bench_serving.py 62.6 KB
Newer Older
zhyncs's avatar
zhyncs committed
1
2
# Adapted from https://github.com/vllm-project/vllm/blob/6366efc67b0aedd2c1721c14385370e50b297fb3/benchmarks/backend_request_func.py
# Adapted from https://github.com/vllm-project/vllm/blob/6366efc67b0aedd2c1721c14385370e50b297fb3/benchmarks/benchmark_serving.py
3

Ying Sheng's avatar
Ying Sheng committed
4
"""
5
Benchmark online serving with dynamic requests.
Ying Sheng's avatar
Ying Sheng committed
6
7

Usage:
8
python3 -m sglang.bench_serving --backend sglang --num-prompt 10
Ying Sheng's avatar
Ying Sheng committed
9

10
python3 -m sglang.bench_serving --backend sglang --dataset-name random --num-prompts 3000 --random-input 1024 --random-output 1024 --random-range-ratio 0.5
Ying Sheng's avatar
Ying Sheng committed
11
"""
zhyncs's avatar
zhyncs committed
12
13
14
15
16

import argparse
import asyncio
import json
import os
17
import pickle
zhyncs's avatar
zhyncs committed
18
19
20
21
22
23
import random
import resource
import sys
import time
import traceback
import warnings
24
from argparse import ArgumentParser
zhyncs's avatar
zhyncs committed
25
from dataclasses import dataclass, field
26
from datetime import datetime
27
from pathlib import Path
28
from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple, Union
zhyncs's avatar
zhyncs committed
29
30
31
32
33
34
35
36
37
38
39
40
41

import aiohttp
import numpy as np
import requests
from tqdm.asyncio import tqdm
from transformers import (
    AutoTokenizer,
    PreTrainedTokenizer,
    PreTrainedTokenizerBase,
    PreTrainedTokenizerFast,
)

AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
42
ASSISTANT_SUFFIX = "Assistant:"
zhyncs's avatar
zhyncs committed
43

44
45
global args

zhyncs's avatar
zhyncs committed
46

Yineng Zhang's avatar
Yineng Zhang committed
47
48
49
50
51
52
# don't want to import sglang package here
def _get_bool_env_var(name: str, default: str = "false") -> bool:
    value = os.getenv(name, default)
    return value.lower() in ("true", "1")


zhyncs's avatar
zhyncs committed
53
54
55
56
57
58
59
@dataclass
class RequestFuncInput:
    prompt: str
    api_url: str
    prompt_len: int
    output_len: int
    model: str
60
    lora_name: str
61
    image_data: str
62
    extra_request_body: Dict[str, Any]
zhyncs's avatar
zhyncs committed
63
64
65
66
67
68
69
70
71
72
73


@dataclass
class RequestFuncOutput:
    generated_text: str = ""
    success: bool = False
    latency: float = 0.0
    ttft: float = 0.0  # Time to first token
    itl: List[float] = field(default_factory=list)  # List of inter-token latencies
    prompt_len: int = 0
    error: str = ""
74
    output_len: int = 0
zhyncs's avatar
zhyncs committed
75

76
77
78
79
80
81
    @staticmethod
    def init_new(request_func_input: RequestFuncInput):
        output = RequestFuncOutput()
        output.prompt_len = request_func_input.prompt_len
        return output

zhyncs's avatar
zhyncs committed
82
83
84
85
86

def remove_prefix(text: str, prefix: str) -> str:
    return text[len(prefix) :] if text.startswith(prefix) else text


87
88
89
90
def remove_suffix(text: str, suffix: str) -> str:
    return text[: -len(suffix)] if text.endswith(suffix) else text


91
92
93
94
95
96
97
98
def get_auth_headers() -> Dict[str, str]:
    api_key = os.environ.get("OPENAI_API_KEY")
    if api_key:
        return {"Authorization": f"Bearer {api_key}"}
    else:
        return {}


99
# trt llm does not support ignore_eos
100
101
102
103
104
105
106
107
108
109
110
111
# https://github.com/triton-inference-server/tensorrtllm_backend/issues/505
async def async_request_trt_llm(
    request_func_input: RequestFuncInput,
    pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith("generate_stream")

    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
        payload = {
            "accumulate_tokens": True,
            "text_input": request_func_input.prompt,
zhyncs's avatar
zhyncs committed
112
            "temperature": 0.000001,
113
114
115
            "top_p": 1.0,
            "max_tokens": request_func_input.output_len,
            "stream": True,
Ying Sheng's avatar
Ying Sheng committed
116
117
            "min_length": request_func_input.output_len,
            "end_id": 1048576,
118
            **request_func_input.extra_request_body,
119
        }
120
121
122
        if args.disable_ignore_eos:
            del payload["min_length"]
            del payload["end_id"]
123
        output = RequestFuncOutput.init_new(request_func_input)
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142

        ttft = 0.0
        st = time.perf_counter()
        most_recent_timestamp = st
        try:
            async with session.post(url=api_url, json=payload) as response:
                if response.status == 200:
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
                        if not chunk_bytes:
                            continue

                        chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data:")

                        data = json.loads(chunk)
                        output.generated_text += data["text_output"]
                        timestamp = time.perf_counter()
                        # First token
                        if ttft == 0.0:
Xu Song's avatar
Xu Song committed
143
                            ttft = timestamp - st
144
145
146
147
148
149
150
151
152
153
                            output.ttft = ttft

                        # Decoding phase
                        else:
                            output.itl.append(timestamp - most_recent_timestamp)

                        most_recent_timestamp = timestamp

                    output.latency = most_recent_timestamp - st
                    output.success = True
Ying Sheng's avatar
Ying Sheng committed
154
                    output.output_len = request_func_input.output_len
155
156
157
158
159
160
161
162
163
164
165
166
167
168

                else:
                    output.error = response.reason or ""
                    output.success = False
        except Exception:
            output.success = False
            exc_info = sys.exc_info()
            output.error = "".join(traceback.format_exception(*exc_info))

        if pbar:
            pbar.update(1)
        return output


zhyncs's avatar
zhyncs committed
169
170
171
172
173
174
175
176
177
178
# set ignore_eos True by default
async def async_request_openai_completions(
    request_func_input: RequestFuncInput,
    pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith(
        "completions"
    ), "OpenAI Completions API URL must end with 'completions'."

Lianmin Zheng's avatar
Lianmin Zheng committed
179
180
    prompt = request_func_input.prompt

zhyncs's avatar
zhyncs committed
181
182
183
    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
        payload = {
            "model": request_func_input.model,
Lianmin Zheng's avatar
Lianmin Zheng committed
184
            "prompt": prompt,
zhyncs's avatar
zhyncs committed
185
186
187
            "temperature": 0.0,
            "best_of": 1,
            "max_tokens": request_func_input.output_len,
188
            "stream": not args.disable_stream,
189
            "ignore_eos": not args.disable_ignore_eos,
190
            **request_func_input.extra_request_body,
zhyncs's avatar
zhyncs committed
191
        }
192
        headers = get_auth_headers()
zhyncs's avatar
zhyncs committed
193

194
        output = RequestFuncOutput.init_new(request_func_input)
zhyncs's avatar
zhyncs committed
195
196

        generated_text = ""
197
        output_len = request_func_input.output_len
zhyncs's avatar
zhyncs committed
198
199
200
201
202
203
204
205
206
207
208
209
210
211
        ttft = 0.0
        st = time.perf_counter()
        most_recent_timestamp = st
        try:
            async with session.post(
                url=api_url, json=payload, headers=headers
            ) as response:
                if response.status == 200:
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
                        if not chunk_bytes:
                            continue

                        chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data: ")
212
                        latency = time.perf_counter() - st
zhyncs's avatar
zhyncs committed
213
                        if chunk == "[DONE]":
214
                            pass
zhyncs's avatar
zhyncs committed
215
216
217
218
219
220
221
222
223
224
225
226
227
228
                        else:
                            data = json.loads(chunk)

                            # NOTE: Some completion API might have a last
                            # usage summary response without a token so we
                            # want to check a token was generated
                            if data["choices"][0]["text"]:
                                timestamp = time.perf_counter()
                                # First token
                                if ttft == 0.0:
                                    ttft = time.perf_counter() - st
                                    output.ttft = ttft

                                # Decoding phase
229
230
                                else:
                                    output.itl.append(timestamp - most_recent_timestamp)
zhyncs's avatar
zhyncs committed
231
232
233

                                most_recent_timestamp = timestamp
                                generated_text += data["choices"][0]["text"]
Lzhang-hub's avatar
Lzhang-hub committed
234
                                output_len = (data.get("usage") or {}).get(
235
236
                                    "completion_tokens", output_len
                                )
zhyncs's avatar
zhyncs committed
237
238
239
240

                    output.generated_text = generated_text
                    output.success = True
                    output.latency = latency
241
                    output.output_len = output_len
zhyncs's avatar
zhyncs committed
242
243
244
245
246
247
248
249
250
251
252
253
254
                else:
                    output.error = response.reason or ""
                    output.success = False
        except Exception:
            output.success = False
            exc_info = sys.exc_info()
            output.error = "".join(traceback.format_exception(*exc_info))

    if pbar:
        pbar.update(1)
    return output


255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
async def async_request_truss(
    request_func_input: RequestFuncInput,
    pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
    api_url = request_func_input.api_url

    prompt = request_func_input.prompt

    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
        payload = {
            "model": request_func_input.model,
            "prompt": prompt,
            "temperature": 0.0,
            "best_of": 1,
            "max_tokens": request_func_input.output_len,
            "stream": not args.disable_stream,
            "ignore_eos": not args.disable_ignore_eos,
            **request_func_input.extra_request_body,
        }
274
        headers = get_auth_headers()
275

276
        output = RequestFuncOutput.init_new(request_func_input)
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301

        generated_text = ""
        ttft = 0.0
        st = time.perf_counter()
        most_recent_timestamp = st
        try:
            async with session.post(
                url=api_url, json=payload, headers=headers
            ) as response:
                if response.status == 200:
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
                        if not chunk_bytes:
                            continue

                        chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data: ")
                        latency = time.perf_counter() - st
                        if chunk == "[DONE]":
                            pass
                        else:
                            data = json.loads(chunk)

                            # NOTE: Some completion API might have a last
                            # usage summary response without a token so we
                            # want to check a token was generated
302
                            if data["choices"][0]["text"]:
303
304
305
306
307
308
309
310
311
312
313
                                timestamp = time.perf_counter()
                                # First token
                                if ttft == 0.0:
                                    ttft = time.perf_counter() - st
                                    output.ttft = ttft

                                # Decoding phase
                                else:
                                    output.itl.append(timestamp - most_recent_timestamp)

                                most_recent_timestamp = timestamp
314
                                generated_text += data["choices"][0]["text"]
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332

                    output.generated_text = generated_text
                    output.success = True
                    output.latency = latency
                    output.output_len = request_func_input.output_len
                else:
                    output.error = response.reason or ""
                    output.success = False
        except Exception:
            output.success = False
            exc_info = sys.exc_info()
            output.error = "".join(traceback.format_exception(*exc_info))

    if pbar:
        pbar.update(1)
    return output


333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
async def async_request_sglang_generate(
    request_func_input: RequestFuncInput,
    pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    prompt = request_func_input.prompt

    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
        payload = {
            "text": prompt,
            "sampling_params": {
                "temperature": 0.0,
                "max_new_tokens": request_func_input.output_len,
                "ignore_eos": not args.disable_ignore_eos,
            },
            "stream": not args.disable_stream,
349
            "lora_path": request_func_input.lora_name,
350
351
            "return_logprob": args.return_logprob,
            "logprob_start_len": -1,
352
353
            **request_func_input.extra_request_body,
        }
354
355
356
357
358

        # Add image data if available
        if request_func_input.image_data:
            payload["image_data"] = request_func_input.image_data

359
        headers = get_auth_headers()
360

361
        output = RequestFuncOutput.init_new(request_func_input)
362
363

        generated_text = ""
364
        output_len = request_func_input.output_len
365
366
367
        ttft = 0.0
        st = time.perf_counter()
        most_recent_timestamp = st
368
        last_output_len = 0
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
        try:
            async with session.post(
                url=api_url, json=payload, headers=headers
            ) as response:
                if response.status == 200:
                    async for chunk_bytes in response.content:
                        chunk_bytes = chunk_bytes.strip()
                        if not chunk_bytes:
                            continue
                        # print(chunk_bytes)

                        chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data: ")
                        latency = time.perf_counter() - st
                        if chunk == "[DONE]":
                            pass
                        else:
                            data = json.loads(chunk)

                            # NOTE: Some completion API might have a last
                            # usage summary response without a token so we
                            # want to check a token was generated
                            if data["text"]:
                                timestamp = time.perf_counter()
392
393
394
                                generated_text = data["text"]
                                output_len = data["meta_info"]["completion_tokens"]

395
396
397
398
399
400
401
                                # First token
                                if ttft == 0.0:
                                    ttft = time.perf_counter() - st
                                    output.ttft = ttft

                                # Decoding phase
                                else:
402
403
404
405
406
407
408
                                    num_new_tokens = output_len - last_output_len
                                    if num_new_tokens == 0:
                                        continue
                                    adjust_itl = (
                                        timestamp - most_recent_timestamp
                                    ) / num_new_tokens
                                    output.itl.extend([adjust_itl] * num_new_tokens)
409
410

                                most_recent_timestamp = timestamp
Lianmin Zheng's avatar
Lianmin Zheng committed
411
                                last_output_len = output_len
412
413
414
415

                    output.generated_text = generated_text
                    output.success = True
                    output.latency = latency
416
                    output.output_len = output_len
417
418
419
420
421
422
423
                else:
                    output.error = response.reason or ""
                    output.success = False
        except Exception:
            output.success = False
            exc_info = sys.exc_info()
            output.error = "".join(traceback.format_exception(*exc_info))
424
            print(f"{output.error=}")
425
426
427
428
429
430

    if pbar:
        pbar.update(1)
    return output


431
async def async_request_gserver(
Lianmin Zheng's avatar
Lianmin Zheng committed
432
433
434
435
436
437
    request_func_input: RequestFuncInput,
    pbar: Optional[tqdm] = None,
) -> RequestFuncOutput:
    raise NotImplementedError()


438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
async def async_request_profile(api_url: str) -> RequestFuncOutput:
    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
        output = RequestFuncOutput()
        try:
            async with session.post(url=api_url) as response:
                if response.status == 200:
                    output.success = True
                else:
                    output.error = response.reason or ""
                    output.success = False
        except Exception:
            output.success = False
            exc_info = sys.exc_info()
            output.error = "".join(traceback.format_exception(*exc_info))

    return output


zhyncs's avatar
zhyncs committed
456
def get_model(pretrained_model_name_or_path: str) -> str:
457
    if os.getenv("SGLANG_USE_MODELSCOPE", "false").lower() == "true":
zhyncs's avatar
zhyncs committed
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
        import huggingface_hub.constants
        from modelscope import snapshot_download

        model_path = snapshot_download(
            model_id=pretrained_model_name_or_path,
            local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
            ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"],
        )

        return model_path
    return pretrained_model_name_or_path


def get_tokenizer(
    pretrained_model_name_or_path: str,
) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
Lianmin Zheng's avatar
Lianmin Zheng committed
474
475
476
477
478
479
480
    if pretrained_model_name_or_path.endswith(
        ".json"
    ) or pretrained_model_name_or_path.endswith(".model"):
        from sglang.srt.hf_transformers_utils import get_tokenizer

        return get_tokenizer(pretrained_model_name_or_path)

zhyncs's avatar
zhyncs committed
481
482
483
484
485
486
487
488
489
    if pretrained_model_name_or_path is not None and not os.path.exists(
        pretrained_model_name_or_path
    ):
        pretrained_model_name_or_path = get_model(pretrained_model_name_or_path)
    return AutoTokenizer.from_pretrained(
        pretrained_model_name_or_path, trust_remote_code=True
    )


490
491
492
493
494
495
496
def get_dataset(args, tokenizer):
    if args.dataset_name == "sharegpt":
        input_requests = sample_sharegpt_requests(
            dataset_path=args.dataset_path,
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            fixed_output_len=args.sharegpt_output_len,
497
            context_len=args.sharegpt_context_len,
498
            prompt_suffix=args.prompt_suffix,
499
            apply_chat_template=args.apply_chat_template,
500
        )
501
    elif args.dataset_name.startswith("random"):
502
503
504
505
506
507
508
        input_requests = sample_random_requests(
            input_len=args.random_input_len,
            output_len=args.random_output_len,
            num_prompts=args.num_prompts,
            range_ratio=args.random_range_ratio,
            tokenizer=tokenizer,
            dataset_path=args.dataset_path,
509
            random_sample=args.dataset_name == "random",
510
511
512
        )
    elif args.dataset_name == "generated-shared-prefix":
        input_requests = sample_generated_shared_prefix_requests(
513
514
515
516
517
            num_groups=args.gsp_num_groups,
            prompts_per_group=args.gsp_prompts_per_group,
            system_prompt_len=args.gsp_system_prompt_len,
            question_len=args.gsp_question_len,
            output_len=args.gsp_output_len,
518
            tokenizer=tokenizer,
519
            args=args,
520
        )
521
522
523
524
525
526
527
    elif args.dataset_name == "mmmu":
        input_requests = sample_mmmu_requests(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            fixed_output_len=args.random_output_len,
            random_sample=True,
        )
528
529
530
531
532
    else:
        raise ValueError(f"Unknown dataset: {args.dataset_name}")
    return input_requests


zhyncs's avatar
zhyncs committed
533
ASYNC_REQUEST_FUNCS = {
534
535
536
    "sglang": async_request_sglang_generate,
    "sglang-native": async_request_sglang_generate,
    "sglang-oai": async_request_openai_completions,
zhyncs's avatar
zhyncs committed
537
538
    "vllm": async_request_openai_completions,
    "lmdeploy": async_request_openai_completions,
539
    "trt": async_request_trt_llm,
540
    "gserver": async_request_gserver,
541
    "truss": async_request_truss,
zhyncs's avatar
zhyncs committed
542
543
544
545
546
547
548
549
}


@dataclass
class BenchmarkMetrics:
    completed: int
    total_input: int
    total_output: int
Ying Sheng's avatar
Ying Sheng committed
550
    total_output_retokenized: int
zhyncs's avatar
zhyncs committed
551
552
553
    request_throughput: float
    input_throughput: float
    output_throughput: float
Ying Sheng's avatar
Ying Sheng committed
554
    output_throughput_retokenized: float
555
556
    total_throughput: float
    total_throughput_retokenized: float
zhyncs's avatar
zhyncs committed
557
558
559
560
561
562
563
564
565
566
567
    mean_ttft_ms: float
    median_ttft_ms: float
    std_ttft_ms: float
    p99_ttft_ms: float
    mean_tpot_ms: float
    median_tpot_ms: float
    std_tpot_ms: float
    p99_tpot_ms: float
    mean_itl_ms: float
    median_itl_ms: float
    std_itl_ms: float
568
    p95_itl_ms: float
zhyncs's avatar
zhyncs committed
569
    p99_itl_ms: float
570
    max_itl_ms: float
zhyncs's avatar
zhyncs committed
571
572
    mean_e2e_latency_ms: float
    median_e2e_latency_ms: float
573
574
    std_e2e_latency_ms: float
    p99_e2e_latency_ms: float
575
    concurrency: float
zhyncs's avatar
zhyncs committed
576
577


Lianmin Zheng's avatar
Lianmin Zheng committed
578
SHAREGPT_URL = "https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json"
Lianmin Zheng's avatar
Lianmin Zheng committed
579
580


Lianmin Zheng's avatar
Lianmin Zheng committed
581
582
583
584
def download_and_cache_file(url: str, filename: Optional[str] = None):
    """Read and cache a file from a url."""
    if filename is None:
        filename = os.path.join("/tmp", url.split("/")[-1])
Lianmin Zheng's avatar
Lianmin Zheng committed
585

Lianmin Zheng's avatar
Lianmin Zheng committed
586
587
588
    # Check if the cache file already exists
    if os.path.exists(filename):
        return filename
Lianmin Zheng's avatar
Lianmin Zheng committed
589

Lianmin Zheng's avatar
Lianmin Zheng committed
590
    print(f"Downloading from {url} to {filename}")
Lianmin Zheng's avatar
Lianmin Zheng committed
591

Lianmin Zheng's avatar
Lianmin Zheng committed
592
593
594
    # Stream the response to show the progress bar
    response = requests.get(url, stream=True)
    response.raise_for_status()  # Check for request errors
Lianmin Zheng's avatar
Lianmin Zheng committed
595

Lianmin Zheng's avatar
Lianmin Zheng committed
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
    # Total size of the file in bytes
    total_size = int(response.headers.get("content-length", 0))
    chunk_size = 1024  # Download in chunks of 1KB

    # Use tqdm to display the progress bar
    with open(filename, "wb") as f, tqdm(
        desc=filename,
        total=total_size,
        unit="B",
        unit_scale=True,
        unit_divisor=1024,
    ) as bar:
        for chunk in response.iter_content(chunk_size=chunk_size):
            f.write(chunk)
            bar.update(len(chunk))

    return filename
Lianmin Zheng's avatar
Lianmin Zheng committed
613
614


615
616
617
618
619
620
621
@dataclass
class DatasetRow:
    prompt: str
    prompt_len: int
    output_len: int


622
623
624
625
626
def sample_mmmu_requests(
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    fixed_output_len: Optional[int] = None,
    random_sample: bool = True,
627
) -> List[DatasetRow]:
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
    """
    Sample requests from the MMMU dataset using HuggingFace datasets.

    Args:
        num_requests: Number of requests to sample.
        tokenizer: Tokenizer to use for token counting.
        fixed_output_len: If provided, use this fixed output length for all requests.
        random_sample: Whether to randomly sample or take the first N.

    Returns:
        List of tuples (prompt, prompt_token_len, output_token_len).
    """
    try:
        import base64
        import io

        from datasets import load_dataset
    except ImportError:
        raise ImportError("Please install datasets: pip install datasets")

    print("Loading MMMU dataset from HuggingFace...")

    try:
        print("Attempting to load MMMU Math dataset...")
        mmmu_dataset = load_dataset("MMMU/MMMU", "Math", split="test")
        print(
            f"Successfully loaded MMMU Math dataset from HuggingFace with {len(mmmu_dataset)} examples"
        )
    except Exception as e:
        print(f"Failed to load MMMU Math dataset: {e}")
        raise ValueError(f"Failed to load MMMU dataset: {e}")

    # Sample from the dataset
    if len(mmmu_dataset) > num_requests:
        if random_sample:
            # Random sample
            indices = random.sample(range(len(mmmu_dataset)), num_requests)
            sample_dataset = mmmu_dataset.select(indices)
        else:
            # Take first N
            sample_dataset = mmmu_dataset.select(
                range(min(num_requests, len(mmmu_dataset)))
            )
    else:
        print(f"Dataset has less than {num_requests} examples, using all examples")
        sample_dataset = mmmu_dataset

    print(f"Selected {len(sample_dataset)} examples for benchmarking")

    # Create prompts
    filtered_dataset = []

    for i, example in enumerate(sample_dataset):
        try:
            # Extract image_1
            image = example.get("image_1")

            if image is not None:
                if hasattr(image, "save"):
                    # Convert RGBA images to RGB before encoding
                    if image.mode == "RGBA":
                        image = image.convert("RGB")

                    # Encode image to base64
                    buffered = io.BytesIO()
                    image.save(buffered, format="JPEG")
                    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
                    image_path = f"data:image/jpeg;base64,{img_str}"
                else:
                    continue

                # Extract the question
                question = example.get("question")

                # Create the prompt with image, question
                prompt = f"Question: {question}\n\nAnswer: "
                prompt = tokenizer.apply_chat_template(
                    [
                        {
                            "role": "user",
                            "content": [
                                {"type": "image_url", "image_url": {"url": image_path}},
                                {"type": "text", "text": prompt},
                            ],
                        }
                    ],
                    add_generation_prompt=True,
                    tokenize=False,
                )
                prompt = f"<image>{image_path}</image>{prompt}"

                # Calculate token lengths
                # Note: This is approximate since we're not rendering the actual image tokens
                prompt_token_ids = tokenizer.encode(prompt)
                prompt_len = (
                    len(prompt_token_ids) + 512
                )  # Add estimate for image tokens

                output_len = fixed_output_len if fixed_output_len is not None else 256

728
729
730
731
732
                filtered_dataset.append(
                    DatasetRow(
                        prompt=prompt, prompt_len=prompt_len, output_len=output_len
                    )
                )
733
734
735
736
737
738
739
740

        except Exception as e:
            print(f"Error processing example {i}: {e}")

    print(f"\nCreated {len(filtered_dataset)} MMMU prompts")
    return filtered_dataset


zhyncs's avatar
zhyncs committed
741
742
743
744
745
def sample_sharegpt_requests(
    dataset_path: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    fixed_output_len: Optional[int] = None,
746
    context_len: Optional[int] = None,
747
    prompt_suffix: Optional[str] = "",
748
    apply_chat_template=False,
749
) -> List[DatasetRow]:
zhyncs's avatar
zhyncs committed
750
751
752
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")

Lianmin Zheng's avatar
Lianmin Zheng committed
753
    # Download sharegpt if necessary
754
    if not os.path.isfile(dataset_path) and dataset_path == "":
Lianmin Zheng's avatar
Lianmin Zheng committed
755
        dataset_path = download_and_cache_file(SHAREGPT_URL)
zhyncs's avatar
zhyncs committed
756
757
758
759

    # Load the dataset.
    with open(dataset_path) as f:
        dataset = json.load(f)
760

zhyncs's avatar
zhyncs committed
761
    # Filter out the conversations with less than 2 turns.
762
763
764
765
766
    dataset = [
        data
        for data in dataset
        if len(data.get("conversations", data.get("conversation", []))) >= 2
    ]
zhyncs's avatar
zhyncs committed
767
768
    # Only keep the first two turns of each conversation.
    dataset = [
769
770
771
772
        (
            data.get("conversations", data.get("conversation", []))[0]["value"],
            data.get("conversations", data.get("conversation", []))[1]["value"],
        )
zhyncs's avatar
zhyncs committed
773
774
775
776
777
778
779
        for data in dataset
    ]

    # Shuffle the dataset.
    random.shuffle(dataset)

    # Filter out sequences that are too long or too short
780
    filtered_dataset: List[DatasetRow] = []
zhyncs's avatar
zhyncs committed
781
782
783
784
785
786
    for i in range(len(dataset)):
        if len(filtered_dataset) == num_requests:
            break

        # Tokenize the prompts and completions.
        prompt = dataset[i][0]
787
        if prompt_suffix:
788
789
790
791
792
            prompt = (
                remove_suffix(prompt, ASSISTANT_SUFFIX)
                + prompt_suffix
                + ASSISTANT_SUFFIX
            )
793
794
795
796
797
798
799
800
801

        if apply_chat_template:
            prompt = tokenizer.apply_chat_template(
                [{"role": "user", "content": prompt}],
                add_generation_prompt=True,
                tokenize=False,
            )
            prompt = prompt.replace(tokenizer.bos_token, "")

Lianmin Zheng's avatar
Lianmin Zheng committed
802
        prompt_token_ids = tokenizer.encode(prompt)
zhyncs's avatar
zhyncs committed
803
        completion = dataset[i][1]
Lianmin Zheng's avatar
Lianmin Zheng committed
804
        completion_token_ids = tokenizer.encode(completion)
zhyncs's avatar
zhyncs committed
805
806
807
808
        prompt_len = len(prompt_token_ids)
        output_len = (
            len(completion_token_ids) if fixed_output_len is None else fixed_output_len
        )
809

810
        if prompt_len < 2 or output_len < 2:
zhyncs's avatar
zhyncs committed
811
812
            # Prune too short sequences.
            continue
813
814

        if context_len and prompt_len + output_len > context_len:
zhyncs's avatar
zhyncs committed
815
816
            # Prune too long sequences.
            continue
817

818
819
820
        filtered_dataset.append(
            DatasetRow(prompt=prompt, prompt_len=prompt_len, output_len=output_len)
        )
zhyncs's avatar
zhyncs committed
821

822
823
    print(f"#Input tokens: {np.sum([x.prompt_len for x in filtered_dataset])}")
    print(f"#Output tokens: {np.sum([x.output_len for x in filtered_dataset])}")
zhyncs's avatar
zhyncs committed
824
825
826
    return filtered_dataset


827
828
829
830
831
832
def sample_random_requests(
    input_len: int,
    output_len: int,
    num_prompts: int,
    range_ratio: float,
    tokenizer: PreTrainedTokenizerBase,
Lianmin Zheng's avatar
Lianmin Zheng committed
833
    dataset_path: str,
834
    random_sample: bool = True,
835
) -> List[DatasetRow]:
836
    input_lens = np.random.randint(
Yineng Zhang's avatar
Yineng Zhang committed
837
        max(int(input_len * range_ratio), 1),
838
839
840
841
842
843
844
845
        input_len + 1,
        size=num_prompts,
    )
    output_lens = np.random.randint(
        int(output_len * range_ratio),
        output_len + 1,
        size=num_prompts,
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
846

847
    if random_sample:
Lianmin Zheng's avatar
Lianmin Zheng committed
848
849
850
        # Sample token ids from ShareGPT and repeat/truncate them to satisfy the input_lens

        # Download sharegpt if necessary
Lianmin Zheng's avatar
Lianmin Zheng committed
851
852
        if not os.path.isfile(dataset_path):
            dataset_path = download_and_cache_file(SHAREGPT_URL)
Lianmin Zheng's avatar
Lianmin Zheng committed
853
854
855
856
857

        # Load the dataset.
        with open(dataset_path) as f:
            dataset = json.load(f)
        # Filter out the conversations with less than 2 turns.
858
859
860
861
862
        dataset = [
            data
            for data in dataset
            if len(data.get("conversations", data.get("conversation", []))) >= 2
        ]
Lianmin Zheng's avatar
Lianmin Zheng committed
863
864
        # Only keep the first two turns of each conversation.
        dataset = [
865
866
867
868
            (
                data.get("conversations", data.get("conversation", []))[0]["value"],
                data.get("conversations", data.get("conversation", []))[1]["value"],
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
869
870
871
872
873
874
            for data in dataset
        ]
        # Shuffle the dataset.
        random.shuffle(dataset)

        # Filter out sequences that are too long or too short
875
        input_requests: List[DatasetRow] = []
876
877
878
879
880
        for data in dataset:
            i = len(input_requests)
            if i == num_prompts:
                break

Lianmin Zheng's avatar
Lianmin Zheng committed
881
            # Tokenize the prompts and completions.
882
            prompt = data[0]
Lianmin Zheng's avatar
Lianmin Zheng committed
883
            prompt_token_ids = tokenizer.encode(prompt)
Lianmin Zheng's avatar
Lianmin Zheng committed
884
885
            prompt_len = len(prompt_token_ids)

886
887
888
889
            # Skip empty prompt
            if prompt_len == 0:
                continue

Yineng Zhang's avatar
Yineng Zhang committed
890
            if prompt_len > input_lens[i]:
Lianmin Zheng's avatar
Lianmin Zheng committed
891
892
893
894
895
                input_ids = prompt_token_ids[: input_lens[i]]
            else:
                ratio = (input_lens[i] + prompt_len - 1) // prompt_len
                input_ids = (prompt_token_ids * ratio)[: input_lens[i]]
            prompt = tokenizer.decode(input_ids)
896
897
898
899
900
901
902
            input_requests.append(
                DatasetRow(
                    prompt=prompt,
                    prompt_len=int(input_lens[i]),
                    output_len=int(output_lens[i]),
                )
            )
Lianmin Zheng's avatar
Lianmin Zheng committed
903
904
905
906
907
908
909
910
911
912
913
    else:
        # Sample token ids from random integers. This can cause some NaN issues.
        offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)
        input_requests = []
        for i in range(num_prompts):
            prompt = tokenizer.decode(
                [
                    (offsets[i] + i + j) % tokenizer.vocab_size
                    for j in range(input_lens[i])
                ]
            )
914
915
916
917
918
919
920
            input_requests.append(
                DatasetRow(
                    prompt=prompt,
                    prompt_len=int(input_lens[i]),
                    output_len=int(output_lens[i]),
                )
            )
921
922
923
924
925
926

    print(f"#Input tokens: {np.sum(input_lens)}")
    print(f"#Output tokens: {np.sum(output_lens)}")
    return input_requests


927
928
929
930
931
932
933
def gen_prompt(tokenizer, token_num):
    """Generate a random prompt of specified token length using tokenizer vocabulary."""
    all_available_tokens = list(tokenizer.get_vocab().values())
    selected_tokens = random.choices(all_available_tokens, k=token_num)
    return tokenizer.decode(selected_tokens)


934
935
936
937
938
939
def get_gen_prefix_cache_path(args, tokenizer):
    """Create cache directory under ~/.cache/sglang/benchmark"""
    cache_dir = Path.home() / ".cache" / "sglang" / "benchmark"

    # Create a unique cache filename based on the generation parameters
    cache_key = (
940
941
        f"gen_shared_prefix_{args.gsp_num_groups}_{args.gsp_prompts_per_group}_"
        f"{args.gsp_system_prompt_len}_{args.gsp_question_len}_{args.gsp_output_len}_"
942
943
944
945
946
        f"{tokenizer.__class__.__name__}.pkl"
    )
    return cache_dir / cache_key


947
948
949
950
951
952
953
def sample_generated_shared_prefix_requests(
    num_groups: int,
    prompts_per_group: int,
    system_prompt_len: int,
    question_len: int,
    output_len: int,
    tokenizer: PreTrainedTokenizerBase,
954
    args: argparse.Namespace,
955
) -> List[DatasetRow]:
956
957
958
959
960
961
962
    """Generate benchmark requests with shared system prompts using random tokens and caching."""
    cache_path = get_gen_prefix_cache_path(args, tokenizer)

    # Try to load from cache first
    if cache_path.exists():
        print(f"\nLoading cached generated input data from {cache_path}")
        with open(cache_path, "rb") as f:
963
964
            return pickle.load(f)

965
966
    print("\nGenerating new input data...")

967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
    # Generate system prompts for each group
    system_prompts = []
    for _ in range(num_groups):
        system_prompt = gen_prompt(tokenizer, system_prompt_len)
        system_prompts.append(system_prompt)

    # Generate questions
    questions = []
    for _ in range(num_groups * prompts_per_group):
        question = gen_prompt(tokenizer, question_len)
        questions.append(question)

    # Combine system prompts with questions
    input_requests = []
    total_input_tokens = 0
    total_output_tokens = 0

984
    for group_idx in tqdm(range(num_groups), desc="Generating system prompt"):
985
        system_prompt = system_prompts[group_idx]
986
987
988
        for prompt_idx in tqdm(
            range(prompts_per_group), desc="Generating questions", leave=False
        ):
989
990
991
992
            question = questions[group_idx * prompts_per_group + prompt_idx]
            full_prompt = f"{system_prompt}\n\n{question}"
            prompt_len = len(tokenizer.encode(full_prompt))

993
994
995
996
997
            input_requests.append(
                DatasetRow(
                    prompt=full_prompt, prompt_len=prompt_len, output_len=output_len
                )
            )
998
999
1000
            total_input_tokens += prompt_len
            total_output_tokens += output_len

1001
1002
1003
1004
    # Shuffle questions
    random.shuffle(input_requests)

    # Print statistics
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
    print(f"\nGenerated shared prefix dataset statistics:")
    print(f"Number of groups: {num_groups}")
    print(f"Prompts per group: {prompts_per_group}")
    print(f"Total prompts: {len(input_requests)}")
    print(f"Total input tokens: {total_input_tokens}")
    print(f"Total output tokens: {total_output_tokens}")
    print(
        f"Average system prompt length: {sum(len(tokenizer.encode(sp)) for sp in system_prompts) / len(system_prompts):.1f} tokens"
    )
    print(
        f"Average question length: {sum(len(tokenizer.encode(q)) for q in questions) / len(questions):.1f} tokens\n"
    )
1017
1018
1019
1020
1021
1022

    # Save to cache
    cache_path.parent.mkdir(parents=True, exist_ok=True)
    print(f"Caching generated input data to {cache_path}")
    with open(cache_path, "wb") as f:
        pickle.dump(input_requests, f)
1023
1024
1025
1026

    return input_requests


zhyncs's avatar
zhyncs committed
1027
async def get_request(
1028
    input_requests: List[DatasetRow],
zhyncs's avatar
zhyncs committed
1029
    request_rate: float,
1030
) -> AsyncGenerator[DatasetRow, None]:
zhyncs's avatar
zhyncs committed
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
    input_requests = iter(input_requests)
    for request in input_requests:
        yield request

        if request_rate == float("inf"):
            # If the request rate is infinity, then we don't need to wait.
            continue

        # Sample the request interval from the exponential distribution.
        interval = np.random.exponential(1.0 / request_rate)
        # The next request will be sent after the interval.
        await asyncio.sleep(interval)


def calculate_metrics(
1046
    input_requests: List[DatasetRow],
zhyncs's avatar
zhyncs committed
1047
1048
1049
    outputs: List[RequestFuncOutput],
    dur_s: float,
    tokenizer: PreTrainedTokenizerBase,
1050
    backend: str,
zhyncs's avatar
zhyncs committed
1051
) -> Tuple[BenchmarkMetrics, List[int]]:
Ying Sheng's avatar
Ying Sheng committed
1052
1053
    output_lens: List[int] = []
    retokenized_output_lens: List[int] = []
zhyncs's avatar
zhyncs committed
1054
1055
1056
1057
1058
    total_input = 0
    completed = 0
    itls: List[float] = []
    tpots: List[float] = []
    ttfts: List[float] = []
zhyncs's avatar
zhyncs committed
1059
    e2e_latencies: List[float] = []
zhyncs's avatar
zhyncs committed
1060
1061
    for i in range(len(outputs)):
        if outputs[i].success:
Ying Sheng's avatar
Ying Sheng committed
1062
1063
1064
            output_len = outputs[i].output_len
            output_lens.append(output_len)
            retokenized_output_len = len(
Lianmin Zheng's avatar
Lianmin Zheng committed
1065
                tokenizer.encode(outputs[i].generated_text, add_special_tokens=False)
Ying Sheng's avatar
Ying Sheng committed
1066
1067
            )
            retokenized_output_lens.append(retokenized_output_len)
1068
            total_input += input_requests[i].prompt_len
zhyncs's avatar
zhyncs committed
1069
1070
1071
1072
            if output_len > 1:
                tpots.append((outputs[i].latency - outputs[i].ttft) / (output_len - 1))
            itls += outputs[i].itl
            ttfts.append(outputs[i].ttft)
zhyncs's avatar
zhyncs committed
1073
1074
1075

            e2e_latencies.append(outputs[i].latency)

zhyncs's avatar
zhyncs committed
1076
1077
            completed += 1
        else:
Ying Sheng's avatar
Ying Sheng committed
1078
1079
            output_lens.append(0)
            retokenized_output_lens.append(0)
zhyncs's avatar
zhyncs committed
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089

    if completed == 0:
        warnings.warn(
            "All requests failed. This is likely due to a misconfiguration "
            "on the benchmark arguments.",
            stacklevel=2,
        )
    metrics = BenchmarkMetrics(
        completed=completed,
        total_input=total_input,
Ying Sheng's avatar
Ying Sheng committed
1090
1091
        total_output=sum(output_lens),
        total_output_retokenized=sum(retokenized_output_lens),
zhyncs's avatar
zhyncs committed
1092
1093
        request_throughput=completed / dur_s,
        input_throughput=total_input / dur_s,
Ying Sheng's avatar
Ying Sheng committed
1094
1095
        output_throughput=sum(output_lens) / dur_s,
        output_throughput_retokenized=sum(retokenized_output_lens) / dur_s,
1096
1097
1098
        total_throughput=(total_input + sum(output_lens)) / dur_s,
        total_throughput_retokenized=(total_input + sum(retokenized_output_lens))
        / dur_s,
zhyncs's avatar
zhyncs committed
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
        mean_ttft_ms=np.mean(ttfts or 0)
        * 1000,  # ttfts is empty if streaming is not supported by backend
        median_ttft_ms=np.median(ttfts or 0) * 1000,
        std_ttft_ms=np.std(ttfts or 0) * 1000,
        p99_ttft_ms=np.percentile(ttfts or 0, 99) * 1000,
        mean_tpot_ms=np.mean(tpots or 0) * 1000,
        median_tpot_ms=np.median(tpots or 0) * 1000,
        std_tpot_ms=np.std(tpots or 0) * 1000,
        p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
        mean_itl_ms=np.mean(itls or 0) * 1000,
        median_itl_ms=np.median(itls or 0) * 1000,
        std_itl_ms=np.std(itls or 0) * 1000,
1111
        p95_itl_ms=np.percentile(itls or 0, 95) * 1000,
zhyncs's avatar
zhyncs committed
1112
        p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
1113
        max_itl_ms=np.max(itls or 0) * 1000,
zhyncs's avatar
zhyncs committed
1114
1115
        mean_e2e_latency_ms=np.mean(e2e_latencies) * 1000,
        median_e2e_latency_ms=np.median(e2e_latencies) * 1000,
1116
1117
        std_e2e_latency_ms=np.std(e2e_latencies) * 1000,
        p99_e2e_latency_ms=np.percentile(e2e_latencies, 99) * 1000,
1118
        concurrency=np.sum(e2e_latencies) / dur_s,
zhyncs's avatar
zhyncs committed
1119
1120
    )

Ying Sheng's avatar
Ying Sheng committed
1121
    return metrics, output_lens
zhyncs's avatar
zhyncs committed
1122
1123
1124
1125
1126


async def benchmark(
    backend: str,
    api_url: str,
1127
    base_url: str,
zhyncs's avatar
zhyncs committed
1128
1129
    model_id: str,
    tokenizer: PreTrainedTokenizerBase,
1130
    input_requests: List[DatasetRow],
zhyncs's avatar
zhyncs committed
1131
    request_rate: float,
1132
    max_concurrency: Optional[int],
zhyncs's avatar
zhyncs committed
1133
    disable_tqdm: bool,
1134
    lora_names: List[str],
1135
    extra_request_body: Dict[str, Any],
1136
    profile: bool,
1137
    pd_separated: bool = False,
Yineng Zhang's avatar
Yineng Zhang committed
1138
    flush_cache: bool = False,
1139
    warmup_requests: int = 1,
zhyncs's avatar
zhyncs committed
1140
1141
1142
1143
1144
1145
):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
    else:
        raise ValueError(f"Unknown backend: {backend}")

1146
    # Limit concurrency
1147
1148
1149
1150
1151
1152
1153
1154
1155
    # From https://github.com/vllm-project/vllm/pull/9390
    semaphore = asyncio.Semaphore(max_concurrency) if max_concurrency else None

    async def limited_request_func(request_func_input, pbar):
        if semaphore is None:
            return await request_func(request_func_input=request_func_input, pbar=pbar)
        async with semaphore:
            return await request_func(request_func_input=request_func_input, pbar=pbar)

1156
    # Warmup
1157
    print(f"Starting warmup with {warmup_requests} sequences...")
1158
1159

    # Use the first request for all warmup iterations
1160
1161
1162
1163
1164
1165
    test_request = input_requests[0]
    test_prompt, test_prompt_len, test_output_len = (
        test_request.prompt,
        test_request.prompt_len,
        test_request.output_len,
    )
1166
    if lora_names is not None and len(lora_names) != 0:
1167
1168
1169
1170
        lora_name = lora_names[0]
    else:
        lora_name = None

1171
1172
1173
1174
1175
1176
1177
1178
1179
    if "<image>" in test_prompt:
        import re

        image_match = re.search(r"<image>(.*?)</image>(.*)", test_prompt)
        image_data = image_match.group(1) if image_match else None
        test_prompt = image_match.group(2) if image_match else test_prompt
    else:
        image_data = None

1180
    # Create the test input once
zhyncs's avatar
zhyncs committed
1181
1182
1183
1184
1185
    test_input = RequestFuncInput(
        model=model_id,
        prompt=test_prompt,
        api_url=api_url,
        prompt_len=test_prompt_len,
1186
        output_len=min(test_output_len, 32),
1187
        lora_name=lora_name,
1188
        image_data=image_data,
1189
        extra_request_body=extra_request_body,
zhyncs's avatar
zhyncs committed
1190
    )
1191
1192
1193

    # Run warmup requests
    warmup_tasks = []
1194
    for _ in range(warmup_requests):
1195
1196
1197
1198
1199
1200
1201
        warmup_tasks.append(
            asyncio.create_task(request_func(request_func_input=test_input))
        )

    warmup_outputs = await asyncio.gather(*warmup_tasks)

    # Check if at least one warmup request succeeded
1202
    if warmup_requests > 0 and not any(output.success for output in warmup_outputs):
zhyncs's avatar
zhyncs committed
1203
        raise ValueError(
1204
1205
            "Warmup failed - Please make sure benchmark arguments "
            f"are correctly specified. Error: {warmup_outputs[0].error}"
zhyncs's avatar
zhyncs committed
1206
1207
        )
    else:
1208
1209
1210
        print(
            f"Warmup completed with {args.warmup_requests} sequences. Starting main benchmark run..."
        )
zhyncs's avatar
zhyncs committed
1211

1212
    # Flush cache
Yineng Zhang's avatar
Yineng Zhang committed
1213
    if ("sglang" in backend and _get_bool_env_var("SGLANG_IS_IN_CI")) or flush_cache:
1214
        requests.post(base_url + "/flush_cache", headers=get_auth_headers())
1215
1216

    time.sleep(1.0)
1217

1218
    # Start profiler
1219
1220
1221
1222
1223
1224
1225
1226
    if profile:
        print("Starting profiler...")
        profile_output = await async_request_profile(
            api_url=base_url + "/start_profile"
        )
        if profile_output.success:
            print("Profiler started")

zhyncs's avatar
zhyncs committed
1227
1228
    pbar = None if disable_tqdm else tqdm(total=len(input_requests))

1229
    # Run all requests
zhyncs's avatar
zhyncs committed
1230
1231
1232
    benchmark_start_time = time.perf_counter()
    tasks: List[asyncio.Task] = []
    async for request in get_request(input_requests, request_rate):
1233
1234
1235
1236
1237
        prompt, prompt_len, output_len = (
            request.prompt,
            request.prompt_len,
            request.output_len,
        )
1238
        if lora_names is not None and len(lora_names) != 0:
1239
1240
1241
1242
1243
            idx = random.randint(0, len(lora_names) - 1)
            lora_name = lora_names[idx]
        else:
            lora_name = None

1244
1245
1246
1247
1248
1249
1250
1251
1252
        if "<image>" in prompt:
            import re

            image_match = re.search(r"<image>(.*?)</image>(.*)", prompt)
            image_data = image_match.group(1) if image_match else None
            prompt = image_match.group(2) if image_match else prompt
        else:
            image_data = None

zhyncs's avatar
zhyncs committed
1253
1254
1255
1256
1257
1258
        request_func_input = RequestFuncInput(
            model=model_id,
            prompt=prompt,
            api_url=api_url,
            prompt_len=prompt_len,
            output_len=output_len,
1259
            lora_name=lora_name,
1260
            image_data=image_data,
1261
            extra_request_body=extra_request_body,
zhyncs's avatar
zhyncs committed
1262
1263
1264
        )
        tasks.append(
            asyncio.create_task(
1265
                limited_request_func(request_func_input=request_func_input, pbar=pbar)
zhyncs's avatar
zhyncs committed
1266
1267
1268
1269
            )
        )
    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)

1270
    # Stop profiler
1271
1272
1273
1274
1275
1276
    if profile:
        print("Stopping profiler...")
        profile_output = await async_request_profile(api_url=base_url + "/stop_profile")
        if profile_output.success:
            print("Profiler stopped")

zhyncs's avatar
zhyncs committed
1277
1278
1279
    if pbar is not None:
        pbar.close()

1280
1281
    if "sglang" in backend:
        server_info = requests.get(base_url + "/get_server_info")
1282
1283
        if pd_separated:
            accept_length = server_info.json()["decode"][0]["internal_states"][0].get(
1284
1285
1286
                "avg_spec_accept_length", None
            )
        else:
1287
1288
1289
            accept_length = server_info.json()["internal_states"][0].get(
                "avg_spec_accept_length", None
            )
1290
1291
1292
    else:
        accept_length = None

1293
    # Compute metrics and print results
zhyncs's avatar
zhyncs committed
1294
    benchmark_duration = time.perf_counter() - benchmark_start_time
Ying Sheng's avatar
Ying Sheng committed
1295
    metrics, output_lens = calculate_metrics(
zhyncs's avatar
zhyncs committed
1296
1297
1298
1299
        input_requests=input_requests,
        outputs=outputs,
        dur_s=benchmark_duration,
        tokenizer=tokenizer,
1300
        backend=backend,
zhyncs's avatar
zhyncs committed
1301
1302
1303
    )

    print("\n{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
1304
    print("{:<40} {:<10}".format("Backend:", backend))
zhyncs's avatar
zhyncs committed
1305
    print("{:<40} {:<10}".format("Traffic request rate:", request_rate))
1306
1307
    print(
        "{:<40} {:<10}".format(
1308
            "Max request concurrency:",
1309
1310
1311
            max_concurrency if max_concurrency else "not set",
        )
    )
zhyncs's avatar
zhyncs committed
1312
1313
1314
1315
    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
    print("{:<40} {:<10.2f}".format("Benchmark duration (s):", benchmark_duration))
    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
    print("{:<40} {:<10}".format("Total generated tokens:", metrics.total_output))
Ying Sheng's avatar
Ying Sheng committed
1316
1317
1318
1319
1320
    print(
        "{:<40} {:<10}".format(
            "Total generated tokens (retokenized):", metrics.total_output_retokenized
        )
    )
zhyncs's avatar
zhyncs committed
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
    print(
        "{:<40} {:<10.2f}".format(
            "Request throughput (req/s):", metrics.request_throughput
        )
    )
    print(
        "{:<40} {:<10.2f}".format(
            "Input token throughput (tok/s):", metrics.input_throughput
        )
    )
    print(
        "{:<40} {:<10.2f}".format(
            "Output token throughput (tok/s):", metrics.output_throughput
        )
    )
1336
1337
1338
1339
1340
    print(
        "{:<40} {:<10.2f}".format(
            "Total token throughput (tok/s):", metrics.total_throughput
        )
    )
1341
    print("{:<40} {:<10.2f}".format("Concurrency:", metrics.concurrency))
1342
1343
    if accept_length:
        print("{:<40} {:<10.2f}".format("Accept length:", accept_length))
zhyncs's avatar
zhyncs committed
1344
1345
1346
1347
1348
1349
1350
1351
1352
    print("{s:{c}^{n}}".format(s="End-to-End Latency", n=50, c="-"))
    print(
        "{:<40} {:<10.2f}".format("Mean E2E Latency (ms):", metrics.mean_e2e_latency_ms)
    )
    print(
        "{:<40} {:<10.2f}".format(
            "Median E2E Latency (ms):", metrics.median_e2e_latency_ms
        )
    )
zhyncs's avatar
zhyncs committed
1353
1354
1355
1356
    print("{s:{c}^{n}}".format(s="Time to First Token", n=50, c="-"))
    print("{:<40} {:<10.2f}".format("Mean TTFT (ms):", metrics.mean_ttft_ms))
    print("{:<40} {:<10.2f}".format("Median TTFT (ms):", metrics.median_ttft_ms))
    print("{:<40} {:<10.2f}".format("P99 TTFT (ms):", metrics.p99_ttft_ms))
1357
    print("{s:{c}^{n}}".format(s="Inter-Token Latency", n=50, c="-"))
zhyncs's avatar
zhyncs committed
1358
1359
    print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
    print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
1360
    print("{:<40} {:<10.2f}".format("P95 ITL (ms):", metrics.p95_itl_ms))
zhyncs's avatar
zhyncs committed
1361
    print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
1362
    print("{:<40} {:<10.2f}".format("Max ITL (ms):", metrics.max_itl_ms))
zhyncs's avatar
zhyncs committed
1363
1364
    print("=" * 50)

zhyncs's avatar
zhyncs committed
1365
1366
1367
1368
1369
1370
    if (
        metrics.median_ttft_ms is not None
        and metrics.mean_itl_ms is not None
        and metrics.output_throughput is not None
    ):
        result = {
1371
            # Arguments
zhyncs's avatar
zhyncs committed
1372
1373
1374
            "backend": args.backend,
            "dataset_name": args.dataset_name,
            "request_rate": request_rate,
1375
            "max_concurrency": max_concurrency,
1376
1377
1378
1379
1380
1381
1382
            "sharegpt_output_len": args.sharegpt_output_len,
            "random_input_len": args.random_input_len,
            "random_output_len": args.random_output_len,
            "random_range_ratio": args.random_range_ratio,
            # Results
            "duration": benchmark_duration,
            "completed": metrics.completed,
1383
1384
1385
            "total_input_tokens": metrics.total_input,
            "total_output_tokens": metrics.total_output,
            "total_output_tokens_retokenized": metrics.total_output_retokenized,
1386
1387
1388
            "request_throughput": metrics.request_throughput,
            "input_throughput": metrics.input_throughput,
            "output_throughput": metrics.output_throughput,
1389
1390
            "mean_e2e_latency_ms": metrics.mean_e2e_latency_ms,
            "median_e2e_latency_ms": metrics.median_e2e_latency_ms,
1391
1392
            "std_e2e_latency_ms": metrics.std_e2e_latency_ms,
            "p99_e2e_latency_ms": metrics.p99_e2e_latency_ms,
1393
            "mean_ttft_ms": metrics.mean_ttft_ms,
1394
            "median_ttft_ms": metrics.median_ttft_ms,
1395
1396
1397
1398
1399
1400
            "std_ttft_ms": metrics.std_ttft_ms,
            "p99_ttft_ms": metrics.p99_ttft_ms,
            "mean_tpot_ms": metrics.mean_tpot_ms,
            "median_tpot_ms": metrics.median_tpot_ms,
            "std_tpot_ms": metrics.std_tpot_ms,
            "p99_tpot_ms": metrics.p99_tpot_ms,
1401
            "mean_itl_ms": metrics.mean_itl_ms,
1402
            "median_itl_ms": metrics.median_itl_ms,
1403
            "std_itl_ms": metrics.std_itl_ms,
1404
            "p95_itl_ms": metrics.p95_itl_ms,
1405
            "p99_itl_ms": metrics.p99_itl_ms,
1406
            "concurrency": metrics.concurrency,
1407
            "accept_length": accept_length,
zhyncs's avatar
zhyncs committed
1408
1409
1410
1411
        }
    else:
        print(f"Error running benchmark for request rate: {request_rate}")
        print("-" * 30)
1412

zhyncs's avatar
zhyncs committed
1413
1414
1415
1416
1417
    # Determine output file name
    if args.output_file:
        output_file_name = args.output_file
    else:
        now = datetime.now().strftime("%m%d")
1418
        if args.dataset_name.startswith("random"):
zhyncs's avatar
zhyncs committed
1419
            output_file_name = f"{args.backend}_{now}_{args.num_prompts}_{args.random_input_len}_{args.random_output_len}.jsonl"
1420
        else:
zhyncs's avatar
zhyncs committed
1421
            output_file_name = f"{args.backend}_{now}_{args.num_prompts}_sharegpt.jsonl"
1422

1423
1424
1425
1426
1427
1428
1429
1430
1431
    result_details = {
        "input_lens": [output.prompt_len for output in outputs],
        "output_lens": 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],
    }

zhyncs's avatar
zhyncs committed
1432
1433
    # Append results to a JSONL file
    with open(output_file_name, "a") as file:
1434
1435
1436
1437
1438
1439
1440
        if args.output_details:
            result_for_dump = result | result_details
        else:
            result_for_dump = result
        file.write(json.dumps(result_for_dump) + "\n")

    return result | result_details
zhyncs's avatar
zhyncs committed
1441
1442


1443
1444
1445
1446
1447
1448
1449
1450
1451
def check_chat_template(model_path):
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
        return "chat_template" in tokenizer.init_kwargs
    except Exception as e:
        print(f"Fail to load tokenizer config with error={e}")
        return False


1452
1453
1454
1455
1456
1457
def set_global_args(args_: argparse.Namespace):
    """Set the global args."""
    global args
    args = args_


1458
1459
1460
1461
def run_benchmark(args_: argparse.Namespace):
    global args
    args = args_

1462
1463
1464
1465
    # Set default value for max_concurrency if not present
    if not hasattr(args, "max_concurrency"):
        args.max_concurrency = None

1466
1467
1468
1469
    # Set default value for warmup_requests if not present
    if not hasattr(args, "warmup_requests"):
        args.warmup_requests = 1

1470
1471
1472
    if not hasattr(args, "output_details"):
        args.output_details = False

1473
1474
    print(f"benchmark_args={args}")

Lianmin Zheng's avatar
Lianmin Zheng committed
1475
    # Set global environments
1476
    set_ulimit()
zhyncs's avatar
zhyncs committed
1477
1478
1479
    random.seed(args.seed)
    np.random.seed(args.seed)

1480
1481
1482
1483
    extra_request_body = {}
    if args.extra_request_body:
        extra_request_body = json.loads(args.extra_request_body)

Lianmin Zheng's avatar
Lianmin Zheng committed
1484
    # Set url
zhyncs's avatar
zhyncs committed
1485
1486
1487
    if args.port is None:
        args.port = {
            "sglang": 30000,
1488
1489
            "sglang-native": 30000,
            "sglang-oai": 30000,
zhyncs's avatar
zhyncs committed
1490
1491
            "lmdeploy": 23333,
            "vllm": 8000,
1492
            "trt": 8000,
1493
            "gserver": 9988,
1494
            "truss": 8080,
zhyncs's avatar
zhyncs committed
1495
1496
1497
1498
1499
1500
1501
1502
        }.get(args.backend, 30000)

    model_url = (
        f"{args.base_url}/v1/models"
        if args.base_url
        else f"http://{args.host}:{args.port}/v1/models"
    )

1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
    if args.backend in ["sglang", "sglang-native"]:
        api_url = (
            f"{args.base_url}/generate"
            if args.base_url
            else f"http://{args.host}:{args.port}/generate"
        )
    elif args.backend in ["sglang-oai", "vllm", "lmdeploy"]:
        api_url = (
            f"{args.base_url}/v1/completions"
            if args.base_url
            else f"http://{args.host}:{args.port}/v1/completions"
        )
    elif args.backend == "trt":
1516
1517
1518
1519
1520
1521
1522
1523
        api_url = (
            f"{args.base_url}/v2/models/ensemble/generate_stream"
            if args.base_url
            else f"http://{args.host}:{args.port}/v2/models/ensemble/generate_stream"
        )
        if args.model is None:
            print("Please provide a model using `--model` when using `trt` backend.")
            sys.exit(1)
1524
    elif args.backend == "gserver":
Lianmin Zheng's avatar
Lianmin Zheng committed
1525
1526
        api_url = args.base_url if args.base_url else f"{args.host}:{args.port}"
        args.model = args.model or "default"
1527
1528
1529
1530
1531
1532
    elif args.backend == "truss":
        api_url = (
            f"{args.base_url}/v1/models/model:predict"
            if args.base_url
            else f"http://{args.host}:{args.port}/v1/models/model:predict"
        )
1533
1534
1535
    base_url = (
        f"http://{args.host}:{args.port}" if args.base_url is None else args.base_url
    )
1536

Lianmin Zheng's avatar
Lianmin Zheng committed
1537
    # Get model name
zhyncs's avatar
zhyncs committed
1538
    if args.model is None:
1539
1540
1541
1542
1543
        if args.backend == "truss":
            print(
                "Please provide a model with `--model` when using truss backend. e.g. --model meta-llama/Llama-3.1-8B-Instruct"
            )
            sys.exit(1)
zhyncs's avatar
zhyncs committed
1544
        try:
1545
            response = requests.get(model_url, headers=get_auth_headers())
zhyncs's avatar
zhyncs committed
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
            model_list = response.json().get("data", [])
            args.model = model_list[0]["id"] if model_list else None
        except Exception as e:
            print(f"Failed to fetch model from {model_url}. Error: {e}")
            print(
                "Please specify the correct host and port using `--host` and `--port`."
            )
            sys.exit(1)

    if args.model is None:
        print("No model specified or found. Please provide a model using `--model`.")
        sys.exit(1)

1559
1560
1561
1562
1563
1564
    if not check_chat_template(args.model):
        print(
            "\nWARNING It is recommended to use the `Chat` or `Instruct` model for benchmarking.\n"
            "Because when the tokenizer counts the output tokens, if there is gibberish, it might count incorrectly.\n"
        )

zhyncs's avatar
zhyncs committed
1565
1566
    print(f"{args}\n")

Lianmin Zheng's avatar
Lianmin Zheng committed
1567
    # Read dataset
zhyncs's avatar
zhyncs committed
1568
1569
1570
1571
    backend = args.backend
    model_id = args.model
    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
    tokenizer = get_tokenizer(tokenizer_id)
1572
    input_requests = get_dataset(args, tokenizer)
zhyncs's avatar
zhyncs committed
1573

Yineng Zhang's avatar
Yineng Zhang committed
1574
1575
1576
1577
    # compatible with SimpleNamespace
    if not hasattr(args, "flush_cache"):
        args.flush_cache = False

1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
    return asyncio.run(
        benchmark(
            backend=backend,
            api_url=api_url,
            base_url=base_url,
            model_id=model_id,
            tokenizer=tokenizer,
            input_requests=input_requests,
            request_rate=args.request_rate,
            max_concurrency=args.max_concurrency,
            disable_tqdm=args.disable_tqdm,
1589
            lora_names=args.lora_name,
1590
1591
            extra_request_body=extra_request_body,
            profile=args.profile,
1592
            pd_separated=args.pd_separated,
Yineng Zhang's avatar
Yineng Zhang committed
1593
            flush_cache=args.flush_cache,
Lianmin Zheng's avatar
Lianmin Zheng committed
1594
        )
1595
    )
zhyncs's avatar
zhyncs committed
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608


def set_ulimit(target_soft_limit=65535):
    resource_type = resource.RLIMIT_NOFILE
    current_soft, current_hard = resource.getrlimit(resource_type)

    if current_soft < target_soft_limit:
        try:
            resource.setrlimit(resource_type, (target_soft_limit, current_hard))
        except ValueError as e:
            print(f"Fail to set RLIMIT_NOFILE: {e}")


1609
1610
1611
1612
1613
1614
1615
class LoRAPathAction(argparse.Action):
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, [])
        for lora_name in values:
            getattr(namespace, self.dest).append(lora_name)


zhyncs's avatar
zhyncs committed
1616
if __name__ == "__main__":
1617
    parser = ArgumentParser(description="Benchmark the online serving throughput.")
zhyncs's avatar
zhyncs committed
1618
1619
1620
1621
    parser.add_argument(
        "--backend",
        type=str,
        choices=list(ASYNC_REQUEST_FUNCS.keys()),
1622
        default="sglang",
zhyncs's avatar
zhyncs committed
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
        help="Must specify a backend, depending on the LLM Inference Engine.",
    )
    parser.add_argument(
        "--base-url",
        type=str,
        default=None,
        help="Server or API base url if not using http host and port.",
    )
    parser.add_argument(
        "--host", type=str, default="0.0.0.0", help="Default host is 0.0.0.0."
    )
    parser.add_argument(
        "--port",
        type=int,
        help="If not set, the default port is configured according to its default value for different LLM Inference Engines.",
    )
    parser.add_argument(
1640
1641
1642
        "--dataset-name",
        type=str,
        default="sharegpt",
1643
        choices=["sharegpt", "random", "random-ids", "generated-shared-prefix", "mmmu"],
1644
1645
1646
1647
        help="Name of the dataset to benchmark on.",
    )
    parser.add_argument(
        "--dataset-path", type=str, default="", help="Path to the dataset."
zhyncs's avatar
zhyncs committed
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
    )
    parser.add_argument(
        "--model",
        type=str,
        help="Name or path of the model. If not set, the default model will request /v1/models for conf.",
    )
    parser.add_argument(
        "--tokenizer",
        type=str,
        help="Name or path of the tokenizer. If not set, using the model conf.",
    )
    parser.add_argument(
        "--num-prompts",
        type=int,
        default=1000,
        help="Number of prompts to process. Default is 1000.",
    )
    parser.add_argument(
        "--sharegpt-output-len",
        type=int,
        default=None,
        help="Output length for each request. Overrides the output length from the ShareGPT dataset.",
    )
1671
1672
1673
1674
1675
1676
    parser.add_argument(
        "--sharegpt-context-len",
        type=int,
        default=None,
        help="The context length of the model for the ShareGPT dataset. Requests longer than the context length will be dropped.",
    )
1677
1678
1679
    parser.add_argument(
        "--random-input-len",
        type=int,
1680
        default=1024,
1681
1682
1683
1684
        help="Number of input tokens per request, used only for random dataset.",
    )
    parser.add_argument(
        "--random-output-len",
1685
        default=1024,
1686
1687
1688
1689
1690
1691
        type=int,
        help="Number of output tokens per request, used only for random dataset.",
    )
    parser.add_argument(
        "--random-range-ratio",
        type=float,
Yineng Zhang's avatar
Yineng Zhang committed
1692
        default=0.0,
1693
1694
1695
        help="Range of sampled ratio of input/output length, "
        "used only for random dataset.",
    )
zhyncs's avatar
zhyncs committed
1696
1697
1698
    parser.add_argument(
        "--request-rate",
        type=float,
1699
        default=float("inf"),
zhyncs's avatar
zhyncs committed
1700
        help="Number of requests per second. If this is inf, then all the requests are sent at time 0. "
min-xu-et's avatar
min-xu-et committed
1701
        "Otherwise, we use Poisson process to synthesize the request arrival times. Default is inf.",
zhyncs's avatar
zhyncs committed
1702
    )
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
    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, "
        "if the server is not processing requests fast enough to keep up.",
    )
1716
    parser.add_argument("--output-file", type=str, help="Output JSONL file name.")
1717
1718
1719
    parser.add_argument(
        "--output-details", action="store_true", help="Output details of benchmarking."
    )
1720
1721
1722
1723
1724
    parser.add_argument(
        "--disable-tqdm",
        action="store_true",
        help="Specify to disable tqdm progress bar.",
    )
1725
1726
1727
1728
1729
    parser.add_argument(
        "--disable-stream",
        action="store_true",
        help="Disable streaming mode.",
    )
1730
    parser.add_argument(
1731
        "--return-logprob",
1732
        action="store_true",
1733
        help="Return logprob.",
1734
    )
1735
    parser.add_argument("--seed", type=int, default=1, help="The random seed.")
1736
    parser.add_argument(
1737
        "--disable-ignore-eos",
1738
        action="store_true",
1739
        help="Disable ignoring EOS.",
1740
    )
1741
1742
1743
1744
1745
1746
1747
    parser.add_argument(
        "--extra-request-body",
        metavar='{"key1": "value1", "key2": "value2"}',
        type=str,
        help="Append given JSON object to the request payload. You can use this to specify"
        "additional generate params like sampling params.",
    )
1748
1749
1750
1751
1752
    parser.add_argument(
        "--apply-chat-template",
        action="store_true",
        help="Apply chat template",
    )
1753
1754
1755
1756
1757
1758
1759
1760
1761
    parser.add_argument(
        "--profile",
        action="store_true",
        help="Use Torch Profiler. The endpoint must be launched with "
        "SGLANG_TORCH_PROFILER_DIR to enable profiler.",
    )
    parser.add_argument(
        "--lora-name",
        type=str,
1762
        nargs="*",
1763
        default=None,
1764
1765
        action=LoRAPathAction,
        help="The names of LoRA adapters. You can provide a list of names in the format {name} {name} {name}...",
1766
    )
1767
1768
1769
1770
1771
1772
1773
    parser.add_argument(
        "--prompt-suffix",
        type=str,
        default="",
        help="Suffix applied to the end of all user prompts, followed by assistant prompt suffix.",
    )
    parser.add_argument(
Yineng Zhang's avatar
Yineng Zhang committed
1774
        "--pd-separated",
1775
1776
1777
        action="store_true",
        help="Benchmark PD disaggregation server",
    )
Yineng Zhang's avatar
Yineng Zhang committed
1778
1779
1780
1781
1782
    parser.add_argument(
        "--flush-cache",
        action="store_true",
        help="Flush the cache before running the benchmark",
    )
1783
1784
1785
1786
1787
1788
    parser.add_argument(
        "--warmup-requests",
        type=int,
        default=1,
        help="Number of warmup requests to run before the benchmark",
    )
1789
1790
1791

    group = parser.add_argument_group("generated-shared-prefix dataset arguments")
    group.add_argument(
1792
        "--gsp-num-groups",
1793
1794
1795
1796
1797
        type=int,
        default=64,
        help="Number of system prompt groups for generated-shared-prefix dataset",
    )
    group.add_argument(
1798
        "--gsp-prompts-per-group",
1799
1800
1801
1802
1803
        type=int,
        default=16,
        help="Number of prompts per system prompt group for generated-shared-prefix dataset",
    )
    group.add_argument(
1804
        "--gsp-system-prompt-len",
1805
1806
1807
1808
1809
        type=int,
        default=2048,
        help="Target length in tokens for system prompts in generated-shared-prefix dataset",
    )
    group.add_argument(
1810
        "--gsp-question-len",
1811
1812
1813
1814
1815
        type=int,
        default=128,
        help="Target length in tokens for questions in generated-shared-prefix dataset",
    )
    group.add_argument(
1816
        "--gsp-output-len",
1817
1818
1819
1820
        type=int,
        default=256,
        help="Target length in tokens for outputs in generated-shared-prefix dataset",
    )
zhyncs's avatar
zhyncs committed
1821
    args = parser.parse_args()
1822
    run_benchmark(args)