serve.py 53.5 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
@dataclass
class BenchmarkMetrics:
    completed: int
67
    failed: int
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
    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]]
93
94
95
96
    # Max output tokens per second and concurrent requests at that peak
    max_output_tokens_per_s: float
    max_concurrent_requests: int

97

98
99
100
@dataclass
class EmbedBenchmarkMetrics:
    completed: int
101
    failed: int
102
103
    total_input: int
    request_throughput: float
104
    total_token_throughput: float
105
106
107
108
    mean_e2el_ms: float
    std_e2el_ms: float
    median_e2el_ms: float
    percentiles_e2el_ms: float
109

110

111
def _get_current_request_rate(
112
113
114
    ramp_up_strategy: Literal["linear", "exponential"] | None,
    ramp_up_start_rps: int | None,
    ramp_up_end_rps: int | None,
115
116
117
118
    request_index: int,
    total_requests: int,
    request_rate: float,
) -> float:
119
120
121
122
123
    if (
        ramp_up_strategy
        and ramp_up_start_rps is not None
        and ramp_up_end_rps is not None
    ):
124
125
126
127
128
129
130
131
132
133
134
135
        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


136
async def get_request(
137
    input_requests: list[SampleRequest],
138
139
    request_rate: float,
    burstiness: float = 1.0,
140
141
142
    ramp_up_strategy: Literal["linear", "exponential"] | None = None,
    ramp_up_start_rps: int | None = None,
    ramp_up_end_rps: int | None = None,
143
) -> AsyncGenerator[tuple[SampleRequest, float], None]:
144
145
    """
    Asynchronously generates requests at a specified rate
146
    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
147
148
149

    Args:
        input_requests:
150
            A list of input requests, each represented as a SampleRequest.
151
152
153
154
155
156
157
158
159
160
        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.
161
        ramp_up_strategy (optional):
162
163
164
165
166
167
            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.
168
169
    """
    assert burstiness > 0, (
170
171
        f"A positive burstiness factor is expected, but given {burstiness}."
    )
172
    # Convert to list to get length for ramp-up calculations
173
    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
174
        input_requests = list(input_requests)
175

176
    total_requests = len(input_requests)
177
    assert total_requests > 0, "No requests provided."
178

179
180
181
182
    # Precompute delays among requests to minimize request send laggings
    request_rates = []
    delay_ts = []
    for request_index, request in enumerate(input_requests):
183
        current_request_rate = _get_current_request_rate(
184
185
186
187
188
189
190
            ramp_up_strategy,
            ramp_up_start_rps,
            ramp_up_end_rps,
            request_index,
            total_requests,
            request_rate,
        )
191
        request_rates.append(current_request_rate)
192
        if current_request_rate == float("inf"):
193
194
195
196
197
198
199
            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))
200

201
202
203
204
205
206
207
208
209
    # 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.
        #
210
211
        # NOTE: If we simply accumulate the random delta values
        # from the gamma distribution, their sum would have 1-2% gap
212
        # from target_total_delay_s. The purpose of the following logic is to
co63oc's avatar
co63oc committed
213
        # close the gap for stabilizing the throughput data
214
        # from different random seeds.
215
216
217
218
219
220
        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):
221
222
223
224
225
        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)
226
        yield request, request_rates[request_index]
227
228


229
def calculate_metrics_for_embeddings(
230
231
    outputs: list[RequestFuncOutput], dur_s: float, selected_percentiles: list[float]
) -> EmbedBenchmarkMetrics:
232
233
234
235
236
237
238
239
240
241
242
243
    """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
244
    failed = 0
245
246
247
248
249
250
    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
251
252
        else:
            failed += 1
253
254
255
256
257

    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
            "on the benchmark arguments.",
258
259
            stacklevel=2,
        )
260
261
    metrics = EmbedBenchmarkMetrics(
        completed=completed,
262
        failed=failed,
263
264
265
266
267
268
        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,
269
270
271
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
272
273
274
275
    )
    return metrics


276
def calculate_metrics(
277
    input_requests: list[SampleRequest],
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
304
305
306
307
308
309
    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

310
            if not output_len:
311
312
313
314
315
316
                # 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(
317
318
319
320
                    tokenizer(
                        outputs[i].generated_text, add_special_tokens=False
                    ).input_ids
                )
321
            actual_output_lens.append(output_len)
322
            total_input += input_requests[i].prompt_len
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
            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)
343
344
345
            slo_values.append(
                goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
346
347
        if "tpot" in goodput_config_dict:
            valid_metrics.append(all_tpots)
348
349
350
            slo_values.append(
                goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
351
352
        if "e2el" in goodput_config_dict:
            valid_metrics.append(e2els)
353
354
355
            slo_values.append(
                goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
356
357
358
359
360
361
362
363
364
365

        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.",
366
367
            stacklevel=2,
        )
368
369
370
371
372
373
374

    # 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]
375
    failed_outputs = [output for output in outputs if not output.success]
376
    if successful_outputs:
377
378
379
380
        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
        )
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403

        # 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)
404
405
406
            request_end_second = int(
                (output.start_time + output.latency) - min_start_time
            )
407
408
409
410
411
412
413
414

            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))
415
            max_concurrent_requests = int(np.max(concurrent_requests_per_second))
416
417
418

        if TERM_PLOTLIB_AVAILABLE:
            import termplotlib as tpl
419

420
            fig = tpl.figure()
421
422
423
424
425
426
427
428
429
430
            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",
            )
431
432
433
434
            fig.show()
        else:
            print("tip: install termplotlib and gnuplot to plot the metrics")

435
436
    metrics = BenchmarkMetrics(
        completed=completed,
437
        failed=len(failed_outputs),
438
439
440
441
442
443
        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,
444
445
        mean_ttft_ms=np.mean(ttfts or 0)
        * 1000,  # ttfts is empty if streaming is not supported by the endpoint
446
447
        std_ttft_ms=np.std(ttfts or 0) * 1000,
        median_ttft_ms=np.median(ttfts or 0) * 1000,
448
449
450
        percentiles_ttft_ms=[
            (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles
        ],
451
452
453
        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,
454
455
456
        percentiles_tpot_ms=[
            (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles
        ],
457
458
459
        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,
460
461
462
        percentiles_itl_ms=[
            (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles
        ],
463
464
465
        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,
466
467
468
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
469
470
        max_output_tokens_per_s=max_output_tokens_per_s,
        max_concurrent_requests=max_concurrent_requests,
471
472
473
474
475
476
    )

    return metrics, actual_output_lens


async def benchmark(
477
    task_type: TaskType,
478
479
480
481
482
483
    endpoint_type: str,
    api_url: str,
    base_url: str,
    model_id: str,
    model_name: str,
    tokenizer: PreTrainedTokenizerBase,
484
    input_requests: list[SampleRequest],
485
    logprobs: int | None,
486
487
488
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
489
    num_warmups: int,
490
491
    profile: bool,
    selected_percentile_metrics: list[str],
492
    selected_percentiles: list[float],
493
494
    ignore_eos: bool,
    goodput_config_dict: dict[str, float],
495
496
497
498
499
500
501
    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,
502
    ready_check_timeout_sec: int = 600,
503
):
504
505
506
507
    try:
        request_func = ASYNC_REQUEST_FUNCS[endpoint_type]
    except KeyError:
        raise ValueError(f"Unknown backend: {endpoint_type}") from None
508

509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
    # 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),
    )

527
528
    print("Starting initial single prompt test run...")
    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
529
530
531
532
533
534
        input_requests[0].prompt,
        input_requests[0].prompt_len,
        input_requests[0].expected_output_len,
        input_requests[0].multi_modal_data,
    )

535
536
537
538
539
540
541
542
    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]"
543
544
545
546
547
548
549
550
551
552
    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,
553
        extra_headers=extra_headers,
554
        extra_body=extra_body,
555
556
    )

557
558
559
560
561
562
563
564
565
566
567
    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. "
568
569
                f"Error: {test_output.error}"
            )
570
        else:
571
            print("Initial test run completed.")
572
    else:
573
        print("Skipping endpoint ready check.")
574

575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
    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...")

602
603
604
    if lora_modules:
        # For each input request, choose a LoRA module at random.
        lora_modules = iter(
605
606
            [random.choice(lora_modules) for _ in range(len(input_requests))]
        )
607
608
609

    if profile:
        print("Starting profiler...")
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
        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
        )
626
627
628
        if profile_output.success:
            print("Profiler started")

629
    distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
630
631
632

    if ramp_up_strategy is not None:
        print(f"Traffic ramp-up strategy: {ramp_up_strategy}.")
633
634
635
636
        print(
            f"Will increase RPS from {ramp_up_start_rps} to "
            f"{ramp_up_end_rps} RPS over the duration of the benchmark."
        )
637
    else:
638
        print(f"Traffic request rate: {request_rate}")
639
640
641
642
643
644

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

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

645
646
647
648
649
    semaphore = (
        asyncio.Semaphore(max_concurrency)
        if max_concurrency
        else contextlib.nullcontext()
    )
650

651
    async def limited_request_func(request_func_input, session, pbar):
652
        async with semaphore:
653
654
655
            return await request_func(
                request_func_input=request_func_input, session=session, pbar=pbar
            )
656
657
658

    benchmark_start_time = time.perf_counter()
    tasks: list[asyncio.Task] = []
659
660
661
662
663

    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
664
665
666
667
668
669
        rps_change_events.append(
            {
                "rps": last_int_rps,
                "timestamp": datetime.now().isoformat(),
            }
        )
670
671

    async for request, current_request_rate in get_request(
672
673
674
675
676
677
678
        input_requests,
        request_rate,
        burstiness,
        ramp_up_strategy,
        ramp_up_start_rps,
        ramp_up_end_rps,
    ):
679
680
681
682
683
        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):
684
                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
685
                last_int_rps = current_int_rps
686
        prompt, prompt_len, output_len, mm_content, request_id = (
687
688
689
690
            request.prompt,
            request.prompt_len,
            request.expected_output_len,
            request.multi_modal_data,
691
            request.request_id,
692
        )
693
694
695
696
697
        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

698
699
700
701
702
703
704
705
706
707
708
709
710
711
        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,
        )
712
713
        tasks.append(
            asyncio.create_task(
714
715
716
717
718
                limited_request_func(
                    request_func_input=request_func_input, session=session, pbar=pbar
                )
            )
        )
719
720
721
722
723
724
725
    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)

    if pbar is not None:
        pbar.close()

    benchmark_duration = time.perf_counter() - benchmark_start_time

726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
    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
742

743
    print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
744
    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
745
    print("{:<40} {:<10}".format("Failed requests:", metrics.failed))
746
    if max_concurrency is not None:
747
748
749
750
        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))
751
    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
752
    if isinstance(metrics, BenchmarkMetrics):
753
754
755
756
757
758
        print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
    print(
        "{:<40} {:<10.2f}".format(
            "Request throughput (req/s):", metrics.request_throughput
        )
    )
759
    if goodput_config_dict:
760
761
762
763
764
        print(
            "{:<40} {:<10.2f}".format(
                "Request goodput (req/s):", metrics.request_goodput
            )
        )
765
    if isinstance(metrics, BenchmarkMetrics):
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
        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
        )
    )
786

787
788
789
790
    if isinstance(metrics, BenchmarkMetrics):
        result = {
            "duration": benchmark_duration,
            "completed": metrics.completed,
791
            "failed": metrics.failed,
792
793
794
            "total_input_tokens": metrics.total_input,
            "total_output_tokens": metrics.total_output,
            "request_throughput": metrics.request_throughput,
795
            "request_goodput": metrics.request_goodput if goodput_config_dict else None,
796
797
798
799
800
801
802
803
            "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],
804
805
            "max_output_tokens_per_s": metrics.max_output_tokens_per_s,
            "max_concurrent_requests": metrics.max_concurrent_requests,
806
807
808
809
810
811
812
813
814
815
816
        }
    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],
        }
817

818
819
820
    if rps_change_events:
        result["rps_change_events"] = rps_change_events

821
822
823
824
825
826
827
828
829
830
831
832
    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
833
834
835
836
837
838
839
840
841
842
843
844
845
        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"),
            )
        )
846
        result[f"mean_{metric_attribute_name}_ms"] = getattr(
847
848
            metrics, f"mean_{metric_attribute_name}_ms"
        )
849
        result[f"median_{metric_attribute_name}_ms"] = getattr(
850
851
            metrics, f"median_{metric_attribute_name}_ms"
        )
852
        result[f"std_{metric_attribute_name}_ms"] = getattr(
853
854
855
            metrics, f"std_{metric_attribute_name}_ms"
        )
        for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"):
856
            p_word = str(int(p)) if int(p) == p else str(p)
857
            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value))
858
859
            result[f"p{p_word}_{metric_attribute_name}_ms"] = value

860
861
    if task_type == TaskType.GENERATION:
        process_one_metric("ttft", "TTFT", "Time to First Token")
862
        process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)")
863
        process_one_metric("itl", "ITL", "Inter-token Latency")
864
865
866
867
    process_one_metric("e2el", "E2EL", "End-to-end Latency")

    print("=" * 50)

868
869
870
871
872
873
874
875
876
877
    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,
        )
878
879
880
        profile_output = await request_func(
            request_func_input=profile_input, session=session
        )
881
882
        if profile_output.success:
            print("Profiler stopped")
883
884

    await session.close()
885
886
887
888
889
890
891
892
893
894
895
896
897
898
    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 "
899
900
                    f"{str(VALID_NAMES)}. "
                )
901
902
903
904
            if slo_val < 0:
                raise ValueError(
                    f"Invalid value found, {slo_name}: {slo_val}. "
                    "The service level objective value should be "
905
906
                    "non-negative."
                )
907
908
909
910
911
912
913
914
915
916
917
918
    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. "
919
            'Specify service level objectives for goodput as "KEY:VALUE" '
920
            "pairs, where the key is a metric name, and the value is a "
921
922
            "number in milliseconds."
        ) from err
923
924
925
    return goodput_config_dict


926
927
928
def save_to_pytorch_benchmark_format(
    args: argparse.Namespace, results: dict[str, Any], file_name: str
) -> None:
929
    metrics = [
930
931
932
933
934
935
936
937
938
939
940
941
        "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",
942
943
944
945
946
947
    ]
    # 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,
948
        metrics={k: [results[k]] for k in metrics if k in results},
949
950
        extra_info={
            k: results[k]
951
952
953
954
            for k in results
            if k not in metrics and k not in ignored_metrics
        },
    )
955
956
957
958
959
960
961
    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):
962
    add_dataset_parser(parser)
963
964
965
966
967
    parser.add_argument(
        "--label",
        type=str,
        default=None,
        help="The label (prefix) of the benchmark results. If not specified, "
968
        "the value of '--backend' will be used as the label.",
969
    )
970
971
972
    parser.add_argument(
        "--backend",
        type=str,
973
974
        default="openai",
        choices=list(ASYNC_REQUEST_FUNCS.keys()),
975
        help="The type of backend or endpoint to use for the benchmark.",
976
    )
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
    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.",
    )
992
993
994
995
996
    parser.add_argument(
        "--header",
        metavar="KEY=VALUE",
        nargs="*",
        help="Key-value pairs (e.g, --header x-additional-info=0.3.3) "
997
998
999
        "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).",
1000
    )
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
    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, "
1012
1013
        "if the server is not processing requests fast enough to keep up.",
    )
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023

    parser.add_argument(
        "--model",
        type=str,
        required=True,
        help="Name of the model.",
    )
    parser.add_argument(
        "--tokenizer",
        type=str,
1024
        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
1025
1026
1027
1028
1029
1030
    )
    parser.add_argument("--use-beam-search", action="store_true")
    parser.add_argument(
        "--logprobs",
        type=int,
        default=None,
1031
1032
1033
1034
1035
1036
1037
        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"
        ),
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
    )
    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.",
    )
1070
1071
1072
1073
1074
1075
    parser.add_argument(
        "--num-warmups",
        type=int,
        default=0,
        help="Number of warmup requests.",
    )
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
    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",
    )
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
    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.",
    )
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
    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."
1126
1127
        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.",
    )
1128
1129
1130
    parser.add_argument(
        "--percentile-metrics",
        type=str,
1131
        default=None,
1132
        help="Comma-separated list of selected metrics to report percentils. "
1133
        "This argument specifies the metrics to report percentiles. "
1134
1135
1136
        'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
        'If not specified, defaults to "ttft,tpot,itl" for generative models '
        'and "e2el" for pooling models.',
1137
    )
1138
1139
1140
1141
    parser.add_argument(
        "--metric-percentiles",
        type=str,
        default="99",
1142
        help="Comma-separated list of percentiles for selected metrics. "
1143
1144
1145
        'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". '
        'Default value is "99".'
        'Use "--percentile-metrics" to select metrics.',
1146
1147
1148
1149
1150
    )
    parser.add_argument(
        "--goodput",
        nargs="+",
        required=False,
1151
        help='Specify service level objectives for goodput as "KEY:VALUE" '
1152
        "pairs, where the key is a metric name, and the value is in "
1153
        'milliseconds. Multiple "KEY:VALUE" pairs can be provided, '
1154
        "separated by spaces. Allowed request level metric names are "
1155
        '"ttft", "tpot", "e2el". For more context on the definition of '
1156
        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
1157
1158
        "and the blog: https://hao-ai-lab.github.io/blogs/distserve",
    )
1159
1160
1161
1162
1163
1164
1165
1166
    parser.add_argument(
        "--request-id-prefix",
        type=str,
        required=False,
        default="benchmark-serving",
        help="Specify the prefix of request id.",
    )

1167
1168
1169
1170
1171
    sampling_group = parser.add_argument_group("sampling parameters")
    sampling_group.add_argument(
        "--top-p",
        type=float,
        default=None,
1172
        help="Top-p sampling parameter. Only has effect on openai-compatible backends.",
1173
1174
1175
1176
1177
    )
    sampling_group.add_argument(
        "--top-k",
        type=int,
        default=None,
1178
        help="Top-k sampling parameter. Only has effect on openai-compatible backends.",
1179
1180
1181
1182
1183
    )
    sampling_group.add_argument(
        "--min-p",
        type=float,
        default=None,
1184
        help="Min-p sampling parameter. Only has effect on openai-compatible backends.",
1185
1186
1187
1188
1189
1190
1191
1192
1193
    )
    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).",
    )
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
    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.",
    )
1215

1216
    parser.add_argument(
1217
        "--tokenizer-mode",
1218
1219
        type=str,
        default="auto",
1220
        choices=["auto", "slow", "mistral", "custom"],
1221
1222
        help='The tokenizer mode.\n\n* "auto" will use the '
        'fast tokenizer if available.\n* "slow" will '
1223
        "always use the slow tokenizer. \n* "
1224
        '"mistral" will always use the `mistral_common` tokenizer. \n*'
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
        '"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.",
    )
1245

1246
1247
1248
1249
1250
1251
1252
1253
    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 "
1254
1255
        "--ramp-up-end-rps.) over the duration of the benchmark.",
    )
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
    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.",
    )
1270
1271
1272
1273
1274
    parser.add_argument(
        "--ready-check-timeout-sec",
        type=int,
        default=600,
        help="Maximum time to wait for the endpoint to become ready "
1275
        "in seconds (default: 600 seconds / 10 minutes). If set to 0, "
1276
        "the ready check will be skipped.",
1277
    )
1278

1279
1280
1281
1282
1283
1284
1285
1286
1287
    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,
    )

1288

1289
1290
1291
def main(args: argparse.Namespace) -> dict[str, Any]:
    return asyncio.run(main_async(args))

1292

1293
async def main_async(args: argparse.Namespace) -> dict[str, Any]:
1294
1295
1296
1297
    print(args)
    random.seed(args.seed)
    np.random.seed(args.seed)

1298
1299
1300
1301
1302
1303
    # 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. "
1304
1305
                "Please remove the --request-rate argument."
            )
1306
1307
1308
        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 "
1309
1310
                "--ramp-up-end-rps must be specified"
            )
1311
1312
1313
1314
        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")
1315
1316
        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.")
1317

1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
    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}"

1331
1332
1333
1334
1335
1336
1337
1338
1339
    # 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:
1340
                raise ValueError("Invalid header format. Please use KEY=VALUE format.")
1341

1342
1343
1344
1345
1346
    tokenizer = get_tokenizer(
        tokenizer_id,
        tokenizer_mode=tokenizer_mode,
        trust_remote_code=args.trust_remote_code,
    )
1347

1348
1349
1350
    if args.dataset_name is None:
        raise ValueError(
            "Please specify '--dataset-name' and the corresponding "
1351
1352
            "'--dataset-path' if required."
        )
1353

1354
1355
    # Load the dataset.
    input_requests = get_samples(args, tokenizer)
1356
1357
    goodput_config_dict = check_goodput_args(args)

1358
    backend = args.backend
1359
1360
1361
1362
1363
    task_type = (
        TaskType.POOLING
        if "embeddings" in backend or "rerank" in backend
        else TaskType.GENERATION
    )
1364

1365
    # Collect the sampling parameters.
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
    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."
            )
1386

1387
1388
        if "temperature" not in sampling_params:
            sampling_params["temperature"] = 0.0  # Default to greedy decoding.
1389
1390

        default_percentile_metrics = "ttft,tpot,itl"
1391
1392
    else:
        sampling_params = {}
1393
        default_percentile_metrics = "e2el"
1394

1395
1396
1397
    extra_body = args.extra_body or {}
    extra_body = {**sampling_params, **extra_body}

1398
1399
    percentile_metrics: str = args.percentile_metrics or default_percentile_metrics

1400
1401
1402
1403
    # Avoid GC processing "static" data - reduce pause times.
    gc.collect()
    gc.freeze()

1404
    benchmark_result = await benchmark(
1405
1406
        task_type=task_type,
        endpoint_type=backend,
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
        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,
1417
        num_warmups=args.num_warmups,
1418
        profile=args.profile,
1419
        selected_percentile_metrics=percentile_metrics.split(","),
1420
        selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")],
1421
1422
1423
1424
        ignore_eos=args.ignore_eos,
        goodput_config_dict=goodput_config_dict,
        max_concurrency=args.max_concurrency,
        lora_modules=args.lora_modules,
1425
        extra_headers=headers,
1426
        extra_body=extra_body,
1427
1428
1429
1430
1431
        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,
    )
1432
1433

    # Save config and results to json
1434
1435
1436
1437
1438
    result_json: dict[str, Any] = {}

    # Setup
    current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
    result_json["date"] = current_dt
1439
    result_json["endpoint_type"] = args.backend  # for backward compatibility
1440
    result_json["backend"] = args.backend
1441
1442
1443
1444
1445
1446
1447
1448
1449
    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:
1450
                kvstring = item.split("=", 1)
1451
1452
1453
                result_json[kvstring[0].strip()] = kvstring[1].strip()
            else:
                raise ValueError(
1454
1455
                    "Invalid metadata format. Please use KEY=VALUE format."
                )
1456

1457
    # Traffic
1458
1459
1460
    result_json["request_rate"] = (
        args.request_rate if args.request_rate < float("inf") else "inf"
    )
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
    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 [
1475
1476
1477
1478
1479
1480
            "input_lens",
            "output_lens",
            "ttfts",
            "itls",
            "generated_texts",
            "errors",
1481
1482
1483
1484
1485
        ]:
            if field in result_json:
                del result_json[field]
            if field in benchmark_result:
                del benchmark_result[field]
1486

1487
        # Save to file
1488
    if args.save_result or args.append_result:
1489
        base_model_id = model_id.split("/")[-1]
1490
1491
1492
1493
1494
        max_concurrency_str = (
            f"-concurrency{args.max_concurrency}"
            if args.max_concurrency is not None
            else ""
        )
1495
        label = label or args.backend
1496
        if args.ramp_up_strategy is not None:
1497
            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
1498
1499
        else:
            file_name = f"{label}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
1500
1501
1502
        if args.result_filename:
            file_name = args.result_filename
        if args.result_dir:
1503
            os.makedirs(args.result_dir, exist_ok=True)
1504
            file_name = os.path.join(args.result_dir, file_name)
1505
1506
1507
        with open(
            file_name, mode="a+" if args.append_result else "w", encoding="utf-8"
        ) as outfile:
1508
1509
1510
            # Append a newline.
            if args.append_result and outfile.tell() != 0:
                outfile.write("\n")
1511
1512
            json.dump(result_json, outfile)
        save_to_pytorch_benchmark_format(args, result_json, file_name)
1513

1514
    return result_json