benchmark_throughput.py 21.7 KB
Newer Older
1
"""Benchmark offline inference throughput."""
2
3
4
5
import argparse
import json
import random
import time
6
from typing import List, Optional, Tuple
7

8
import torch
9
import uvloop
10
from tqdm import tqdm
11
12
from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          PreTrainedTokenizerBase)
13

14
15
16
from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
from vllm.entrypoints.openai.api_server import (
    build_async_engine_client_from_engine_args)
17
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
18
from vllm.utils import FlexibleArgumentParser, merge_async_iterators
19

20
21
22
23
24

def sample_requests(
    dataset_path: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
25
    fixed_output_len: Optional[int],
26
) -> List[Tuple[str, int, int]]:
27
28
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")
29

30
31
32
33
    # Load the dataset.
    with open(dataset_path) as f:
        dataset = json.load(f)
    # Filter out the conversations with less than 2 turns.
34
    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
35
    # Only keep the first two turns of each conversation.
36
37
    dataset = [(data["conversations"][0]["value"],
                data["conversations"][1]["value"]) for data in dataset]
38

39
40
    # Shuffle the dataset.
    random.shuffle(dataset)
41

42
    # Filter out sequences that are too long or too short
43
    filtered_dataset: List[Tuple[str, int, int]] = []
44
45
46
47
48
49
50
51
52
    for i in range(len(dataset)):
        if len(filtered_dataset) == num_requests:
            break

        # Tokenize the prompts and completions.
        prompt = dataset[i][0]
        prompt_token_ids = tokenizer(prompt).input_ids
        completion = dataset[i][1]
        completion_token_ids = tokenizer(completion).input_ids
53
        prompt_len = len(prompt_token_ids)
54
55
        output_len = len(completion_token_ids
                         ) if fixed_output_len is None else fixed_output_len
56
57
58
59
60
61
62
        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
        filtered_dataset.append((prompt, prompt_len, output_len))
63

64
    return filtered_dataset
65
66


Woosuk Kwon's avatar
Woosuk Kwon committed
67
def run_vllm(
68
69
    requests: List[Tuple[str, int, int]],
    model: str,
70
    tokenizer: str,
71
    quantization: Optional[str],
72
73
74
75
    tensor_parallel_size: int,
    seed: int,
    n: int,
    use_beam_search: bool,
76
    trust_remote_code: bool,
77
    dtype: str,
78
79
    max_model_len: Optional[int],
    enforce_eager: bool,
80
    kv_cache_dtype: str,
81
    quantization_param_path: Optional[str],
82
    device: str,
83
    enable_prefix_caching: bool,
84
85
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
86
    distributed_executor_backend: Optional[str],
87
    gpu_memory_utilization: float = 0.9,
88
89
    num_scheduler_steps: int = 1,
    use_v2_block_manager: bool = False,
90
    download_dir: Optional[str] = None,
91
    load_format: str = EngineArgs.load_format,
92
    disable_async_output_proc: bool = False,
93
) -> float:
94
    from vllm import LLM, SamplingParams
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
    llm = LLM(
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
113
        distributed_executor_backend=distributed_executor_backend,
114
        load_format=load_format,
115
116
        num_scheduler_steps=num_scheduler_steps,
        use_v2_block_manager=use_v2_block_manager,
117
        disable_async_output_proc=disable_async_output_proc,
118
    )
119

Zhuohan Li's avatar
Zhuohan Li committed
120
    # Add the requests to the engine.
121
122
    prompts: List[str] = []
    sampling_params: List[SamplingParams] = []
123
    for prompt, _, output_len in requests:
124
125
126
127
128
129
130
131
132
133
        prompts.append(prompt)
        sampling_params.append(
            SamplingParams(
                n=n,
                temperature=0.0 if use_beam_search else 1.0,
                top_p=1.0,
                use_beam_search=use_beam_search,
                ignore_eos=True,
                max_tokens=output_len,
            ))
134

135
    start = time.perf_counter()
136
    llm.generate(prompts, sampling_params, use_tqdm=True)
137
    end = time.perf_counter()
138
139
140
    return end - start


141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
async def run_vllm_async(
    requests: List[Tuple[str, int, int]],
    model: str,
    tokenizer: str,
    quantization: Optional[str],
    tensor_parallel_size: int,
    seed: int,
    n: int,
    use_beam_search: bool,
    trust_remote_code: bool,
    dtype: str,
    max_model_len: Optional[int],
    enforce_eager: bool,
    kv_cache_dtype: str,
    quantization_param_path: Optional[str],
    device: str,
    enable_prefix_caching: bool,
    enable_chunked_prefill: bool,
    max_num_batched_tokens: int,
    distributed_executor_backend: Optional[str],
    gpu_memory_utilization: float = 0.9,
    num_scheduler_steps: int = 1,
    use_v2_block_manager: bool = False,
    download_dir: Optional[str] = None,
    load_format: str = EngineArgs.load_format,
    disable_async_output_proc: bool = False,
    disable_frontend_multiprocessing: bool = False,
) -> float:
    from vllm import SamplingParams
    engine_args = AsyncEngineArgs(
        model=model,
        tokenizer=tokenizer,
        quantization=quantization,
        tensor_parallel_size=tensor_parallel_size,
        seed=seed,
        trust_remote_code=trust_remote_code,
        dtype=dtype,
        max_model_len=max_model_len,
        gpu_memory_utilization=gpu_memory_utilization,
        enforce_eager=enforce_eager,
        kv_cache_dtype=kv_cache_dtype,
        quantization_param_path=quantization_param_path,
        device=device,
        enable_prefix_caching=enable_prefix_caching,
        download_dir=download_dir,
        enable_chunked_prefill=enable_chunked_prefill,
        max_num_batched_tokens=max_num_batched_tokens,
        distributed_executor_backend=distributed_executor_backend,
        load_format=load_format,
        num_scheduler_steps=num_scheduler_steps,
        use_v2_block_manager=use_v2_block_manager,
        disable_async_output_proc=disable_async_output_proc,
        worker_use_ray=False,
        engine_use_ray=False,
        disable_log_requests=True,
    )

    async with build_async_engine_client_from_engine_args(
            engine_args, disable_frontend_multiprocessing) as llm:

        # Add the requests to the engine.
        prompts: List[str] = []
        sampling_params: List[SamplingParams] = []
        for prompt, _, output_len in requests:
            prompts.append(prompt)
            sampling_params.append(
                SamplingParams(
                    n=n,
                    temperature=0.0 if use_beam_search else 1.0,
                    top_p=1.0,
                    use_beam_search=use_beam_search,
                    ignore_eos=True,
                    max_tokens=output_len,
                ))

        generators = []
        start = time.perf_counter()
        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
            generator = llm.generate(prompt, sp, request_id=f"test{i}")
            generators.append(generator)
        all_gens = merge_async_iterators(*generators)
        async for i, res in all_gens:
            pass
        end = time.perf_counter()
        return end - start


228
229
230
231
232
233
234
def run_hf(
    requests: List[Tuple[str, int, int]],
    model: str,
    tokenizer: PreTrainedTokenizerBase,
    n: int,
    use_beam_search: bool,
    max_batch_size: int,
235
    trust_remote_code: bool,
236
237
) -> float:
    assert not use_beam_search
238
239
    llm = AutoModelForCausalLM.from_pretrained(
        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
240
241
242
    if llm.config.model_type == "llama":
        # To enable padding in the HF backend.
        tokenizer.pad_token = tokenizer.eos_token
243
244
245
    llm = llm.cuda()

    pbar = tqdm(total=len(requests))
246
    start = time.perf_counter()
247
248
249
250
251
252
253
254
255
256
257
258
    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]
259
260
            if (max(max_prompt_len, next_prompt_len) +
                    max(max_output_len, next_output_len)) <= 2048:
261
262
263
264
                # We can add more requests to the batch.
                continue

        # Generate the sequences.
265
266
        input_ids = tokenizer(batch, return_tensors="pt",
                              padding=True).input_ids
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
        llm_outputs = llm.generate(
            input_ids=input_ids.cuda(),
            do_sample=not use_beam_search,
            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
284
    end = time.perf_counter()
285
286
287
    return end - start


288
289
290
291
292
293
def run_mii(
    requests: List[Tuple[str, int, int]],
    model: str,
    tensor_parallel_size: int,
    output_len: int,
) -> float:
294
295
    from mii import client, serve
    llm = serve(model, tensor_parallel=tensor_parallel_size)
296
297
298
    prompts = [prompt for prompt, _, _ in requests]

    start = time.perf_counter()
299
    llm.generate(prompts, max_new_tokens=output_len)
300
    end = time.perf_counter()
301
302
    client = client(model)
    client.terminate_server()
303
304
305
    return end - start


306
307
308
309
310
def main(args: argparse.Namespace):
    print(args)
    random.seed(args.seed)

    # Sample the requests.
311
312
313
314
315
316
317
318
319
320
    tokenizer = AutoTokenizer.from_pretrained(
        args.tokenizer, trust_remote_code=args.trust_remote_code)
    if args.dataset is None:
        # Synthesize a prompt with the given input length.
        prompt = "hi" * (args.input_len - 1)
        requests = [(prompt, args.input_len, args.output_len)
                    for _ in range(args.num_prompts)]
    else:
        requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
                                   args.output_len)
321

Woosuk Kwon's avatar
Woosuk Kwon committed
322
    if args.backend == "vllm":
323
        run_args = [
324
325
326
327
328
329
            requests, args.model, args.tokenizer, args.quantization,
            args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
            args.trust_remote_code, args.dtype, args.max_model_len,
            args.enforce_eager, args.kv_cache_dtype,
            args.quantization_param_path, args.device,
            args.enable_prefix_caching, args.enable_chunked_prefill,
330
            args.max_num_batched_tokens, args.distributed_executor_backend,
331
            args.gpu_memory_utilization, args.num_scheduler_steps,
332
            args.use_v2_block_manager, args.download_dir, args.load_format,
333
334
335
336
337
338
339
340
            args.disable_async_output_proc
        ]

        if args.async_engine:
            run_args.append(args.disable_frontend_multiprocessing)
            elapsed_time = uvloop.run(run_vllm_async(*run_args))
        else:
            elapsed_time = run_vllm(*run_args)
341
342
    elif args.backend == "hf":
        assert args.tensor_parallel_size == 1
343
344
345
        elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
                              args.use_beam_search, args.hf_max_batch_size,
                              args.trust_remote_code)
346
347
348
    elif args.backend == "mii":
        elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
                               args.output_len)
349
350
    else:
        raise ValueError(f"Unknown backend: {args.backend}")
351
352
    total_num_tokens = sum(prompt_len + output_len
                           for _, prompt_len, output_len in requests)
Woosuk Kwon's avatar
Woosuk Kwon committed
353
354
    print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
          f"{total_num_tokens / elapsed_time:.2f} tokens/s")
355

356
357
358
359
360
361
362
363
364
365
366
367
    # 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)

368
369

if __name__ == "__main__":
370
    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
371
372
    parser.add_argument("--backend",
                        type=str,
373
                        choices=["vllm", "hf", "mii"],
Woosuk Kwon's avatar
Woosuk Kwon committed
374
                        default="vllm")
375
376
    parser.add_argument("--dataset",
                        type=str,
377
                        default=None,
378
                        help="Path to the dataset.")
379
380
381
382
383
384
385
386
387
    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.")
388
    parser.add_argument("--model", type=str, default="facebook/opt-125m")
389
    parser.add_argument("--tokenizer", type=str, default=None)
390
391
    parser.add_argument('--quantization',
                        '-q',
392
                        choices=[*QUANTIZATION_METHODS, None],
393
                        default=None)
394
    parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
395
396
397
    parser.add_argument("--n",
                        type=int,
                        default=1,
398
399
                        help="Number of generated sequences per prompt.")
    parser.add_argument("--use-beam-search", action="store_true")
400
401
402
    parser.add_argument("--num-prompts",
                        type=int,
                        default=1000,
403
404
                        help="Number of prompts to process.")
    parser.add_argument("--seed", type=int, default=0)
405
406
407
    parser.add_argument("--hf-max-batch-size",
                        type=int,
                        default=None,
408
                        help="Maximum batch size for HF backend.")
409
410
411
    parser.add_argument('--trust-remote-code',
                        action='store_true',
                        help='trust remote code from huggingface')
412
413
414
415
416
417
    parser.add_argument(
        '--max-model-len',
        type=int,
        default=None,
        help='Maximum length of a sequence (including prompt and output). '
        'If None, will be derived from the model.')
418
419
420
421
422
423
424
425
426
    parser.add_argument(
        '--dtype',
        type=str,
        default='auto',
        choices=['auto', 'half', 'float16', 'bfloat16', 'float', 'float32'],
        help='data type for model weights and activations. '
        'The "auto" option will use FP16 precision '
        'for FP32 and FP16 models, and BF16 precision '
        'for BF16 models.')
427
428
429
430
431
432
    parser.add_argument('--gpu-memory-utilization',
                        type=float,
                        default=0.9,
                        help='the fraction of GPU memory to be used for '
                        'the model executor, which can range from 0 to 1.'
                        'If unspecified, will use the default value of 0.9.')
433
434
435
    parser.add_argument("--enforce-eager",
                        action="store_true",
                        help="enforce eager execution")
436
    parser.add_argument(
437
        '--kv-cache-dtype',
438
        type=str,
439
        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
440
        default="auto",
441
442
443
        help='Data type for kv cache storage. If "auto", will use model '
        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
        'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
444
445
446
447
448
449
450
451
452
453
    parser.add_argument(
        '--quantization-param-path',
        type=str,
        default=None,
        help='Path to the JSON file containing the KV cache scaling factors. '
        'This should generally be supplied, when KV cache dtype is FP8. '
        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
        'instead supported for common inference criteria.')
454
455
456
    parser.add_argument(
        "--device",
        type=str,
457
458
459
460
        default="auto",
        choices=["auto", "cuda", "cpu", "openvino", "tpu", "xpu"],
        help='device type for vLLM execution, supporting CUDA, OpenVINO and '
        'CPU.')
461
462
463
464
465
466
467
468
    parser.add_argument(
        "--num-scheduler-steps",
        type=int,
        default=1,
        help="Maximum number of forward steps per scheduler call.")
    parser.add_argument("--use-v2-block-manager",
                        action='store_true',
                        help="Enable block manager v2.")
469
470
471
    parser.add_argument(
        "--enable-prefix-caching",
        action='store_true',
472
        help="Enable automatic prefix caching for vLLM backend.")
473
474
475
476
477
478
479
480
    parser.add_argument("--enable-chunked-prefill",
                        action='store_true',
                        help="enable chunked prefill for vLLM backend.")
    parser.add_argument('--max-num-batched-tokens',
                        type=int,
                        default=None,
                        help='maximum number of batched tokens per '
                        'iteration')
481
482
483
484
485
    parser.add_argument('--download-dir',
                        type=str,
                        default=None,
                        help='directory to download and load the weights, '
                        'default to the default cache dir of huggingface')
486
487
488
489
490
    parser.add_argument(
        '--output-json',
        type=str,
        default=None,
        help='Path to save the throughput results in JSON format.')
491
492
493
494
495
496
497
    parser.add_argument(
        '--distributed-executor-backend',
        choices=['ray', 'mp'],
        default=None,
        help='Backend to use for distributed serving. When more than 1 GPU '
        'is used, will be automatically set to "ray" if installed '
        'or "mp" (multiprocessing) otherwise.')
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
    parser.add_argument(
        '--load-format',
        type=str,
        default=EngineArgs.load_format,
        choices=[
            'auto', 'pt', 'safetensors', 'npcache', 'dummy', 'tensorizer',
            'bitsandbytes'
        ],
        help='The format of the model weights to load.\n\n'
        '* "auto" will try to load the weights in the safetensors format '
        'and fall back to the pytorch bin format if safetensors format '
        'is not available.\n'
        '* "pt" will load the weights in the pytorch bin format.\n'
        '* "safetensors" will load the weights in the safetensors format.\n'
        '* "npcache" will load the weights in pytorch format and store '
        'a numpy cache to speed up the loading.\n'
        '* "dummy" will initialize the weights with random values, '
        'which is mainly for profiling.\n'
        '* "tensorizer" will load the weights using tensorizer from '
        'CoreWeave. See the Tensorize vLLM Model script in the Examples'
        'section for more information.\n'
        '* "bitsandbytes" will load the weights using bitsandbytes '
        'quantization.\n')
521
522
523
524
525
    parser.add_argument(
        "--disable-async-output-proc",
        action='store_true',
        default=False,
        help="Disable async output processor for vLLM backend.")
526
527
528
529
530
531
532
533
    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.")
534
    args = parser.parse_args()
535
536
537
538
539
540
541
    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
542

Woosuk Kwon's avatar
Woosuk Kwon committed
543
    if args.backend == "vllm":
544
545
546
547
548
        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.")
549
550
        if args.quantization is not None:
            raise ValueError("Quantization is only for vLLM backend.")
551
552
553
554
555
556
557
558
559
560
561
562
563
564
    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.use_beam_search:
            raise ValueError("Beam search is not supported 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.")
565
    main(args)