benchmark_throughput.py 21.6 KB
Newer Older
1
"""Benchmark offline inference throughput."""
2
import argparse
3
import dataclasses
4
5
6
import json
import random
import time
7
8
from functools import cache
from typing import Dict, List, Optional, Tuple
9

zhuwenwen's avatar
zhuwenwen committed
10
import numpy as np
11
import torch
12
import uvloop
13
from PIL import Image
14
from tqdm import tqdm
15
16
from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          PreTrainedTokenizerBase)
17

zhuwenwen's avatar
zhuwenwen committed
18
19

from vllm.inputs import PromptType
20
from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
21
22
from vllm.entrypoints.openai.api_server import (
    build_async_engine_client_from_engine_args)
23
from vllm.inputs import TextPrompt
24
25
from vllm.lora.request import LoRARequest
from vllm.lora.utils import get_adapter_absolute_path
26
from vllm.multimodal import MultiModalDataDict
27
from vllm.sampling_params import BeamSearchParams
28
from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
29
from vllm.utils import FlexibleArgumentParser, merge_async_iterators
30

31

32
33
34
35
36
37
38
39
@dataclasses.dataclass
class SampleRequest:
    """A class representing a single inference request for benchmarking.

    Attributes:
        prompt: The input text prompt for the model.
        prompt_len: The length of the prompt in tokens.
        expected_output_len: The expected length of the output in tokens.
40
41
42
        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
            images).
        lora_request: Optional LoRARequest specifying the LoRA to use. 
43
44
45
46
47
    """
    prompt: str
    prompt_len: int
    expected_output_len: int
    multi_modal_data: Optional[MultiModalDataDict] = None
48
    lora_request: Optional[LoRARequest] = None
49
50


51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
def _get_prompt_for_image_model(question: str, *, model: str) -> str:
    """Prepend and append special tokens around the question to form a prompt.

    Args:
        question: The input question text to wrap with special tokens
        model: The name of the model being used, to determine which special
            tokens to add

    Returns:
        The formatted prompt string with appropriate special tokens for the
            model

    Raises:
        ValueError: If an unsupported model name is provided
    """
    model = model.lower()
    if "pixtral" in model:
        return f"<s>[INST]{question}\n[IMG][/INST]"
    raise ValueError(f"Unsupported model {model}")


72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
@cache
def lora_path_on_disk(lora_path: str) -> str:
    return get_adapter_absolute_path(lora_path)


lora_tokenizer_cache: Dict[int, AnyTokenizer] = {}


def get_random_lora_request(
        args: argparse.Namespace
) -> Tuple[LoRARequest, Optional[AnyTokenizer]]:
    global lora_tokenizer_cache
    lora_id = random.randint(1, args.max_loras)
    lora_request = LoRARequest(lora_name=str(lora_id),
                               lora_int_id=lora_id,
                               lora_path=lora_path_on_disk(args.lora_path))
    if lora_id not in lora_tokenizer_cache:
        lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
    return lora_request, lora_tokenizer_cache[lora_id]


93
94
def sample_requests(tokenizer: PreTrainedTokenizerBase,
                    args: argparse.Namespace) -> List[SampleRequest]:
95

96
97
98
99
    dataset_path: str = args.dataset
    num_requests: int = args.num_prompts
    fixed_output_len: Optional[int] = args.output_len
    model: str = args.model
100
101
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")
102

103
104
105
106
    # Load the dataset.
    with open(dataset_path) as f:
        dataset = json.load(f)
    # Filter out the conversations with less than 2 turns.
107
    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
108
109
    # Shuffle the dataset.
    random.shuffle(dataset)
110

111
    # Filter out sequences that are too long or too short
112
    filtered_dataset: List[SampleRequest] = []
113
114
115
    for data in tqdm(dataset,
                     total=len(filtered_dataset),
                     desc="sampling requests"):
116
117
118
        if len(filtered_dataset) == num_requests:
            break

119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
        # Only keep the first two turns of each conversation.
        prompt = data["conversations"][0]["value"]
        completion = data["conversations"][1]["value"]

        multi_modal_data: Optional[MultiModalDataDict] = None
        if "image" in data:
            multi_modal_data = multi_modal_data or {}
            image_path = data["image"]
            # TODO(vllm-project/vllm/issues/9778): Support multiple images.
            assert isinstance(image_path,
                              str), "Only support single image input"
            try:
                multi_modal_data["image"] = Image.open(image_path).convert(
                    "RGB")
            except FileNotFoundError:
                # Ignore datapoint where asset is missing
                continue
            prompt = _get_prompt_for_image_model(question=prompt, model=model)

138
139
140
141
142
143
144
        request_tokenizer = tokenizer
        lora_request: Optional[LoRARequest] = None
        if args.enable_lora:
            lora_request, lora_tokenizer = get_random_lora_request(args)
            if lora_tokenizer:
                request_tokenizer = lora_tokenizer

145
        # Tokenize the prompts and completions.
146
147
        prompt_token_ids = request_tokenizer(prompt).input_ids
        completion_token_ids = request_tokenizer(completion).input_ids
148
        prompt_len = len(prompt_token_ids)
149
150
        output_len = len(completion_token_ids
                         ) if fixed_output_len is None else fixed_output_len
151
152
153
154
155
156
        if prompt_len < 4 or output_len < 4:
            # Prune too short sequences.
            continue
        if prompt_len > 1024 or prompt_len + output_len > 2048:
            # Prune too long sequences.
            continue
157
158
159
        filtered_dataset.append(
            SampleRequest(prompt=prompt,
                          prompt_len=prompt_len,
160
                          expected_output_len=output_len,
161
162
                          multi_modal_data=multi_modal_data,
                          lora_request=lora_request))
163

164
    return filtered_dataset
165
166


Woosuk Kwon's avatar
Woosuk Kwon committed
167
def run_vllm(
zhuwenwen's avatar
zhuwenwen committed
168
    warmup_requests: List[SampleRequest],
169
    requests: List[SampleRequest],
170
    n: int,
171
    engine_args: EngineArgs,
172
) -> float:
173
    from vllm import LLM, SamplingParams
174
    llm = LLM(**dataclasses.asdict(engine_args))
175

Zhuohan Li's avatar
Zhuohan Li committed
176
    # Add the requests to the engine.
177
    prompts: List[TextPrompt] = []
178
    sampling_params: List[SamplingParams] = []
179
    for request in requests:
180
181
182
        prompts.append(
            TextPrompt(prompt=request.prompt,
                       multi_modal_data=request.multi_modal_data))
183
184
185
        sampling_params.append(
            SamplingParams(
                n=n,
186
                temperature=1.0,
187
188
                top_p=1.0,
                ignore_eos=True,
189
                max_tokens=request.expected_output_len,
190
            ))
191
192
193
    lora_requests: Optional[List[LoRARequest]] = None
    if engine_args.enable_lora:
        lora_requests = [request.lora_request for request in requests]
194

zhuwenwen's avatar
zhuwenwen committed
195
    # warmup
zhuwenwen's avatar
zhuwenwen committed
196
197
    warmup_prompts: List[TextPrompt] = []
    warmup_sampling_params: List[SamplingParams] = []
198
    for request in warmup_requests:
zhuwenwen's avatar
zhuwenwen committed
199
200
201
        warmup_prompts.append(
            TextPrompt(prompt=request.prompt,
                       multi_modal_data=request.multi_modal_data))
zhuwenwen's avatar
zhuwenwen committed
202
203
204
        warmup_sampling_params.append(
            SamplingParams(
                n=n,
zhuwenwen's avatar
zhuwenwen committed
205
                temperature=1.0,
zhuwenwen's avatar
zhuwenwen committed
206
207
                top_p=1.0,
                ignore_eos=True,
zhuwenwen's avatar
zhuwenwen committed
208
                max_tokens=request.expected_output_len,
zhuwenwen's avatar
zhuwenwen committed
209
            ))
210
        
zhuwenwen's avatar
zhuwenwen committed
211
212
    print("Warming up...")
    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
213
        llm.generate(warmup_prompts, warmup_sampling_params, use_tqdm=True)
zhuwenwen's avatar
zhuwenwen committed
214
215
216
217
    
    # dummy_prompt_token_ids = np.random.randint(10000,
    #                                            size=(args.num_prompts,
    #                                                  args.input_len))
zhuwenwen's avatar
zhuwenwen committed
218
    # dummy_prompts: List[PromptType] = [{
zhuwenwen's avatar
zhuwenwen committed
219
220
    #     "prompt_token_ids": batch
    # } for batch in dummy_prompt_token_ids.tolist()]
zhuwenwen's avatar
zhuwenwen committed
221
222
223
    
    # def run_to_completion(profile_dir: Optional[str] = None):
    #     llm.generate(dummy_prompts,
zhuwenwen's avatar
zhuwenwen committed
224
225
    #                     sampling_params=sampling_params,
    #                     use_tqdm=False)
zhuwenwen's avatar
zhuwenwen committed
226
    
zhuwenwen's avatar
zhuwenwen committed
227
228
    # print("Warming up...")
    # for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
zhuwenwen's avatar
zhuwenwen committed
229
    #     run_to_completion(profile_dir=None)
zhuwenwen's avatar
zhuwenwen committed
230

231
232
233
    use_beam_search = False

    if not use_beam_search:
234
        start = time.perf_counter()
235
236
237
238
        llm.generate(prompts,
                     sampling_params,
                     lora_request=lora_requests,
                     use_tqdm=True)
239
240
        end = time.perf_counter()
    else:
241
        assert lora_requests is None, "BeamSearch API does not support LoRA"
242
        prompts = [request.prompt for request in requests]
243
244
        # output_len should be the same for all requests.
        output_len = requests[0][2]
245
246
        for request in requests:
            assert request.expected_output_len == output_len
247
        start = time.perf_counter()
248
249
250
251
252
253
254
        llm.beam_search(
            prompts,
            BeamSearchParams(
                beam_width=n,
                max_tokens=output_len,
                ignore_eos=True,
            ))
255
        end = time.perf_counter()
256
257
258
    return end - start


259
async def run_vllm_async(
260
    requests: List[SampleRequest],
261
    n: int,
262
    engine_args: AsyncEngineArgs,
263
264
265
266
267
268
269
270
    disable_frontend_multiprocessing: bool = False,
) -> float:
    from vllm import SamplingParams

    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:

        # Add the requests to the engine.
271
        prompts: List[TextPrompt] = []
272
        sampling_params: List[SamplingParams] = []
273
        lora_requests: List[Optional[LoRARequest]] = []
274
        for request in requests:
275
276
277
            prompts.append(
                TextPrompt(prompt=request.prompt,
                           multi_modal_data=request.multi_modal_data))
278
279
280
            sampling_params.append(
                SamplingParams(
                    n=n,
281
                    temperature=1.0,
282
283
                    top_p=1.0,
                    ignore_eos=True,
284
                    max_tokens=request.expected_output_len,
285
                ))
286
            lora_requests.append(request.lora_request)
287
288
289

        generators = []
        start = time.perf_counter()
290
291
292
293
294
295
        for i, (prompt, sp,
                lr) in enumerate(zip(prompts, sampling_params, lora_requests)):
            generator = llm.generate(prompt,
                                     sp,
                                     lora_request=lr,
                                     request_id=f"test{i}")
296
297
298
299
300
301
302
303
            generators.append(generator)
        all_gens = merge_async_iterators(*generators)
        async for i, res in all_gens:
            pass
        end = time.perf_counter()
        return end - start


304
def run_hf(
305
    requests: List[SampleRequest],
306
307
308
309
    model: str,
    tokenizer: PreTrainedTokenizerBase,
    n: int,
    max_batch_size: int,
310
    trust_remote_code: bool,
311
) -> float:
312
313
    llm = AutoModelForCausalLM.from_pretrained(
        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
314
315
316
    if llm.config.model_type == "llama":
        # To enable padding in the HF backend.
        tokenizer.pad_token = tokenizer.eos_token
317
318
319
    llm = llm.cuda()

    pbar = tqdm(total=len(requests))
320
    start = time.perf_counter()
321
322
323
324
325
326
327
328
329
330
331
332
    batch: List[str] = []
    max_prompt_len = 0
    max_output_len = 0
    for i in range(len(requests)):
        prompt, prompt_len, output_len = requests[i]
        # Add the prompt to the batch.
        batch.append(prompt)
        max_prompt_len = max(max_prompt_len, prompt_len)
        max_output_len = max(max_output_len, output_len)
        if len(batch) < max_batch_size and i != len(requests) - 1:
            # Check if we can add more requests to the batch.
            _, next_prompt_len, next_output_len = requests[i + 1]
333
334
            if (max(max_prompt_len, next_prompt_len) +
                    max(max_output_len, next_output_len)) <= 2048:
335
336
337
338
                # We can add more requests to the batch.
                continue

        # Generate the sequences.
339
340
        input_ids = tokenizer(batch, return_tensors="pt",
                              padding=True).input_ids
341
342
        llm_outputs = llm.generate(
            input_ids=input_ids.cuda(),
343
            do_sample=True,
344
345
346
347
348
349
350
351
352
353
354
355
356
357
            num_return_sequences=n,
            temperature=1.0,
            top_p=1.0,
            use_cache=True,
            max_new_tokens=max_output_len,
        )
        # Include the decoding time.
        tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
        pbar.update(len(batch))

        # Clear the batch.
        batch = []
        max_prompt_len = 0
        max_output_len = 0
358
    end = time.perf_counter()
359
360
361
    return end - start


362
def run_mii(
363
    requests: List[SampleRequest],
364
365
366
367
    model: str,
    tensor_parallel_size: int,
    output_len: int,
) -> float:
368
369
    from mii import client, serve
    llm = serve(model, tensor_parallel=tensor_parallel_size)
370
    prompts = [request.prompt for request in requests]
371
372

    start = time.perf_counter()
373
    llm.generate(prompts, max_new_tokens=output_len)
374
    end = time.perf_counter()
375
376
    client = client(model)
    client.terminate_server()
377
378
379
    return end - start


380
381
382
383
384
def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)

    # Sample the requests.
385
386
    tokenizer = AutoTokenizer.from_pretrained(
        args.tokenizer, trust_remote_code=args.trust_remote_code)
387
388
389
    warmup_prompt = "hi" * 10
    warmup_requests = [(warmup_prompt, 10, 10)
                for _ in range(1)]
390
    if args.dataset is None:
391
392
393
        vocab_size = tokenizer.vocab_size
        requests = []
        for _ in range(args.num_prompts):
394
395
396
397
398
399
400
401

            request_tokenizer = tokenizer
            lora_request: Optional[LoRARequest] = None
            if args.enable_lora:
                lora_request, lora_tokenizer = get_random_lora_request(args)
                if lora_tokenizer:
                    request_tokenizer = lora_tokenizer

402
403
404
405
406
407
408
409
            # Synthesize a prompt with the given input length.
            candidate_ids = [
                random.randint(0, vocab_size - 1)
                for _ in range(args.input_len)
            ]
            # As tokenizer may add additional tokens like BOS, we need to try
            # different lengths to get the desired input length.
            for _ in range(5):  # Max attempts to correct
410
411
                candidate_prompt = request_tokenizer.decode(candidate_ids)
                tokenized_len = len(request_tokenizer.encode(candidate_prompt))
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427

                if tokenized_len == args.input_len:
                    break

                # Adjust length based on difference
                diff = args.input_len - tokenized_len
                if diff > 0:
                    candidate_ids.extend([
                        random.randint(100, vocab_size - 100)
                        for _ in range(diff)
                    ])
                else:
                    candidate_ids = candidate_ids[:diff]
            requests.append(
                SampleRequest(prompt=candidate_prompt,
                              prompt_len=args.input_len,
428
429
                              expected_output_len=args.output_len,
                              lora_request=lora_request))
430
    else:
431
        requests = sample_requests(tokenizer, args)
432

433
434
    is_multi_modal = any(request.multi_modal_data is not None
                         for request in requests)
Woosuk Kwon's avatar
Woosuk Kwon committed
435
    if args.backend == "vllm":
436
        if args.async_engine:
437
438
439
440
441
442
443
            elapsed_time = uvloop.run(
                run_vllm_async(
                    requests,
                    args.n,
                    AsyncEngineArgs.from_cli_args(args),
                    args.disable_frontend_multiprocessing,
                ))
444
        else:
445
            elapsed_time = run_vllm(warmup_requests, requests, args.n,
446
                                    EngineArgs.from_cli_args(args))
447
448
    elif args.backend == "hf":
        assert args.tensor_parallel_size == 1
449
        elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
450
                              args.hf_max_batch_size, args.trust_remote_code)
451
452
453
    elif args.backend == "mii":
        elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
                               args.output_len)
454
455
    else:
        raise ValueError(f"Unknown backend: {args.backend}")
456
457
458
    total_num_tokens = sum(request.prompt_len + request.expected_output_len
                           for request in requests)
    total_output_tokens = sum(request.expected_output_len
zhuwenwen's avatar
zhuwenwen committed
459
                            for request in requests)
460
461
462
463
464
    if is_multi_modal:
        print("\033[91mWARNING\033[0m: Multi-modal request detected. The "
              "following metrics are not accurate because image tokens are not"
              " counted. See vllm-project/vllm/issues/9778 for details.")
        # TODO(vllm-project/vllm/issues/9778): Count molti-modal token length.
Woosuk Kwon's avatar
Woosuk Kwon committed
465
    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
466
467
          f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
          f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
468

469
470
471
472
473
474
475
476
477
478
479
480
    # Output JSON results if specified
    if args.output_json:
        results = {
            "elapsed_time": elapsed_time,
            "num_requests": len(requests),
            "total_num_tokens": total_num_tokens,
            "requests_per_second": len(requests) / elapsed_time,
            "tokens_per_second": total_num_tokens / elapsed_time,
        }
        with open(args.output_json, "w") as f:
            json.dump(results, f, indent=4)

481
482

if __name__ == "__main__":
483
    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
484
485
    parser.add_argument("--backend",
                        type=str,
486
                        choices=["vllm", "hf", "mii"],
Woosuk Kwon's avatar
Woosuk Kwon committed
487
                        default="vllm")
488
489
    parser.add_argument("--dataset",
                        type=str,
490
                        default=None,
491
492
493
                        help="Path to the dataset. The dataset is expected to "
                        "be a json in form of List[Dict[..., conversations: "
                        "List[Dict[..., value: <prompt_or_response>]]]]")
494
495
496
497
498
499
500
501
502
    parser.add_argument("--input-len",
                        type=int,
                        default=None,
                        help="Input prompt length for each request")
    parser.add_argument("--output-len",
                        type=int,
                        default=None,
                        help="Output length for each request. Overrides the "
                        "output length from the dataset.")
503
504
505
    parser.add_argument("--n",
                        type=int,
                        default=1,
506
                        help="Number of generated sequences per prompt.")
zhuwenwen's avatar
zhuwenwen committed
507
508
509
510
    parser.add_argument('--num-iters-warmup',
                        type=int,
                        default=1,
                        help='Number of iterations to run for warmup.')
511
512
513
    parser.add_argument("--num-prompts",
                        type=int,
                        default=1000,
514
                        help="Number of prompts to process.")
515
516
517
    parser.add_argument("--hf-max-batch-size",
                        type=int,
                        default=None,
518
                        help="Maximum batch size for HF backend.")
519
520
521
522
523
    parser.add_argument(
        '--output-json',
        type=str,
        default=None,
        help='Path to save the throughput results in JSON format.')
524
525
526
527
528
529
530
531
    parser.add_argument("--async-engine",
                        action='store_true',
                        default=False,
                        help="Use vLLM async engine rather than LLM class.")
    parser.add_argument("--disable-frontend-multiprocessing",
                        action='store_true',
                        default=False,
                        help="Disable decoupled async engine frontend.")
532
533
534
535
536
537
538
539
    # LoRA
    parser.add_argument(
        "--lora-path",
        type=str,
        default=None,
        help="Path to the lora adapters to use. This can be an absolute path, "
        "a relative path, or a Hugging Face model identifier.")

540
    parser = AsyncEngineArgs.add_cli_args(parser)
541
    args = parser.parse_args()
542
543
544
545
546
547
548
    if args.tokenizer is None:
        args.tokenizer = args.model
    if args.dataset is None:
        assert args.input_len is not None
        assert args.output_len is not None
    else:
        assert args.input_len is None
549
550
    if args.enable_lora:
        assert args.lora_path is not None
551

Woosuk Kwon's avatar
Woosuk Kwon committed
552
    if args.backend == "vllm":
553
554
555
556
557
        if args.hf_max_batch_size is not None:
            raise ValueError("HF max batch size is only for HF backend.")
    elif args.backend == "hf":
        if args.hf_max_batch_size is None:
            raise ValueError("HF max batch size is required for HF backend.")
558
559
        if args.quantization is not None:
            raise ValueError("Quantization is only for vLLM backend.")
560
561
562
        if args.enable_lora is not None:
            raise ValueError("LoRA benchmarking is only supported for vLLM"
                             " backend")
563
564
565
566
567
568
569
570
571
572
573
574
    elif args.backend == "mii":
        if args.dtype != "auto":
            raise ValueError("dtype must be auto for MII backend.")
        if args.n != 1:
            raise ValueError("n must be 1 for MII backend.")
        if args.quantization is not None:
            raise ValueError("Quantization is only for vLLM backend.")
        if args.hf_max_batch_size is not None:
            raise ValueError("HF max batch size is only for HF backend.")
        if args.tokenizer != args.model:
            raise ValueError("Tokenizer must be the same as the model for MII "
                             "backend.")
575
576
577
        if args.enable_lora is not None:
            raise ValueError("LoRA benchmarking is only supported for vLLM"
                             " backend")
578
    main(args)