.lora_modules,
1725
        extra_headers=headers,
1726
        extra_body=extra_body,
1727
1728
1729
1730
        ramp_up_strategy=args.ramp_up_strategy,
        ramp_up_start_rps=args.ramp_up_start_rps,
        ramp_up_end_rps=args.ramp_up_end_rps,
        ready_check_timeout_sec=args.ready_check_timeout_sec,
1731
        ssl_context=ssl_context,
1732
    )
1733
1734

    # Save config and results to json
1735
1736
1737
1738
1739
    result_json: dict[str, Any] = {}

    # Setup
    current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
    result_json["date"] = current_dt
1740
    result_json["endpoint_type"] = args.backend  # for backward compatibility
1741
    result_json["backend"] = args.backend
1742
1743
1744
1745
1746
1747
1748
1749
1750
    result_json["label"] = label
    result_json["model_id"] = model_id
    result_json["tokenizer_id"] = tokenizer_id
    result_json["num_prompts"] = args.num_prompts

    # Metadata
    if args.metadata:
        for item in args.metadata:
            if "=" in item:
1751
                kvstring = item.split("=", 1)
1752
1753
1754
                result_json[kvstring[0].strip()] = kvstring[1].strip()
            else:
                raise ValueError(
1755
1756
                    "Invalid metadata format. Please use KEY=VALUE format."
                )
1757

1758
    # Traffic
1759
1760
1761
    result_json["request_rate"] = (
        args.request_rate if args.request_rate < float("inf") else "inf"
    )
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
    result_json["burstiness"] = args.burstiness
    result_json["max_concurrency"] = args.max_concurrency

    if args.ramp_up_strategy is not None:
        result_json["ramp_up_strategy"] = args.ramp_up_strategy
        result_json["ramp_up_start_rps"] = args.ramp_up_start_rps
        result_json["ramp_up_end_rps"] = args.ramp_up_end_rps

    # Merge with benchmark result
    result_json = {**result_json, **benchmark_result}

    if not args.save_detailed:
        # Remove fields with too many data points
        for field in [
1776
1777
            "input_lens",
            "output_lens",
1778
            "start_times",
1779
1780
1781
1782
            "ttfts",
            "itls",
            "generated_texts",
            "errors",
1783
1784
1785
1786
1787
        ]:
            if field in result_json:
                del result_json[field]
            if field in benchmark_result:
                del benchmark_result[field]
1788

1789
        # Save to file
1790
    if args.save_result or args.append_result:
1791
        base_model_id = model_id.split("/")[-1]
1792
1793
1794
1795
1796
        max_concurrency_str = (
            f"-concurrency{args.max_concurrency}"
            if args.max_concurrency is not None
            else ""
        )
1797
        label = label or args.backend
1798
        if args.ramp_up_strategy is not None:
1799
            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
1800
1801
        else:
            file_name = f"{label}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
1802
1803
1804
        if args.result_filename:
            file_name = args.result_filename
        if args.result_dir:
1805
            os.makedirs(args.result_dir, exist_ok=True)
1806
            file_name = os.path.join(args.result_dir, file_name)
1807
1808
1809
        with open(
            file_name, mode="a+" if args.append_result else "w", encoding="utf-8"
        ) as outfile:
1810
1811
1812
            # Append a newline.
            if args.append_result and outfile.tell() != 0:
                outfile.write("\n")
1813
1814
            json.dump(result_json, outfile)
        save_to_pytorch_benchmark_format(args, result_json, file_name)
1815

1816
    return result_json