benchmark_serving.py 45.6 KB
Newer Older
zhuwenwen's avatar
zhuwenwen committed
1
# SPDX-License-Identifier: Apache-2.0
zhuwenwen's avatar
zhuwenwen committed
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
zhuwenwen's avatar
zhuwenwen committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
r"""Benchmark online serving throughput.

On the server side, run one of the following commands:
    vLLM OpenAI API server
    vllm serve <your_model> \
        --swap-space 16 \
        --disable-log-requests

On the client side, run:
    python benchmarks/benchmark_serving.py \
        --backend <backend> \
        --model <your_model> \
        --dataset-name sharegpt \
        --dataset-path <path to dataset> \
        --request-rate <request_rate> \ # By default <request_rate> is inf
        --num-prompts <num_prompts> # By default <num_prompts> is 1000

    when using tgi backend, add
        --endpoint /generate_stream
    to the end of the command above.
"""
zhuwenwen's avatar
zhuwenwen committed
24

zhuwenwen's avatar
zhuwenwen committed
25
26
27
28
29
30
31
32
import argparse
import asyncio
import gc
import json
import os
import random
import time
import warnings
33
from collections.abc import Iterable
zhuwenwen's avatar
zhuwenwen committed
34
35
from dataclasses import dataclass
from datetime import datetime
zhuwenwen's avatar
zhuwenwen committed
36
from typing import Any, Literal, Optional
zhuwenwen's avatar
zhuwenwen committed
37
38
39
40
41

import numpy as np
from tqdm.asyncio import tqdm
from transformers import PreTrainedTokenizerBase

zhuwenwen's avatar
zhuwenwen committed
42
43
44
45
46
47
48
from backend_request_func import (
    ASYNC_REQUEST_FUNCS,
    OPENAI_COMPATIBLE_BACKENDS,
    RequestFuncInput,
    RequestFuncOutput,
)

zhuwenwen's avatar
zhuwenwen committed
49
50
51
52
53
54
55
56
57
58
try:
    from vllm.transformers_utils.tokenizer import get_tokenizer
except ImportError:
    from backend_request_func import get_tokenizer

try:
    from vllm.utils import FlexibleArgumentParser
except ImportError:
    from argparse import ArgumentParser as FlexibleArgumentParser

zhuwenwen's avatar
zhuwenwen committed
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
from benchmark_dataset import (
    AIMODataset,
    ASRDataset,
    BurstGPTDataset,
    ConversationDataset,
    CustomDataset,
    HuggingFaceDataset,
    InstructCoderDataset,
    MTBenchDataset,
    NextEditPredictionDataset,
    RandomDataset,
    SampleRequest,
    ShareGPTDataset,
    SonnetDataset,
    VisionArenaDataset,
)
zhuwenwen's avatar
zhuwenwen committed
75
from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
76
from vllm.benchmarks.serve import get_request
zhuwenwen's avatar
zhuwenwen committed
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92

MILLISECONDS_TO_SECONDS_CONVERSION = 1000


@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
zhuwenwen's avatar
zhuwenwen committed
93
    percentiles_ttft_ms: list[tuple[float, float]]
zhuwenwen's avatar
zhuwenwen committed
94
95
96
    mean_tpot_ms: float
    median_tpot_ms: float
    std_tpot_ms: float
zhuwenwen's avatar
zhuwenwen committed
97
    percentiles_tpot_ms: list[tuple[float, float]]
zhuwenwen's avatar
zhuwenwen committed
98
99
100
    mean_itl_ms: float
    median_itl_ms: float
    std_itl_ms: float
zhuwenwen's avatar
zhuwenwen committed
101
    percentiles_itl_ms: list[tuple[float, float]]
zhuwenwen's avatar
zhuwenwen committed
102
103
104
105
106
107
    # 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
zhuwenwen's avatar
zhuwenwen committed
108
    percentiles_e2el_ms: list[tuple[float, float]]
zhuwenwen's avatar
zhuwenwen committed
109
110
111


def calculate_metrics(
zhuwenwen's avatar
zhuwenwen committed
112
113
    input_requests: list[SampleRequest],
    outputs: list[RequestFuncOutput],
zhuwenwen's avatar
zhuwenwen committed
114
115
    dur_s: float,
    tokenizer: PreTrainedTokenizerBase,
zhuwenwen's avatar
zhuwenwen committed
116
117
118
119
120
    selected_percentile_metrics: list[str],
    selected_percentiles: list[float],
    goodput_config_dict: dict[str, float],
) -> tuple[BenchmarkMetrics, list[int]]:
    actual_output_lens: list[int] = []
zhuwenwen's avatar
zhuwenwen committed
121
122
123
    total_input = 0
    completed = 0
    good_completed = 0
zhuwenwen's avatar
zhuwenwen committed
124
125
126
127
128
    itls: list[float] = []
    tpots: list[float] = []
    all_tpots: list[float] = []
    ttfts: list[float] = []
    e2els: list[float] = []
zhuwenwen's avatar
zhuwenwen committed
129
130
131
132
    for i in range(len(outputs)):
        if outputs[i].success:
            output_len = outputs[i].output_tokens

zhuwenwen's avatar
zhuwenwen committed
133
            if not output_len:
zhuwenwen's avatar
zhuwenwen committed
134
135
136
137
138
139
                # 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(
zhuwenwen's avatar
zhuwenwen committed
140
141
142
143
                    tokenizer(
                        outputs[i].generated_text, add_special_tokens=False
                    ).input_ids
                )
zhuwenwen's avatar
zhuwenwen committed
144
            actual_output_lens.append(output_len)
zhuwenwen's avatar
zhuwenwen committed
145
            total_input += input_requests[i].prompt_len
zhuwenwen's avatar
zhuwenwen committed
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
            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)
zhuwenwen's avatar
zhuwenwen committed
166
167
168
            slo_values.append(
                goodput_config_dict["ttft"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
zhuwenwen's avatar
zhuwenwen committed
169
170
        if "tpot" in goodput_config_dict:
            valid_metrics.append(all_tpots)
zhuwenwen's avatar
zhuwenwen committed
171
172
173
            slo_values.append(
                goodput_config_dict["tpot"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
zhuwenwen's avatar
zhuwenwen committed
174
175
        if "e2el" in goodput_config_dict:
            valid_metrics.append(e2els)
zhuwenwen's avatar
zhuwenwen committed
176
177
178
            slo_values.append(
                goodput_config_dict["e2el"] / MILLISECONDS_TO_SECONDS_CONVERSION
            )
zhuwenwen's avatar
zhuwenwen committed
179
180
181
182
183
184
185
186
187
188

        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.",
zhuwenwen's avatar
zhuwenwen committed
189
190
            stacklevel=2,
        )
zhuwenwen's avatar
zhuwenwen committed
191
192
193
194
195
196
197
198
    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,
zhuwenwen's avatar
zhuwenwen committed
199
200
        mean_ttft_ms=np.mean(ttfts or 0)
        * 1000,  # ttfts is empty if streaming is not supported by backend
zhuwenwen's avatar
zhuwenwen committed
201
202
        std_ttft_ms=np.std(ttfts or 0) * 1000,
        median_ttft_ms=np.median(ttfts or 0) * 1000,
zhuwenwen's avatar
zhuwenwen committed
203
204
205
        percentiles_ttft_ms=[
            (p, np.percentile(ttfts or 0, p) * 1000) for p in selected_percentiles
        ],
zhuwenwen's avatar
zhuwenwen committed
206
207
208
        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,
zhuwenwen's avatar
zhuwenwen committed
209
210
211
        percentiles_tpot_ms=[
            (p, np.percentile(tpots or 0, p) * 1000) for p in selected_percentiles
        ],
zhuwenwen's avatar
zhuwenwen committed
212
213
214
        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,
zhuwenwen's avatar
zhuwenwen committed
215
216
217
        percentiles_itl_ms=[
            (p, np.percentile(itls or 0, p) * 1000) for p in selected_percentiles
        ],
zhuwenwen's avatar
zhuwenwen committed
218
219
220
        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,
zhuwenwen's avatar
zhuwenwen committed
221
222
223
        percentiles_e2el_ms=[
            (p, np.percentile(e2els or 0, p) * 1000) for p in selected_percentiles
        ],
zhuwenwen's avatar
zhuwenwen committed
224
225
226
227
228
229
230
231
232
233
234
235
    )

    return metrics, actual_output_lens


async def benchmark(
    backend: str,
    api_url: str,
    base_url: str,
    model_id: str,
    model_name: str,
    tokenizer: PreTrainedTokenizerBase,
zhuwenwen's avatar
zhuwenwen committed
236
    input_requests: list[SampleRequest],
zhuwenwen's avatar
zhuwenwen committed
237
238
239
240
241
    logprobs: Optional[int],
    request_rate: float,
    burstiness: float,
    disable_tqdm: bool,
    profile: bool,
zhuwenwen's avatar
zhuwenwen committed
242
243
    selected_percentile_metrics: list[str],
    selected_percentiles: list[float],
zhuwenwen's avatar
zhuwenwen committed
244
    ignore_eos: bool,
zhuwenwen's avatar
zhuwenwen committed
245
    goodput_config_dict: dict[str, float],
zhuwenwen's avatar
zhuwenwen committed
246
    max_concurrency: Optional[int],
zhuwenwen's avatar
zhuwenwen committed
247
    lora_modules: Optional[Iterable[str]],
zhuwenwen's avatar
zhuwenwen committed
248
    extra_body: Optional[dict],
zhuwenwen's avatar
zhuwenwen committed
249
250
251
    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
    ramp_up_start_rps: Optional[int] = None,
    ramp_up_end_rps: Optional[int] = None,
zhuwenwen's avatar
zhuwenwen committed
252
253
254
255
256
257
258
):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
    else:
        raise ValueError(f"Unknown backend: {backend}")

    print("Starting initial single prompt test run...")
zhuwenwen's avatar
zhuwenwen committed
259
260
261
262
263
264
    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
        input_requests[0].prompt,
        input_requests[0].prompt_len,
        input_requests[0].expected_output_len,
        input_requests[0].multi_modal_data,
    )
zhuwenwen's avatar
zhuwenwen committed
265
266

    assert test_mm_content is None or isinstance(test_mm_content, dict)
zhuwenwen's avatar
zhuwenwen committed
267
268
269
270
271
272
273
274
275
276
    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,
zhuwenwen's avatar
zhuwenwen committed
277
        extra_body=extra_body,
zhuwenwen's avatar
zhuwenwen committed
278
279
280
281
282
283
    )

    test_output = await request_func(request_func_input=test_input)
    if not test_output.success:
        raise ValueError(
            "Initial test run failed - Please make sure benchmark arguments "
zhuwenwen's avatar
zhuwenwen committed
284
285
            f"are correctly specified. Error: {test_output.error}"
        )
zhuwenwen's avatar
zhuwenwen committed
286
287
288
289
290
291
    else:
        print("Initial test run completed. Starting main benchmark run...")

    if lora_modules:
        # For each input request, choose a LoRA module at random.
        lora_modules = iter(
zhuwenwen's avatar
zhuwenwen committed
292
293
            [random.choice(lora_modules) for _ in range(len(input_requests))]
        )
zhuwenwen's avatar
zhuwenwen committed
294
295
296

    if profile:
        print("Starting profiler...")
zhuwenwen's avatar
zhuwenwen committed
297
298
299
300
301
302
303
304
305
306
307
308
        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_body=extra_body,
        )
zhuwenwen's avatar
zhuwenwen committed
309
310
311
312
        profile_output = await request_func(request_func_input=profile_input)
        if profile_output.success:
            print("Profiler started")

zhuwenwen's avatar
zhuwenwen committed
313
    distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
zhuwenwen's avatar
zhuwenwen committed
314

zhuwenwen's avatar
zhuwenwen committed
315
316
317
318
319
320
321
322
323
    if ramp_up_strategy is not None:
        print(
            f"Traffic ramp-up strategy: {ramp_up_strategy}. Will increase "
            f"RPS from {ramp_up_start_rps} to {ramp_up_end_rps} RPS over "
            "the duration of the benchmark."
        )
    else:
        print(f"Traffic request rate: {request_rate} RPS.")

zhuwenwen's avatar
zhuwenwen committed
324
325
326
327
328
329
330
331
332
    print(f"Burstiness factor: {burstiness} ({distribution})")
    print(f"Maximum request concurrency: {max_concurrency}")

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

    # This can be used once the minimum Python version is 3.10 or higher,
    # and it will simplify the code in limited_request_func.
    #    semaphore = (asyncio.Semaphore(max_concurrency)
    #                 if max_concurrency else contextlib.nullcontext())
zhuwenwen's avatar
zhuwenwen committed
333
    semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None
zhuwenwen's avatar
zhuwenwen committed
334
335
336

    async def limited_request_func(request_func_input, pbar):
        if semaphore is None:
zhuwenwen's avatar
zhuwenwen committed
337
            return await request_func(request_func_input=request_func_input, pbar=pbar)
zhuwenwen's avatar
zhuwenwen committed
338
        async with semaphore:
zhuwenwen's avatar
zhuwenwen committed
339
            return await request_func(request_func_input=request_func_input, pbar=pbar)
zhuwenwen's avatar
zhuwenwen committed
340
341

    benchmark_start_time = time.perf_counter()
zhuwenwen's avatar
zhuwenwen committed
342
    tasks: list[asyncio.Task] = []
zhuwenwen's avatar
zhuwenwen committed
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370

    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
        rps_change_events.append(
            {
                "rps": last_int_rps,
                "timestamp": datetime.now().isoformat(),
            }
        )

    async for request, current_request_rate in get_request(
        input_requests,
        request_rate,
        burstiness,
        ramp_up_strategy,
        ramp_up_start_rps,
        ramp_up_end_rps,
    ):
        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):
                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
                last_int_rps = current_int_rps

zhuwenwen's avatar
zhuwenwen committed
371
372
373
374
375
376
        prompt, prompt_len, output_len, mm_content = (
            request.prompt,
            request.prompt_len,
            request.expected_output_len,
            request.multi_modal_data,
        )
zhuwenwen's avatar
zhuwenwen committed
377
378
379
380
381
        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

zhuwenwen's avatar
zhuwenwen committed
382
383
384
385
386
387
388
389
390
391
392
393
        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_body=extra_body,
        )
zhuwenwen's avatar
zhuwenwen committed
394
395
        task = limited_request_func(request_func_input=request_func_input, pbar=pbar)
        tasks.append(asyncio.create_task(task))
zhuwenwen's avatar
zhuwenwen committed
396
    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
zhuwenwen's avatar
zhuwenwen committed
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426

    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,
        )
        profile_output = await request_func(request_func_input=profile_input)
        if profile_output.success:
            print("Profiler stopped")

    if pbar is not None:
        pbar.close()

    benchmark_duration = time.perf_counter() - benchmark_start_time

    metrics, actual_output_lens = calculate_metrics(
        input_requests=input_requests,
        outputs=outputs,
        dur_s=benchmark_duration,
        tokenizer=tokenizer,
        selected_percentile_metrics=selected_percentile_metrics,
        selected_percentiles=selected_percentiles,
        goodput_config_dict=goodput_config_dict,
    )

zhuwenwen's avatar
zhuwenwen committed
427
    print("{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
zhuwenwen's avatar
zhuwenwen committed
428
    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
zhuwenwen's avatar
zhuwenwen committed
429
    print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
zhuwenwen's avatar
zhuwenwen committed
430
    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
zhuwenwen's avatar
zhuwenwen committed
431
432
433
434
435
436
    print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
    print(
        "{:<40} {:<10.2f}".format(
            "Request throughput (req/s):", metrics.request_throughput
        )
    )
zhuwenwen's avatar
zhuwenwen committed
437
    if goodput_config_dict:
zhuwenwen's avatar
zhuwenwen committed
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
        print(
            "{:<40} {:<10.2f}".format(
                "Request goodput (req/s):", metrics.request_goodput
            )
        )
    print(
        "{:<40} {:<10.2f}".format(
            "Output token throughput (tok/s):", metrics.output_throughput
        )
    )
    print(
        "{:<40} {:<10.2f}".format(
            "Total Token throughput (tok/s):", metrics.total_token_throughput
        )
    )
zhuwenwen's avatar
zhuwenwen committed
453
454
455
456
457
458
459

    result = {
        "duration": benchmark_duration,
        "completed": metrics.completed,
        "total_input_tokens": metrics.total_input,
        "total_output_tokens": metrics.total_output,
        "request_throughput": metrics.request_throughput,
zhuwenwen's avatar
zhuwenwen committed
460
        "request_goodput": metrics.request_goodput if goodput_config_dict else None,
zhuwenwen's avatar
zhuwenwen committed
461
462
463
464
465
466
467
468
469
470
        "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],
    }

zhuwenwen's avatar
zhuwenwen committed
471
472
473
    if rps_change_events:
        result["rps_change_events"] = rps_change_events

zhuwenwen's avatar
zhuwenwen committed
474
475
476
477
478
479
480
481
482
483
484
485
    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
zhuwenwen's avatar
zhuwenwen committed
486
487
488
489
490
491
492
493
494
495
496
497
498
        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"),
            )
        )
zhuwenwen's avatar
zhuwenwen committed
499
        result[f"mean_{metric_attribute_name}_ms"] = getattr(
zhuwenwen's avatar
zhuwenwen committed
500
501
            metrics, f"mean_{metric_attribute_name}_ms"
        )
zhuwenwen's avatar
zhuwenwen committed
502
        result[f"median_{metric_attribute_name}_ms"] = getattr(
zhuwenwen's avatar
zhuwenwen committed
503
504
            metrics, f"median_{metric_attribute_name}_ms"
        )
zhuwenwen's avatar
zhuwenwen committed
505
        result[f"std_{metric_attribute_name}_ms"] = getattr(
zhuwenwen's avatar
zhuwenwen committed
506
507
508
            metrics, f"std_{metric_attribute_name}_ms"
        )
        for p, value in getattr(metrics, f"percentiles_{metric_attribute_name}_ms"):
zhuwenwen's avatar
zhuwenwen committed
509
            p_word = str(int(p)) if int(p) == p else str(p)
zhuwenwen's avatar
zhuwenwen committed
510
            print("{:<40} {:<10.2f}".format(f"P{p_word} {metric_name} (ms):", value))
zhuwenwen's avatar
zhuwenwen committed
511
512
513
            result[f"p{p_word}_{metric_attribute_name}_ms"] = value

    process_one_metric("ttft", "TTFT", "Time to First Token")
zhuwenwen's avatar
zhuwenwen committed
514
    process_one_metric("tpot", "TPOT", "Time per Output Token (excl. 1st token)")
zhuwenwen's avatar
zhuwenwen committed
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
    process_one_metric("itl", "ITL", "Inter-token Latency")
    process_one_metric("e2el", "E2EL", "End-to-end Latency")

    print("=" * 50)

    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 "
zhuwenwen's avatar
zhuwenwen committed
534
535
                    f"{str(VALID_NAMES)}. "
                )
zhuwenwen's avatar
zhuwenwen committed
536
537
538
539
            if slo_val < 0:
                raise ValueError(
                    f"Invalid value found, {slo_name}: {slo_val}. "
                    "The service level objective value should be "
zhuwenwen's avatar
zhuwenwen committed
540
541
                    "non-negative."
                )
zhuwenwen's avatar
zhuwenwen committed
542
543
544
545
546
547
548
549
550
551
552
553
    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. "
zhuwenwen's avatar
zhuwenwen committed
554
            'Specify service level objectives for goodput as "KEY:VALUE" '
zhuwenwen's avatar
zhuwenwen committed
555
            "pairs, where the key is a metric name, and the value is a "
zhuwenwen's avatar
zhuwenwen committed
556
557
            "number in milliseconds."
        ) from err
zhuwenwen's avatar
zhuwenwen committed
558
559
560
    return goodput_config_dict


zhuwenwen's avatar
zhuwenwen committed
561
562
563
def save_to_pytorch_benchmark_format(
    args: argparse.Namespace, results: dict[str, Any], file_name: str
) -> None:
zhuwenwen's avatar
zhuwenwen committed
564
    metrics = [
zhuwenwen's avatar
zhuwenwen committed
565
566
567
568
569
570
571
572
573
574
575
576
        "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",
zhuwenwen's avatar
zhuwenwen committed
577
578
579
580
581
582
    ]
    # 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,
zhuwenwen's avatar
zhuwenwen committed
583
        metrics={k: [results[k]] for k in metrics},
zhuwenwen's avatar
zhuwenwen committed
584
585
        extra_info={
            k: results[k]
zhuwenwen's avatar
zhuwenwen committed
586
587
588
589
            for k in results
            if k not in metrics and k not in ignored_metrics
        },
    )
zhuwenwen's avatar
zhuwenwen committed
590
591
592
    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"
zhuwenwen's avatar
zhuwenwen committed
593
        write_to_json(pt_file, pt_records)
zhuwenwen's avatar
zhuwenwen committed
594
595
596
597
598
599
600
601
602
603
604
605
606


def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)
    np.random.seed(args.seed)

    backend = args.backend
    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

zhuwenwen's avatar
zhuwenwen committed
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
    # 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. "
                "Please remove the --request-rate argument."
            )
        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 "
                "--ramp-up-end-rps must be specified"
            )
        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")
        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.")

zhuwenwen's avatar
zhuwenwen committed
627
628
629
630
631
632
633
    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}"

zhuwenwen's avatar
zhuwenwen committed
634
635
636
637
638
    tokenizer = get_tokenizer(
        tokenizer_id,
        tokenizer_mode=tokenizer_mode,
        trust_remote_code=args.trust_remote_code,
    )
zhuwenwen's avatar
zhuwenwen committed
639

zhuwenwen's avatar
zhuwenwen committed
640
641
642
    if args.dataset_name is None:
        raise ValueError(
            "Please specify '--dataset-name' and the corresponding "
zhuwenwen's avatar
zhuwenwen committed
643
644
            "'--dataset-path' if required."
        )
zhuwenwen's avatar
zhuwenwen committed
645

zhuwenwen's avatar
zhuwenwen committed
646
647
648
649
650
651
652
653
654
655
    if args.dataset_name == "custom":
        dataset = CustomDataset(dataset_path=args.dataset_path)
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
            skip_chat_template=args.custom_skip_chat_template,
        )

    elif args.dataset_name == "sonnet":
zhuwenwen's avatar
zhuwenwen committed
656
657
        dataset = SonnetDataset(dataset_path=args.dataset_path)
        # For the "sonnet" dataset, formatting depends on the backend.
zhuwenwen's avatar
zhuwenwen committed
658
        if args.backend == "openai-chat":
zhuwenwen's avatar
zhuwenwen committed
659
660
661
662
663
664
665
666
            input_requests = dataset.sample(
                num_requests=args.num_prompts,
                input_len=args.sonnet_input_len,
                output_len=args.sonnet_output_len,
                prefix_len=args.sonnet_prefix_len,
                tokenizer=tokenizer,
                return_prompt_formatted=False,
            )
zhuwenwen's avatar
zhuwenwen committed
667
        else:
zhuwenwen's avatar
zhuwenwen committed
668
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
zhuwenwen's avatar
zhuwenwen committed
669
670
671
672
673
674
675
676
677
678
                "Tokenizer/model must have chat template for sonnet dataset."
            )
            input_requests = dataset.sample(
                num_requests=args.num_prompts,
                input_len=args.sonnet_input_len,
                output_len=args.sonnet_output_len,
                prefix_len=args.sonnet_prefix_len,
                tokenizer=tokenizer,
                return_prompt_formatted=True,
            )
zhuwenwen's avatar
zhuwenwen committed
679
680

    elif args.dataset_name == "hf":
zhuwenwen's avatar
zhuwenwen committed
681
682
683
684
685
686
687
688
689
        # all following datasets are implemented from the
        # HuggingFaceDataset base class
        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = VisionArenaDataset
            args.hf_split = "train"
            args.hf_subset = None
        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = InstructCoderDataset
            args.hf_split = "train"
zhuwenwen's avatar
zhuwenwen committed
690
691
692
        elif args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = MTBenchDataset
            args.hf_split = "train"
zhuwenwen's avatar
zhuwenwen committed
693
694
695
696
697
        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = ConversationDataset
        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
            dataset_class = AIMODataset
            args.hf_split = "train"
zhuwenwen's avatar
zhuwenwen committed
698
699
700
701
702
703
        elif args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS:  # noqa: E501
            dataset_class = NextEditPredictionDataset
            args.hf_split = "train"
        elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = ASRDataset
            args.hf_split = "train"
zhuwenwen's avatar
zhuwenwen committed
704
        else:
zhuwenwen's avatar
zhuwenwen committed
705
706
707
708
709
710
711
            supported_datasets = set(
                [
                    dataset_name
                    for cls in HuggingFaceDataset.__subclasses__()
                    for dataset_name in cls.SUPPORTED_DATASET_PATHS
                ]
            )
zhuwenwen's avatar
zhuwenwen committed
712
713
714
715
716
            raise ValueError(
                f"Unsupported dataset path: {args.dataset_path}. "
                "Huggingface dataset only supports dataset_path"
                f" from one of following: {supported_datasets}. "
                "Please consider contributing if you would "
zhuwenwen's avatar
zhuwenwen committed
717
718
719
720
721
722
723
724
725
726
727
728
                "like to add support for additional dataset formats."
            )

        if dataset_class.IS_MULTIMODAL and backend not in [
            "openai-chat",
            "openai-audio",
        ]:
            # multi-modal benchmark is only available on OpenAI Chat backend.
            raise ValueError(
                "Multi-modal content is only supported on 'openai-chat' and "
                "'openai-audio' backend."
            )
zhuwenwen's avatar
zhuwenwen committed
729
        input_requests = dataset_class(
zhuwenwen's avatar
zhuwenwen committed
730
731
732
            dataset_path=args.dataset_path,
            dataset_subset=args.hf_subset,
            dataset_split=args.hf_split,
zhuwenwen's avatar
zhuwenwen committed
733
            random_seed=args.seed,
734
            no_stream=args.no_stream,
zhuwenwen's avatar
zhuwenwen committed
735
        ).sample(
zhuwenwen's avatar
zhuwenwen committed
736
737
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
zhuwenwen's avatar
zhuwenwen committed
738
            output_len=args.hf_output_len,
zhuwenwen's avatar
zhuwenwen committed
739
740
741
        )

    else:
zhuwenwen's avatar
zhuwenwen committed
742
743
        # For datasets that follow a similar structure, use a mapping.
        dataset_mapping = {
zhuwenwen's avatar
zhuwenwen committed
744
745
746
747
748
749
750
751
752
753
754
            "sharegpt": lambda: ShareGPTDataset(
                random_seed=args.seed, dataset_path=args.dataset_path
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
            ),
            "burstgpt": lambda: BurstGPTDataset(
                random_seed=args.seed, dataset_path=args.dataset_path
            ).sample(tokenizer=tokenizer, num_requests=args.num_prompts),
            "random": lambda: RandomDataset(dataset_path=args.dataset_path).sample(
zhuwenwen's avatar
zhuwenwen committed
755
756
757
758
759
760
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                prefix_len=args.random_prefix_len,
                input_len=args.random_input_len,
                output_len=args.random_output_len,
                range_ratio=args.random_range_ratio,
zhuwenwen's avatar
zhuwenwen committed
761
            ),
zhuwenwen's avatar
zhuwenwen committed
762
        }
zhuwenwen's avatar
zhuwenwen committed
763

zhuwenwen's avatar
zhuwenwen committed
764
765
766
767
        try:
            input_requests = dataset_mapping[args.dataset_name]()
        except KeyError as err:
            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err
zhuwenwen's avatar
zhuwenwen committed
768
769
    goodput_config_dict = check_goodput_args(args)

zhuwenwen's avatar
zhuwenwen committed
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
    # Collect the sampling parameters.
    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,
        }.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."
        )

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

    if args.backend == "llama.cpp":
        # Disable prompt caching in llama.cpp backend
        sampling_params["cache_prompt"] = False

zhuwenwen's avatar
zhuwenwen committed
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
    # Avoid GC processing "static" data - reduce pause times.
    gc.collect()
    gc.freeze()

    benchmark_result = asyncio.run(
        benchmark(
            backend=backend,
            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,
            selected_percentile_metrics=args.percentile_metrics.split(","),
zhuwenwen's avatar
zhuwenwen committed
814
            selected_percentiles=[float(p) for p in args.metric_percentiles.split(",")],
zhuwenwen's avatar
zhuwenwen committed
815
816
817
818
            ignore_eos=args.ignore_eos,
            goodput_config_dict=goodput_config_dict,
            max_concurrency=args.max_concurrency,
            lora_modules=args.lora_modules,
zhuwenwen's avatar
zhuwenwen committed
819
            extra_body=sampling_params,
zhuwenwen's avatar
zhuwenwen committed
820
821
822
            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,
zhuwenwen's avatar
zhuwenwen committed
823
824
        )
    )
zhuwenwen's avatar
zhuwenwen committed
825
826

    # Save config and results to json
zhuwenwen's avatar
zhuwenwen committed
827
    if args.save_result or args.append_result:
zhuwenwen's avatar
zhuwenwen committed
828
        result_json: dict[str, Any] = {}
zhuwenwen's avatar
zhuwenwen committed
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847

        # Setup
        current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
        result_json["date"] = current_dt
        result_json["backend"] = backend
        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:
                    kvstring = item.split("=")
                    result_json[kvstring[0].strip()] = kvstring[1].strip()
                else:
                    raise ValueError(
                        "Invalid metadata format. Please use KEY=VALUE format."
                    )
zhuwenwen's avatar
zhuwenwen committed
848
849
850
851
852
853
854
        # Traffic
        result_json["request_rate"] = (
            args.request_rate if args.request_rate < float("inf") else "inf"
        )
        result_json["burstiness"] = args.burstiness
        result_json["max_concurrency"] = args.max_concurrency

zhuwenwen's avatar
zhuwenwen committed
855
856
857
858
859
        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

zhuwenwen's avatar
zhuwenwen committed
860
861
        # Merge with benchmark result
        result_json = {**result_json, **benchmark_result}
zhuwenwen's avatar
zhuwenwen committed
862

zhuwenwen's avatar
zhuwenwen committed
863
864
865
        if not args.save_detailed:
            # Remove fields with too many data points
            for field in [
zhuwenwen's avatar
zhuwenwen committed
866
867
868
869
870
871
                "input_lens",
                "output_lens",
                "ttfts",
                "itls",
                "generated_texts",
                "errors",
zhuwenwen's avatar
zhuwenwen committed
872
873
874
            ]:
                if field in result_json:
                    del result_json[field]
zhuwenwen's avatar
zhuwenwen committed
875
876
                if field in benchmark_result:
                    del benchmark_result[field]
zhuwenwen's avatar
zhuwenwen committed
877
878
879

        # Save to file
        base_model_id = model_id.split("/")[-1]
zhuwenwen's avatar
zhuwenwen committed
880
881
882
883
884
        max_concurrency_str = (
            f"-concurrency{args.max_concurrency}"
            if args.max_concurrency is not None
            else ""
        )
zhuwenwen's avatar
zhuwenwen committed
885
886
887
888
        if args.ramp_up_strategy is not None:
            file_name = f"{backend}-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
        else:
            file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
zhuwenwen's avatar
zhuwenwen committed
889
890
891
        if args.result_filename:
            file_name = args.result_filename
        if args.result_dir:
zhuwenwen's avatar
zhuwenwen committed
892
            os.makedirs(args.result_dir, exist_ok=True)
zhuwenwen's avatar
zhuwenwen committed
893
            file_name = os.path.join(args.result_dir, file_name)
zhuwenwen's avatar
zhuwenwen committed
894
895
896
897
898
899
        with open(
            file_name, mode="a+" if args.append_result else "w", encoding="utf-8"
        ) as outfile:
            # Append a newline.
            if args.append_result and outfile.tell() != 0:
                outfile.write("\n")
zhuwenwen's avatar
zhuwenwen committed
900
901
902
903
            json.dump(result_json, outfile)
        save_to_pytorch_benchmark_format(args, result_json, file_name)


zhuwenwen's avatar
zhuwenwen committed
904
def create_argument_parser():
zhuwenwen's avatar
zhuwenwen committed
905
    parser = FlexibleArgumentParser(
zhuwenwen's avatar
zhuwenwen committed
906
907
        description="Benchmark the online serving throughput."
    )
zhuwenwen's avatar
zhuwenwen committed
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
    parser.add_argument(
        "--backend",
        type=str,
        default="vllm",
        choices=list(ASYNC_REQUEST_FUNCS.keys()),
    )
    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.",
    )
    parser.add_argument(
        "--dataset-name",
        type=str,
        default="sharegpt",
zhuwenwen's avatar
zhuwenwen committed
933
        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf", "custom"],
zhuwenwen's avatar
zhuwenwen committed
934
935
        help="Name of the dataset to benchmark on.",
    )
zhuwenwen's avatar
zhuwenwen committed
936
937
938
939
940
941
942
    parser.add_argument(
        "--dataset-path",
        type=str,
        default=None,
        help="Path to the sharegpt/sonnet dataset. "
        "Or the huggingface dataset ID if using HF dataset.",
    )
943
944
945
946
947
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
zhuwenwen's avatar
zhuwenwen committed
948
949
950
951
952
953
954
955
956
957
958
    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, "
zhuwenwen's avatar
zhuwenwen committed
959
960
        "if the server is not processing requests fast enough to keep up.",
    )
zhuwenwen's avatar
zhuwenwen committed
961
962
963
964
965
966
967
968
969
970

    parser.add_argument(
        "--model",
        type=str,
        required=True,
        help="Name of the model.",
    )
    parser.add_argument(
        "--tokenizer",
        type=str,
zhuwenwen's avatar
zhuwenwen committed
971
        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
972
973
974
975
976
977
978
979
980
981
982
983
    )
    parser.add_argument("--use-beam-search", action="store_true")
    parser.add_argument(
        "--num-prompts",
        type=int,
        default=1000,
        help="Number of prompts to process.",
    )
    parser.add_argument(
        "--logprobs",
        type=int,
        default=None,
zhuwenwen's avatar
zhuwenwen committed
984
985
986
987
988
989
990
        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"
        ),
zhuwenwen's avatar
zhuwenwen committed
991
992
993
994
995
996
997
998
999
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
    )
    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("--seed", type=int, default=0)
    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",
    )
zhuwenwen's avatar
zhuwenwen committed
1035
1036
1037
1038
1039
1040
    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.",
    )
zhuwenwen's avatar
zhuwenwen committed
1041
1042
1043
1044
1045
    parser.add_argument(
        "--append-result",
        action="store_true",
        help="Append the benchmark result to the existing json file.",
    )
zhuwenwen's avatar
zhuwenwen committed
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
    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 "
        "{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
        " format.",
    )
    parser.add_argument(
        "--ignore-eos",
        action="store_true",
        help="Set ignore_eos flag when sending the benchmark request."
zhuwenwen's avatar
zhuwenwen committed
1074
1075
        "Warning: ignore_eos is not supported in deepspeed_mii and tgi.",
    )
zhuwenwen's avatar
zhuwenwen committed
1076
1077
1078
1079
    parser.add_argument(
        "--percentile-metrics",
        type=str,
        default="ttft,tpot,itl",
zhuwenwen's avatar
zhuwenwen committed
1080
        help="Comma-separated list of selected metrics to report percentils. "
zhuwenwen's avatar
zhuwenwen committed
1081
        "This argument specifies the metrics to report percentiles. "
zhuwenwen's avatar
zhuwenwen committed
1082
1083
1084
        'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
        'Default value is "ttft,tpot,itl".',
    )
zhuwenwen's avatar
zhuwenwen committed
1085
1086
1087
1088
    parser.add_argument(
        "--metric-percentiles",
        type=str,
        default="99",
zhuwenwen's avatar
zhuwenwen committed
1089
1090
1091
1092
        help="Comma-separated list of percentiles for selected metrics. "
        'To report 25-th, 50-th, and 75-th percentiles, use "25,50,75". '
        'Default value is "99". '
        'Use "--percentile-metrics" to select metrics.',
zhuwenwen's avatar
zhuwenwen committed
1093
1094
1095
1096
1097
    )
    parser.add_argument(
        "--goodput",
        nargs="+",
        required=False,
zhuwenwen's avatar
zhuwenwen committed
1098
        help='Specify service level objectives for goodput as "KEY:VALUE" '
zhuwenwen's avatar
zhuwenwen committed
1099
        "pairs, where the key is a metric name, and the value is in "
zhuwenwen's avatar
zhuwenwen committed
1100
        'milliseconds. Multiple "KEY:VALUE" pairs can be provided, '
zhuwenwen's avatar
zhuwenwen committed
1101
        "separated by spaces. Allowed request level metric names are "
zhuwenwen's avatar
zhuwenwen committed
1102
        '"ttft", "tpot", "e2el". For more context on the definition of '
zhuwenwen's avatar
zhuwenwen committed
1103
        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
zhuwenwen's avatar
zhuwenwen committed
1104
1105
        "and the blog: https://hao-ai-lab.github.io/blogs/distserve",
    )
zhuwenwen's avatar
zhuwenwen committed
1106
1107

    # group for dataset specific arguments
zhuwenwen's avatar
zhuwenwen committed
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
    custom_group = parser.add_argument_group("custom dataset options")
    custom_group.add_argument(
        "--custom-output-len",
        type=int,
        default=256,
        help="Number of output tokens per request, used only for custom dataset.",
    )
    custom_group.add_argument(
        "--custom-skip-chat-template",
        action="store_true",
        help="Skip applying chat template to prompt, used only for custom dataset.",
    )

zhuwenwen's avatar
zhuwenwen committed
1121
1122
1123
1124
1125
    sonnet_group = parser.add_argument_group("sonnet dataset options")
    sonnet_group.add_argument(
        "--sonnet-input-len",
        type=int,
        default=550,
zhuwenwen's avatar
zhuwenwen committed
1126
        help="Number of input tokens per request, used only for sonnet dataset.",
zhuwenwen's avatar
zhuwenwen committed
1127
1128
1129
1130
1131
    )
    sonnet_group.add_argument(
        "--sonnet-output-len",
        type=int,
        default=150,
zhuwenwen's avatar
zhuwenwen committed
1132
        help="Number of output tokens per request, used only for sonnet dataset.",
zhuwenwen's avatar
zhuwenwen committed
1133
1134
1135
1136
1137
    )
    sonnet_group.add_argument(
        "--sonnet-prefix-len",
        type=int,
        default=200,
zhuwenwen's avatar
zhuwenwen committed
1138
        help="Number of prefix tokens per request, used only for sonnet dataset.",
zhuwenwen's avatar
zhuwenwen committed
1139
1140
1141
1142
1143
1144
1145
1146
    )

    sharegpt_group = parser.add_argument_group("sharegpt dataset options")
    sharegpt_group.add_argument(
        "--sharegpt-output-len",
        type=int,
        default=None,
        help="Output length for each request. Overrides the output length "
zhuwenwen's avatar
zhuwenwen committed
1147
1148
        "from the ShareGPT dataset.",
    )
zhuwenwen's avatar
zhuwenwen committed
1149
1150
1151
1152
1153
1154

    random_group = parser.add_argument_group("random dataset options")
    random_group.add_argument(
        "--random-input-len",
        type=int,
        default=1024,
zhuwenwen's avatar
zhuwenwen committed
1155
        help="Number of input tokens per request, used only for random sampling.",
zhuwenwen's avatar
zhuwenwen committed
1156
1157
1158
1159
1160
    )
    random_group.add_argument(
        "--random-output-len",
        type=int,
        default=128,
zhuwenwen's avatar
zhuwenwen committed
1161
        help="Number of output tokens per request, used only for random sampling.",
zhuwenwen's avatar
zhuwenwen committed
1162
1163
1164
1165
    )
    random_group.add_argument(
        "--random-range-ratio",
        type=float,
zhuwenwen's avatar
zhuwenwen committed
1166
1167
1168
1169
1170
        default=0.0,
        help="Range ratio for sampling input/output length, "
        "used only for random sampling. Must be in the range [0, 1) to define "
        "a symmetric sampling range"
        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
zhuwenwen's avatar
zhuwenwen committed
1171
1172
1173
1174
1175
    )
    random_group.add_argument(
        "--random-prefix-len",
        type=int,
        default=0,
zhuwenwen's avatar
zhuwenwen committed
1176
1177
1178
1179
1180
1181
1182
1183
1184
        help=(
            "Number of fixed prefix tokens before the random context "
            "in a request. "
            "The total input length is the sum of `random-prefix-len` and "
            "a random "
            "context length sampled from [input_len * (1 - range_ratio), "
            "input_len * (1 + range_ratio)]."
        ),
    )
zhuwenwen's avatar
zhuwenwen committed
1185
1186

    hf_group = parser.add_argument_group("hf dataset options")
zhuwenwen's avatar
zhuwenwen committed
1187
1188
1189
1190
1191
1192
    hf_group.add_argument(
        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
    )
    hf_group.add_argument(
        "--hf-split", type=str, default=None, help="Split of the HF dataset."
    )
zhuwenwen's avatar
zhuwenwen committed
1193
1194
1195
1196
1197
1198
1199
1200
    hf_group.add_argument(
        "--hf-output-len",
        type=int,
        default=None,
        help="Output length for each request. Overrides the output lengths "
        "from the sampled HF dataset.",
    )

zhuwenwen's avatar
zhuwenwen committed
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
    sampling_group = parser.add_argument_group("sampling parameters")
    sampling_group.add_argument(
        "--top-p",
        type=float,
        default=None,
        help="Top-p sampling parameter. Only has effect on openai-compatible backends.",
    )
    sampling_group.add_argument(
        "--top-k",
        type=int,
        default=None,
        help="Top-k sampling parameter. Only has effect on openai-compatible backends.",
    )
    sampling_group.add_argument(
        "--min-p",
        type=float,
        default=None,
        help="Min-p sampling parameter. Only has effect on openai-compatible backends.",
    )
    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).",
    )

zhuwenwen's avatar
zhuwenwen committed
1229
    parser.add_argument(
zhuwenwen's avatar
zhuwenwen committed
1230
        "--tokenizer-mode",
zhuwenwen's avatar
zhuwenwen committed
1231
1232
        type=str,
        default="auto",
zhuwenwen's avatar
zhuwenwen committed
1233
        choices=["auto", "slow", "mistral", "custom"],
zhuwenwen's avatar
zhuwenwen committed
1234
1235
        help='The tokenizer mode.\n\n* "auto" will use the '
        'fast tokenizer if available.\n* "slow" will '
zhuwenwen's avatar
zhuwenwen committed
1236
        "always use the slow tokenizer. \n* "
zhuwenwen's avatar
zhuwenwen committed
1237
        '"mistral" will always use the `mistral_common` tokenizer. \n*'
zhuwenwen's avatar
zhuwenwen committed
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
        '"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.",
    )
zhuwenwen's avatar
zhuwenwen committed
1258

zhuwenwen's avatar
zhuwenwen committed
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
    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 --ramp-up-end-rps). "
        "over the duration of the benchmark.",
    )
    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.",
    )

    return parser
zhuwenwen's avatar
zhuwenwen committed
1285

zhuwenwen's avatar
zhuwenwen committed
1286
1287
1288
1289

if __name__ == "__main__":
    parser = create_argument_parser()
    args = parser.parse_args()
zhuwenwen's avatar
zhuwenwen committed
1290
    main(args)