serve.py 52.1 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
14
15
16
17
        --model <your_model> \
        --dataset-name <dataset_name. Default 'random'> \
        --request-rate <request_rate. Default inf> \
        --num-prompts <num_prompts. Default 1000>
"""
18

19
20
import argparse
import asyncio
21
import contextlib
22
import gc
23
import importlib.util
24
25
26
import json
import os
import random
27
import shutil
28
29
import time
import warnings
30
from collections.abc import AsyncGenerator, Iterable
31
32
from dataclasses import dataclass
from datetime import datetime
33
from enum import Enum
34
from typing import Any, Literal
35

36
import aiohttp
37
38
39
40
import numpy as np
from tqdm.asyncio import tqdm
from transformers import PreTrainedTokenizerBase

41
from vllm.benchmarks.datasets import SampleRequest, add_dataset_parser, get_samples
42
from vllm.benchmarks.lib.endpoint_request_func import (
43
44
45
46
47
    ASYNC_REQUEST_FUNCS,
    OPENAI_COMPATIBLE_BACKENDS,
    RequestFuncInput,
    RequestFuncOutput,
)
48
from vllm.benchmarks.lib.ready_checker import wait_for_endpoint
49
from vllm.benchmarks.lib.utils import convert_to_pytorch_benchmark_format, write_to_json
50
51
52
53
from vllm.transformers_utils.tokenizer import get_tokenizer

MILLISECONDS_TO_SECONDS_CONVERSION = 1000

54
55
56
TERM_PLOTLIB_AVAILABLE = (importlib.util.find_spec("termplotlib") is not None) and (
    shutil.which("gnuplot") is not None
)
57

58

59
60
class TaskType(Enum):
    GENERATION = "generation"
61
    POOLING = "pooling"
62
63


64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
@dataclass
class BenchmarkMetrics:
    completed: int
    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]]
92
93
94
95
    # Max output tokens per second and concurrent requests at that peak
    max_output_tokens_per_s: float
    max_concurrent_requests: int

96

97
98
99
100
101
@dataclass
class EmbedBenchmarkMetrics:
    completed: int
    total_input: int
    request_throughput: float
102
    total_token_throughput: float
103
104
105
106
    mean_e2el_ms: float
    std_e2el_ms: float
    median_e2el_ms: float
    percentiles_e2el_ms: float
107

108

109
def _get_current_request_rate(
110
111
112
    ramp_up_strategy: Literal["linear", "exponential"] | None,
    ramp_up_start_rps: int | None,
    ramp_up_end_rps: int | None,
113
114
115
116
    request_index: int,
    total_requests: int,
    request_rate: float,
) -> float:
117
118
119
120
121
    if (
        ramp_up_strategy
        and ramp_up_start_rps is not None
        and ramp_up_end_rps is not None
    ):
122
123
124
125
126
127
128
129
130
131
132
133
        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


134
async def get_request(
135
    input_requests: list[SampleRequest],
136
137
    request_rate: float,
    burstiness: float = 1.0,
138
139
140
    ramp_up_strategy: Literal["linear", "exponential"] | None = None,
    ramp_up_start_rps: int | None = None,
    ramp_up_end_rps: int | None = None,
141
) -> AsyncGenerator[tuple[SampleRequest, float], None]:
142
143
    """
    Asynchronously generates requests at a specified rate
144
    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
145
146
147

    Args:
        input_requests:
148
            A list of input requests, each represented as a SampleRequest.
149
150
151
152
153
154
155
156
157
158
        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.
159
        ramp_up_strategy (optional):
160
161
162
163
164
165
            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.
166
167
    """
    assert burstiness > 0, (
168
169
        f"A positive burstiness factor is expected, but given {burstiness}."
    )
170
    # Convert to list to get length for ramp-up calculations
171
    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
172
        input_requests = list(input_requests)
173

174
    total_requests = len(input_requests)
175
    assert total_requests > 0, "No requests provided."
176

177
178
179
180
    # Precompute delays among requests to minimize request send laggings
    request_rates = []
    delay_ts = []
    for request_index, request in enumerate(input_requests):
181
        current_request_rate = _get_current_request_rate(
182
183
184
185
186
187
188
            ramp_up_strategy,
            ramp_up_start_rps,
            ramp_up_end_rps,
            request_index,
            total_requests,
            request_rate,
        )
189
        request_rates.append(current_request_rate)
190
        if current_request_rate == float("inf"):
191
192
193
194
195
196
197
            delay_ts.append(0)
        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))
198

199
200
201
202
203
204
205
206
207
    # 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.
        #
208
209
        # NOTE: If we simply accumulate the random delta values
        # from the gamma distribution, their sum would have 1-2% gap
210
        # from target_total_delay_s. The purpose of the following logic is to
co63oc's avatar
co63oc committed
211
        # close the gap for stabilizing the throughput data
212
        # from different random seeds.
213
214
215
216
217
218
        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):
219
220
221
222
223
        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)
224
        yield request, request_rates[request_index]
225
226


227
def calculate_metrics_for_embeddings(
228
229
    outputs: list[RequestFuncOutput], dur_s: float, selected_percentiles: list[float]
) -> EmbedBenchmarkMetrics:
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
    """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
    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

    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
            "on the benchmark arguments.",
253
254
            stacklevel=2,
        )
255
256
257
258
259
260
261
262
    metrics = EmbedBenchmarkMetrics(
        completed=completed,
        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,
263
264
265
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
266
267
268
269
    )
    return metrics


270
def calculate_metrics(
271
    input_requests: list[SampleRequest],
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
    outputs: list[RequestFuncOutput],
    dur_s: float,
    tokenizer: PreTrainedTokenizerBase,
    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] = []
    for i in range(len(outputs)):
        if outputs[i].success:
            output_len = outputs[i].output_tokens

304
            if not output_len:
305
306
307
308
309
310
                # 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(
311
312
313
314
                    tokenizer(
                        outputs[i].generated_text, add_special_tokens=False
                    ).input_ids
                )
315
            actual_output_lens.append(output_len)
316
            total_input += input_requests[i].prompt_len
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
            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)
            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)
337
338
339
            slo_values.append(
                goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
340
341
        if "tpot" in goodput_config_dict:
            valid_metrics.append(all_tpots)
342
343
344
            slo_values.append(
                goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
345
346
        if "e2el" in goodput_config_dict:
            valid_metrics.append(e2els)
347
348
349
            slo_values.append(
                goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
350
351
352
353
354
355
356
357
358
359

        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.",
360
361
            stacklevel=2,
        )
362
363
364
365
366
367
368
369

    # 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]
    if successful_outputs:
370
371
372
373
        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
        )
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396

        # 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)
397
398
399
            request_end_second = int(
                (output.start_time + output.latency) - min_start_time
            )
400
401
402
403
404
405
406
407

            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))
408
            max_concurrent_requests = int(np.max(concurrent_requests_per_second))
409
410
411

        if TERM_PLOTLIB_AVAILABLE:
            import termplotlib as tpl
412

413
            fig = tpl.figure()
414
415
416
417
418
419
420
421
422
423
            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",
            )
424
425
426
427
            fig.show()
        else:
            print("tip: install termplotlib and gnuplot to plot the metrics")

428
429
430
431
432
433
434
435
    metrics = BenchmarkMetrics(
        completed=completed,
        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,
436
437
        mean_ttft_ms=np.mean(ttfts or 0)
        * 1000,  # ttfts is empty if streaming is not supported by the endpoint
438
439
        std_ttft_ms=np.std(ttfts or 0) * 1000,
        median_ttft_ms=np.median(ttfts or 0) * 1000,
440
441
442
        percentiles_ttft_ms=[
            (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles
        ],
443
444
445
        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,
446
447
448
        percentiles_tpot_ms=[
            (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles
        ],
449
450
451
        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,
452
453
454
        percentiles_itl_ms=[
            (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles
        ],
455
456
457
        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,
458
459
460
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
461
462
        max_output_tokens_per_s=max_output_tokens_per_s,
        max_concurrent_requests=max_concurrent_requests,
463
464
465
466
467
468
    )

    return metrics, actual_output_lens


async def benchmark(
469
    task_type: TaskType,
470
471
472
473
474
475
    endpoint_type: str,
    api_url: str,
    base_url: str,
    model_id: str,
    model_name: str,
    tokenizer: PreTrainedTokenizerBase,
476
    input_requests: list[SampleRequest],
477
    logprobs: int | None,
478
479
480
481
482
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
    profile: bool,
    selected_percentile_metrics: list[str],
483
    selected_percentiles: list[float],
484
485
    ignore_eos: bool,
    goodput_config_dict: dict[str, float],
486
487
488
489
490
491
492
    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,
493
    ready_check_timeout_sec: int = 600,
494
):
495
496
497
498
    try:
        request_func = ASYNC_REQUEST_FUNCS[endpoint_type]
    except KeyError:
        raise ValueError(f"Unknown backend: {endpoint_type}") from None
499

500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
    # Reuses connections across requests to reduce TLS handshake overhead.
    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,
        ssl=("https://" in api_url),
    )

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

518
519
    print("Starting initial single prompt test run...")
    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
520
521
522
523
524
525
        input_requests[0].prompt,
        input_requests[0].prompt_len,
        input_requests[0].expected_output_len,
        input_requests[0].multi_modal_data,
    )

526
527
528
529
530
531
532
533
    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]"
534
535
536
537
538
539
540
541
542
543
    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,
544
        extra_headers=extra_headers,
545
        extra_body=extra_body,
546
547
    )

548
549
550
551
552
553
554
555
556
557
558
    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. "
559
560
                f"Error: {test_output.error}"
            )
561
562
        else:
            print("Initial test run completed. Starting main benchmark run...")
563
    else:
564
        print("Skipping endpoint ready check.")
565
566
567
568

    if lora_modules:
        # For each input request, choose a LoRA module at random.
        lora_modules = iter(
569
570
            [random.choice(lora_modules) for _ in range(len(input_requests))]
        )
571
572
573

    if profile:
        print("Starting profiler...")
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
        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
        )
590
591
592
        if profile_output.success:
            print("Profiler started")

593
    distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
594
595
596

    if ramp_up_strategy is not None:
        print(f"Traffic ramp-up strategy: {ramp_up_strategy}.")
597
598
599
600
        print(
            f"Will increase RPS from {ramp_up_start_rps} to "
            f"{ramp_up_end_rps} RPS over the duration of the benchmark."
        )
601
    else:
602
        print(f"Traffic request rate: {request_rate}")
603
604
605
606
607
608

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

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

609
610
611
612
613
    semaphore = (
        asyncio.Semaphore(max_concurrency)
        if max_concurrency
        else contextlib.nullcontext()
    )
614

615
    async def limited_request_func(request_func_input, session, pbar):
616
        async with semaphore:
617
618
619
            return await request_func(
                request_func_input=request_func_input, session=session, pbar=pbar
            )
620
621
622

    benchmark_start_time = time.perf_counter()
    tasks: list[asyncio.Task] = []
623
624
625
626
627

    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
628
629
630
631
632
633
        rps_change_events.append(
            {
                "rps": last_int_rps,
                "timestamp": datetime.now().isoformat(),
            }
        )
634
635

    async for request, current_request_rate in get_request(
636
637
638
639
640
641
642
        input_requests,
        request_rate,
        burstiness,
        ramp_up_strategy,
        ramp_up_start_rps,
        ramp_up_end_rps,
    ):
643
644
645
646
647
        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):
648
                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
649
                last_int_rps = current_int_rps
650
        prompt, prompt_len, output_len, mm_content, request_id = (
651
652
653
654
            request.prompt,
            request.prompt_len,
            request.expected_output_len,
            request.multi_modal_data,
655
            request.request_id,
656
        )
657
658
659
660
661
        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

662
663
664
665
666
667
668
669
670
671
672
673
674
675
        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,
        )
676
677
        tasks.append(
            asyncio.create_task(
678
679
680
681
682
                limited_request_func(
                    request_func_input=request_func_input, session=session, pbar=pbar
                )
            )
        )
683
684
685
686
687
688
689
    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)

    if pbar is not None:
        pbar.close()

    benchmark_duration = time.perf_counter() - benchmark_start_time

690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
    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
706

707
    print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
708
    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
709
    if max_concurrency is not None:
710
711
712
713
        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))
714
    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
715
    if isinstance(metrics, BenchmarkMetrics):
716
717
718
719
720
721
        print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
    print(
        "{:<40} {:<10.2f}".format(
            "Request throughput (req/s):", metrics.request_throughput
        )
    )
722
    if goodput_config_dict:
723
724
725
726
727
        print(
            "{:<40} {:<10.2f}".format(
                "Request goodput (req/s):", metrics.request_goodput
            )
        )
728
    if isinstance(metrics, BenchmarkMetrics):
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
        print(
            "{:<40} {:<10.2f}".format(
                "Output token throughput (tok/s):", metrics.output_throughput
            )
        )
        print(
            "{:<40} {:<10.2f}".format(
                "Peak output token throughput (tok/s):", metrics.max_output_tokens_per_s
            )
        )
        print(
            "{:<40} {:<10.2f}".format(
                "Peak concurrent requests:", metrics.max_concurrent_requests
            )
        )
    print(
        "{:<40} {:<10.2f}".format(
            "Total Token throughput (tok/s):", metrics.total_token_throughput
        )
    )
749

750
751
752
753
754
755
756
    if isinstance(metrics, BenchmarkMetrics):
        result = {
            "duration": benchmark_duration,
            "completed": metrics.completed,
            "total_input_tokens": metrics.total_input,
            "total_output_tokens": metrics.total_output,
            "request_throughput": metrics.request_throughput,
757
            "request_goodput": metrics.request_goodput if goodput_config_dict else None,
758
759
760
761
762
763
764
765
            "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],
            "generated_texts": [output.generated_text for output in outputs],
            "errors": [output.error for output in outputs],
766
767
            "max_output_tokens_per_s": metrics.max_output_tokens_per_s,
            "max_concurrent_requests": metrics.max_concurrent_requests,
768
769
770
771
772
773
774
775
776
777
778
        }
    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],
        }
779

780
781
782
    if rps_change_events:
        result["rps_change_events"] = rps_change_events

783
784
785
786
787
788
789
790
791
792
793
794
    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
795
796
797
798
799
800
801
802
803
804
805
806
807
        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"),
            )
        )
808
        result[f"mean_{metric_attribute_name}_ms"] = getattr(
809
810
            metrics, f"mean_{metric_attribute_name}_ms"
        )
811
        result[f"median_{metric_attribute_name}_ms"] = getattr(
812
813
            metrics, f"median_{metric_attribute_name}_ms"
        )
814
        result[f"std_{metric_attribute_name}_ms"] = getattr(
815
816
817
            metrics, f"std_{metric_attribute_name}_ms"
        )
        for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"):
818
            p_word = str(int(p)) if int(p) == p else str(p)
819
            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value))
820
821
            result[f"p{p_word}_{metric_attribute_name}_ms"] = value

822
823
    if task_type == TaskType.GENERATION:
        process_one_metric("ttft", "TTFT", "Time to First Token")
824
        process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)")
825
        process_one_metric("itl", "ITL", "Inter-token Latency")
826
827
828
829
    process_one_metric("e2el", "E2EL", "End-to-end Latency")

    print("=" * 50)

830
831
832
833
834
835
836
837
838
839
    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,
        )
840
841
842
        profile_output = await request_func(
            request_func_input=profile_input, session=session
        )
843
844
        if profile_output.success:
            print("Profiler stopped")
845
846

    await session.close()
847
848
849
850
851
852
853
854
855
856
857
858
859
860
    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 "
861
862
                    f"{str(VALID_NAMES)}. "
                )
863
864
865
866
            if slo_val < 0:
                raise ValueError(
                    f"Invalid value found, {slo_name}: {slo_val}. "
                    "The service level objective value should be "
867
868
                    "non-negative."
                )
869
870
871
872
873
874
875
876
877
878
879
880
    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. "
881
            'Specify service level objectives for goodput as "KEY:VALUE" '
882
            "pairs, where the key is a metric name, and the value is a "
883
884
            "number in milliseconds."
        ) from err
885
886
887
    return goodput_config_dict


888
889
890
def save_to_pytorch_benchmark_format(
    args: argparse.Namespace, results: dict[str, Any], file_name: str
) -> None:
891
    metrics = [
892
893
894
895
896
897
898
899
900
901
902
903
        "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",
904
905
906
907
908
909
    ]
    # 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,
910
        metrics={k: [results[k]] for k in metrics if k in results},
911
912
        extra_info={
            k: results[k]
913
914
915
916
            for k in results
            if k not in metrics and k not in ignored_metrics
        },
    )
917
918
919
920
921
922
923
    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):
924
    add_dataset_parser(parser)
925
926
927
928
929
    parser.add_argument(
        "--label",
        type=str,
        default=None,
        help="The label (prefix) of the benchmark results. If not specified, "
930
        "the value of '--backend' will be used as the label.",
931
    )
932
933
934
    parser.add_argument(
        "--backend",
        type=str,
935
936
        default="openai",
        choices=list(ASYNC_REQUEST_FUNCS.keys()),
937
        help="The type of backend or endpoint to use for the benchmark.",
938
    )
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
    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.",
    )
954
955
956
957
958
    parser.add_argument(
        "--header",
        metavar="KEY=VALUE",
        nargs="*",
        help="Key-value pairs (e.g, --header x-additional-info=0.3.3) "
959
960
961
        "for headers to be passed with each request. These headers override "
        "per backend constants and values set via environment variable, and "
        "will be overriden by other arguments (such as request ids).",
962
    )
963
964
965
966
967
968
969
970
971
972
973
    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, "
974
975
        "if the server is not processing requests fast enough to keep up.",
    )
976
977
978
979
980
981
982
983
984
985

    parser.add_argument(
        "--model",
        type=str,
        required=True,
        help="Name of the model.",
    )
    parser.add_argument(
        "--tokenizer",
        type=str,
986
        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
987
988
989
990
991
992
    )
    parser.add_argument("--use-beam-search", action="store_true")
    parser.add_argument(
        "--logprobs",
        type=int,
        default=None,
993
994
995
996
997
998
999
        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"
        ),
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
    )
    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(
        "--trust-remote-code",
        action="store_true",
        help="Trust remote code from huggingface",
    )
    parser.add_argument(
        "--disable-tqdm",
        action="store_true",
        help="Specify to disable tqdm progress bar.",
    )
    parser.add_argument(
        "--profile",
        action="store_true",
        help="Use Torch Profiler. The endpoint must be launched with "
        "VLLM_TORCH_PROFILER_DIR to enable profiler.",
    )
    parser.add_argument(
        "--save-result",
        action="store_true",
        help="Specify to save benchmark results to a json file",
    )
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
    parser.add_argument(
        "--save-detailed",
        action="store_true",
        help="When saving the results, whether to include per request "
        "information such as response, error, ttfs, tpots, etc.",
    )
    parser.add_argument(
        "--append-result",
        action="store_true",
        help="Append the benchmark result to the existing json file.",
    )
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
    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."
1082
1083
        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.",
    )
1084
1085
1086
    parser.add_argument(
        "--percentile-metrics",
        type=str,
1087
        default=None,
1088
        help="Comma-separated list of selected metrics to report percentils. "
1089
        "This argument specifies the metrics to report percentiles. "
1090
1091
1092
        'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
        'If not specified, defaults to "ttft,tpot,itl" for generative models '
        'and "e2el" for pooling models.',
1093
    )
1094
1095
1096
1097
    parser.add_argument(
        "--metric-percentiles",
        type=str,
        default="99",
1098
        help="Comma-separated list of percentiles for selected metrics. "
1099
1100
1101
        'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". '
        'Default value is "99".'
        'Use "--percentile-metrics" to select metrics.',
1102
1103
1104
1105
1106
    )
    parser.add_argument(
        "--goodput",
        nargs="+",
        required=False,
1107
        help='Specify service level objectives for goodput as "KEY:VALUE" '
1108
        "pairs, where the key is a metric name, and the value is in "
1109
        'milliseconds. Multiple "KEY:VALUE" pairs can be provided, '
1110
        "separated by spaces. Allowed request level metric names are "
1111
        '"ttft", "tpot", "e2el". For more context on the definition of '
1112
        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
1113
1114
        "and the blog: https://hao-ai-lab.github.io/blogs/distserve",
    )
1115
1116
1117
1118
1119
1120
1121
1122
    parser.add_argument(
        "--request-id-prefix",
        type=str,
        required=False,
        default="benchmark-serving",
        help="Specify the prefix of request id.",
    )

1123
1124
1125
1126
1127
    sampling_group = parser.add_argument_group("sampling parameters")
    sampling_group.add_argument(
        "--top-p",
        type=float,
        default=None,
1128
        help="Top-p sampling parameter. Only has effect on openai-compatible backends.",
1129
1130
1131
1132
1133
    )
    sampling_group.add_argument(
        "--top-k",
        type=int,
        default=None,
1134
        help="Top-k sampling parameter. Only has effect on openai-compatible backends.",
1135
1136
1137
1138
1139
    )
    sampling_group.add_argument(
        "--min-p",
        type=float,
        default=None,
1140
        help="Min-p sampling parameter. Only has effect on openai-compatible backends.",
1141
1142
1143
1144
1145
1146
1147
1148
1149
    )
    sampling_group.add_argument(
        "--temperature",
        type=float,
        default=None,
        help="Temperature sampling parameter. Only has effect on "
        "openai-compatible backends. If not specified, default to greedy "
        "decoding (i.e. temperature==0.0).",
    )
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
    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.",
    )
1171

1172
    parser.add_argument(
1173
        "--tokenizer-mode",
1174
1175
        type=str,
        default="auto",
1176
        choices=["auto", "slow", "mistral", "custom"],
1177
1178
        help='The tokenizer mode.\n\n* "auto" will use the '
        'fast tokenizer if available.\n* "slow" will '
1179
        "always use the slow tokenizer. \n* "
1180
        '"mistral" will always use the `mistral_common` tokenizer. \n*'
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
        '"custom" will use --tokenizer to select the preregistered tokenizer.',
    )

    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 "
        "same as the ``--model`` argument. ",
    )

    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.",
    )
1201

1202
1203
1204
1205
1206
1207
1208
1209
    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 "
1210
1211
        "--ramp-up-end-rps.) over the duration of the benchmark.",
    )
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
    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.",
    )
1226
1227
1228
1229
1230
    parser.add_argument(
        "--ready-check-timeout-sec",
        type=int,
        default=600,
        help="Maximum time to wait for the endpoint to become ready "
1231
        "in seconds (default: 600 seconds / 10 minutes). If set to 0, "
1232
        "the ready check will be skipped.",
1233
    )
1234

1235
1236
1237
1238
1239
1240
1241
1242
1243
    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,
    )

1244

1245
1246
1247
def main(args: argparse.Namespace) -> dict[str, Any]:
    return asyncio.run(main_async(args))

1248

1249
async def main_async(args: argparse.Namespace) -> dict[str, Any]:
1250
1251
1252
1253
    print(args)
    random.seed(args.seed)
    np.random.seed(args.seed)

1254
1255
1256
1257
1258
1259
    # 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. "
1260
1261
                "Please remove the --request-rate argument."
            )
1262
1263
1264
        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 "
1265
1266
                "--ramp-up-end-rps must be specified"
            )
1267
1268
1269
1270
        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")
1271
1272
        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.")
1273

1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
    label = args.label
    model_id = args.model
    model_name = args.served_model_name
    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
    tokenizer_mode = args.tokenizer_mode

    if args.base_url is not None:
        api_url = f"{args.base_url}{args.endpoint}"
        base_url = f"{args.base_url}"
    else:
        api_url = f"http://{args.host}:{args.port}{args.endpoint}"
        base_url = f"http://{args.host}:{args.port}"

1287
1288
1289
1290
1291
1292
1293
1294
1295
    # 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:
1296
                raise ValueError("Invalid header format. Please use KEY=VALUE format.")
1297

1298
1299
1300
1301
1302
    tokenizer = get_tokenizer(
        tokenizer_id,
        tokenizer_mode=tokenizer_mode,
        trust_remote_code=args.trust_remote_code,
    )
1303

1304
1305
1306
    if args.dataset_name is None:
        raise ValueError(
            "Please specify '--dataset-name' and the corresponding "
1307
1308
            "'--dataset-path' if required."
        )
1309

1310
1311
    # Load the dataset.
    input_requests = get_samples(args, tokenizer)
1312
1313
    goodput_config_dict = check_goodput_args(args)

1314
    backend = args.backend
1315
1316
1317
1318
1319
    task_type = (
        TaskType.POOLING
        if "embeddings" in backend or "rerank" in backend
        else TaskType.GENERATION
    )
1320

1321
    # Collect the sampling parameters.
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
    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."
            )
1342

1343
1344
        if "temperature" not in sampling_params:
            sampling_params["temperature"] = 0.0  # Default to greedy decoding.
1345
1346

        default_percentile_metrics = "ttft,tpot,itl"
1347
1348
    else:
        sampling_params = {}
1349
        default_percentile_metrics = "e2el"
1350

1351
1352
1353
    extra_body = args.extra_body or {}
    extra_body = {**sampling_params, **extra_body}

1354
1355
    percentile_metrics: str = args.percentile_metrics or default_percentile_metrics

1356
1357
1358
1359
    # Avoid GC processing "static" data - reduce pause times.
    gc.collect()
    gc.freeze()

1360
    benchmark_result = await benchmark(
1361
1362
        task_type=task_type,
        endpoint_type=backend,
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
        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,
        profile=args.profile,
1374
        selected_percentile_metrics=percentile_metrics.split(","),
1375
        selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")],
1376
1377
1378
1379
        ignore_eos=args.ignore_eos,
        goodput_config_dict=goodput_config_dict,
        max_concurrency=args.max_concurrency,
        lora_modules=args.lora_modules,
1380
        extra_headers=headers,
1381
        extra_body=extra_body,
1382
1383
1384
1385
1386
        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,
    )
1387
1388

    # Save config and results to json
1389
1390
1391
1392
1393
    result_json: dict[str, Any] = {}

    # Setup
    current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
    result_json["date"] = current_dt
1394
    result_json["endpoint_type"] = args.backend  # for backward compatibility
1395
    result_json["backend"] = args.backend
1396
1397
1398
1399
1400
1401
1402
1403
1404
    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:
1405
                kvstring = item.split("=", 1)
1406
1407
1408
                result_json[kvstring[0].strip()] = kvstring[1].strip()
            else:
                raise ValueError(
1409
1410
                    "Invalid metadata format. Please use KEY=VALUE format."
                )
1411

1412
    # Traffic
1413
1414
1415
    result_json["request_rate"] = (
        args.request_rate if args.request_rate < float("inf") else "inf"
    )
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
    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 [
1430
1431
1432
1433
1434
1435
            "input_lens",
            "output_lens",
            "ttfts",
            "itls",
            "generated_texts",
            "errors",
1436
1437
1438
1439
1440
        ]:
            if field in result_json:
                del result_json[field]
            if field in benchmark_result:
                del benchmark_result[field]
1441

1442
        # Save to file
1443
    if args.save_result or args.append_result:
1444
        base_model_id = model_id.split("/")[-1]
1445
1446
1447
1448
1449
        max_concurrency_str = (
            f"-concurrency{args.max_concurrency}"
            if args.max_concurrency is not None
            else ""
        )
1450
        label = label or args.backend
1451
        if args.ramp_up_strategy is not None:
1452
            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
1453
1454
        else:
            file_name = f"{label}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
1455
1456
1457
        if args.result_filename:
            file_name = args.result_filename
        if args.result_dir:
1458
            os.makedirs(args.result_dir, exist_ok=True)
1459
            file_name = os.path.join(args.result_dir, file_name)
1460
1461
1462
        with open(
            file_name, mode="a+" if args.append_result else "w", encoding="utf-8"
        ) as outfile:
1463
1464
1465
            # Append a newline.
            if args.append_result and outfile.tell() != 0:
                outfile.write("\n")
1466
1467
            json.dump(result_json, outfile)
        save_to_pytorch_benchmark_format(args, result_json, file_name)
1468

1469
    return result_json