benchmark_throughput.py 28.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Benchmark offline inference throughput."""
4

5
import argparse
6
import dataclasses
7
import json
8
import os
9
10
import random
import time
zhuwenwen's avatar
zhuwenwen committed
11

zhuwenwen's avatar
zhuwenwen committed
12
from pathlib import Path
13
import warnings
14
from typing import Any, Optional, Union
15

zhuwenwen's avatar
zhuwenwen committed
16
import numpy as np
17
import torch
18
import uvloop
19
from tqdm import tqdm
zhuwenwen's avatar
zhuwenwen committed
20
21

from vllm.inputs import PromptType
22
from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
23
from typing_extensions import deprecated
24
25
26
27
28
29
30
31
32
33
34
35
36

from benchmark_dataset import (
    AIMODataset,
    BurstGPTDataset,
    ConversationDataset,
    InstructCoderDataset,
    RandomDataset,
    SampleRequest,
    ShareGPTDataset,
    SonnetDataset,
    VisionArenaDataset,
)
from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
37
from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
38
from vllm.entrypoints.openai.api_server import (
39
40
    build_async_engine_client_from_engine_args,
)
41
from vllm.inputs import TextPrompt, TokensPrompt
42
from vllm.lora.request import LoRARequest
43
from vllm.outputs import RequestOutput
44
from vllm.sampling_params import BeamSearchParams
45
from vllm.utils import FlexibleArgumentParser, merge_async_iterators
46

47

Woosuk Kwon's avatar
Woosuk Kwon committed
48
def run_vllm(
49
    requests: list[SampleRequest],
50
    n: int,
51
    num_iters_warmup: int,
52
    engine_args: EngineArgs,
53
    disable_detokenize: bool = False,
54
) -> tuple[float, Optional[list[RequestOutput]]]:
55
    from vllm import LLM, SamplingParams
56

57
    llm = LLM(**dataclasses.asdict(engine_args))
58
    assert all(
59
60
61
62
63
64
65
        llm.llm_engine.model_config.max_model_len
        >= (request.prompt_len + request.expected_output_len)
        for request in requests
    ), (
        "Please ensure that max_model_len is greater than the sum of"
        " prompt_len and expected_output_len for all requests."
    )
Zhuohan Li's avatar
Zhuohan Li committed
66
    # Add the requests to the engine.
67
    prompts: list[Union[TextPrompt, TokensPrompt]] = []
68
    sampling_params: list[SamplingParams] = []
69
    for request in requests:
70
        prompts.append(
71
72
73
74
75
76
77
78
79
            TokensPrompt(
                prompt_token_ids=request.prompt["prompt_token_ids"],
                multi_modal_data=request.multi_modal_data,
            )
            if "prompt_token_ids" in request.prompt
            else TextPrompt(
                prompt=request.prompt, multi_modal_data=request.multi_modal_data
            )
        )
80
81
82
        sampling_params.append(
            SamplingParams(
                n=n,
83
                temperature=1.0,
84
85
                top_p=1.0,
                ignore_eos=True,
86
                max_tokens=request.expected_output_len,
87
                detokenize=not disable_detokenize,
88
89
            )
        )
90
    lora_requests: Optional[list[LoRARequest]] = None
91
92
    if engine_args.enable_lora:
        lora_requests = [request.lora_request for request in requests]
93

zhuwenwen's avatar
zhuwenwen committed
94
    # warmup
95
    warmup_sampling_params = SamplingParams(
zhuwenwen's avatar
zhuwenwen committed
96
        n=n,
97
98
99
100
101
102
        temperature=1.0,
        top_p=1.0,
        ignore_eos=True,
        max_tokens=10,
    )
    dummy_prompt_token_ids = np.random.randint(10000, size=(1,10))
zhuwenwen's avatar
zhuwenwen committed
103
    dummy_prompts: list[PromptType] = [{
104
105
        "prompt_token_ids": batch
    } for batch in dummy_prompt_token_ids.tolist()]
zhuwenwen's avatar
zhuwenwen committed
106
    
zhuwenwen's avatar
zhuwenwen committed
107
108
    use_beam_search = False
    
109
110
    print("Warming up...")
    for _ in tqdm(range(num_iters_warmup), desc="Warmup iterations"):
zhuwenwen's avatar
zhuwenwen committed
111
        if not use_beam_search:
zhuwenwen's avatar
zhuwenwen committed
112
            llm.generate(dummy_prompts, sampling_params=warmup_sampling_params, use_tqdm=False)
zhuwenwen's avatar
zhuwenwen committed
113
114
115
116
117
118
119
120
121
        else:
            llm.beam_search(
                dummy_prompts,
                BeamSearchParams(
                    beam_width=args.n,
                    max_tokens=args.output_len,
                    ignore_eos=True,
                ),
            )
zhuwenwen's avatar
zhuwenwen committed
122

123
    outputs = None
124
    if not use_beam_search:
zhuwenwen's avatar
zhuwenwen committed
125
126
127
128
129
130
131
132
133
134
135
136
137
138
        if args.profile:
            profile_dir = args.profile_result_dir
            if not profile_dir:
                profile_dir = Path(
                    "."
                ) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
            print(f"Profiling (results will be saved to '{profile_dir}')...")
            with torch.profiler.profile(
                        activities=[torch.profiler.ProfilerActivity.CPU,
                                    torch.profiler.ProfilerActivity.CUDA,
                        ],record_shapes=True,
                        on_trace_ready=torch.profiler.tensorboard_trace_handler(str(profile_dir))
                        ) as prof:
                start = time.perf_counter()
zhuwenwen's avatar
zhuwenwen committed
139
140
141
                outputs = llm.generate(
                    prompts, sampling_params, lora_request=lora_requests, use_tqdm=True
                )
zhuwenwen's avatar
zhuwenwen committed
142
143
144
145
146
                end = time.perf_counter()
            print('Prepare time report')
            print(prof.key_averages(group_by_input_shape=True).table(sort_by="self_cuda_time_total", row_limit=-1))
        else:
            start = time.perf_counter()
zhuwenwen's avatar
zhuwenwen committed
147
148
149
            outputs = llm.generate(
                prompts, sampling_params, lora_request=lora_requests, use_tqdm=True
            )
zhuwenwen's avatar
zhuwenwen committed
150
            end = time.perf_counter()
151
    else:
152
        assert lora_requests is None, "BeamSearch API does not support LoRA"
153
        # output_len should be the same for all requests.
154
        output_len = requests[0].expected_output_len
155
156
        for request in requests:
            assert request.expected_output_len == output_len
157
        start = time.perf_counter()
158
159
160
161
162
163
        llm.beam_search(
            prompts,
            BeamSearchParams(
                beam_width=n,
                max_tokens=output_len,
                ignore_eos=True,
164
165
            ),
        )
166
        end = time.perf_counter()
167
168
169
170
    return end - start, outputs


def run_vllm_chat(
171
172
173
174
175
    requests: list[SampleRequest],
    n: int,
    engine_args: EngineArgs,
    disable_detokenize: bool = False,
) -> tuple[float, list[RequestOutput]]:
176
177
178
179
180
181
    """
    Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
    multimodal models as it properly handles multimodal inputs and chat
    formatting. For non-multimodal models, use run_vllm() instead.
    """
    from vllm import LLM, SamplingParams
182

183
184
185
    llm = LLM(**dataclasses.asdict(engine_args))

    assert all(
186
187
188
189
190
191
192
        llm.llm_engine.model_config.max_model_len
        >= (request.prompt_len + request.expected_output_len)
        for request in requests
    ), (
        "Please ensure that max_model_len is greater than the sum of "
        "prompt_len and expected_output_len for all requests."
    )
193
194
195
196
197
198
199
200
201
202
203
204
205

    prompts = []
    sampling_params: list[SamplingParams] = []
    for request in requests:
        prompts.append(request.prompt)
        sampling_params.append(
            SamplingParams(
                n=n,
                temperature=1.0,
                top_p=1.0,
                ignore_eos=True,
                max_tokens=request.expected_output_len,
                detokenize=not disable_detokenize,
206
207
            )
        )
208
209
210
211
    start = time.perf_counter()
    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
    end = time.perf_counter()
    return end - start, outputs
212
213


214
async def run_vllm_async(
215
    requests: list[SampleRequest],
216
    n: int,
217
    engine_args: AsyncEngineArgs,
218
    disable_frontend_multiprocessing: bool = False,
219
    disable_detokenize: bool = False,
220
221
222
223
) -> float:
    from vllm import SamplingParams

    async with build_async_engine_client_from_engine_args(
224
225
        engine_args,
        disable_frontend_multiprocessing=disable_frontend_multiprocessing,
226
    ) as llm:
227
        model_config = await llm.get_model_config()
228
        assert all(
229
230
231
232
233
234
235
            model_config.max_model_len
            >= (request.prompt_len + request.expected_output_len)
            for request in requests
        ), (
            "Please ensure that max_model_len is greater than the sum of"
            " prompt_len and expected_output_len for all requests."
        )
236
237

        # Add the requests to the engine.
238
        prompts: list[Union[TextPrompt, TokensPrompt]] = []
239
240
        sampling_params: list[SamplingParams] = []
        lora_requests: list[Optional[LoRARequest]] = []
241
        for request in requests:
242
            prompts.append(
243
244
245
246
247
248
249
250
251
                TokensPrompt(
                    prompt_token_ids=request.prompt["prompt_token_ids"],
                    multi_modal_data=request.multi_modal_data,
                )
                if "prompt_token_ids" in request.prompt
                else TextPrompt(
                    prompt=request.prompt, multi_modal_data=request.multi_modal_data
                )
            )
252
253
254
            sampling_params.append(
                SamplingParams(
                    n=n,
255
                    temperature=1.0,
256
257
                    top_p=1.0,
                    ignore_eos=True,
258
                    max_tokens=request.expected_output_len,
259
                    detokenize=not disable_detokenize,
260
261
                )
            )
262
            lora_requests.append(request.lora_request)
263
264
265

        generators = []
        start = time.perf_counter()
266
267
268
269
        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}")
270
271
272
273
274
275
276
277
            generators.append(generator)
        all_gens = merge_async_iterators(*generators)
        async for i, res in all_gens:
            pass
        end = time.perf_counter()
        return end - start


278
def run_hf(
279
    requests: list[SampleRequest],
280
281
282
283
    model: str,
    tokenizer: PreTrainedTokenizerBase,
    n: int,
    max_batch_size: int,
284
    trust_remote_code: bool,
285
    disable_detokenize: bool = False,
286
) -> float:
287
    llm = AutoModelForCausalLM.from_pretrained(
288
289
        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code
    )
290
291
292
    if llm.config.model_type == "llama":
        # To enable padding in the HF backend.
        tokenizer.pad_token = tokenizer.eos_token
293
294
295
    llm = llm.cuda()

    pbar = tqdm(total=len(requests))
296
    start = time.perf_counter()
297
    batch: list[str] = []
298
299
300
    max_prompt_len = 0
    max_output_len = 0
    for i in range(len(requests)):
301
302
303
        prompt = requests[i].prompt
        prompt_len = requests[i].prompt_len
        output_len = requests[i].expected_output_len
304
305
306
307
308
309
        # 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.
310
311
            next_prompt_len = requests[i + 1].prompt_len
            next_output_len = requests[i + 1].expected_output_len
312
313
314
315
            if (
                max(max_prompt_len, next_prompt_len)
                + max(max_output_len, next_output_len)
            ) <= 2048:
316
317
318
319
                # We can add more requests to the batch.
                continue

        # Generate the sequences.
320
        input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids
321
322
        llm_outputs = llm.generate(
            input_ids=input_ids.cuda(),
323
            do_sample=True,
324
325
326
327
328
329
            num_return_sequences=n,
            temperature=1.0,
            top_p=1.0,
            use_cache=True,
            max_new_tokens=max_output_len,
        )
330
331
332
        if not disable_detokenize:
            # Include the decoding time.
            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
333
334
335
336
337
338
        pbar.update(len(batch))

        # Clear the batch.
        batch = []
        max_prompt_len = 0
        max_output_len = 0
339
    end = time.perf_counter()
340
341
342
    return end - start


343
def run_mii(
344
    requests: list[SampleRequest],
345
346
347
348
    model: str,
    tensor_parallel_size: int,
    output_len: int,
) -> float:
349
    from mii import client, serve
350

351
    llm = serve(model, tensor_parallel=tensor_parallel_size)
352
    prompts = [request.prompt for request in requests]
353
354

    start = time.perf_counter()
355
    llm.generate(prompts, max_new_tokens=output_len)
356
    end = time.perf_counter()
357
358
    client = client(model)
    client.terminate_server()
359
360
361
    return end - start


362
363
364
def save_to_pytorch_benchmark_format(
    args: argparse.Namespace, results: dict[str, Any]
) -> None:
365
366
367
368
369
370
371
    pt_records = convert_to_pytorch_benchmark_format(
        args=args,
        metrics={
            "requests_per_second": [results["requests_per_second"]],
            "tokens_per_second": [results["tokens_per_second"]],
        },
        extra_info={
372
373
374
            k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"]
        },
    )
375
376
377
    if pt_records:
        # Don't use json suffix here as we don't want CI to pick it up
        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
378
        write_to_json(pt_file, pt_records)
379
380


381
382
383
384
385
386
387
388
389
390
391
392
393
394
def get_requests(args, tokenizer):
    # Common parameters for all dataset types.
    common_kwargs = {
        "dataset_path": args.dataset_path,
        "random_seed": args.seed,
    }
    sample_kwargs = {
        "tokenizer": tokenizer,
        "lora_path": args.lora_path,
        "max_loras": args.max_loras,
        "num_requests": args.num_prompts,
        "input_len": args.input_len,
        "output_len": args.output_len,
    }
395

396
397
398
399
400
401
    if args.dataset_path is None or args.dataset_name == "random":
        sample_kwargs["range_ratio"] = args.random_range_ratio
        sample_kwargs["prefix_len"] = args.prefix_len
        dataset_cls = RandomDataset
    elif args.dataset_name == "sharegpt":
        dataset_cls = ShareGPTDataset
402
403
        if args.backend == "vllm-chat":
            sample_kwargs["enable_multimodal_chat"] = True
404
405
    elif args.dataset_name == "sonnet":
        assert tokenizer.chat_template or tokenizer.default_chat_template, (
406
407
            "Tokenizer/model must have chat template for sonnet dataset."
        )
408
409
410
411
412
        dataset_cls = SonnetDataset
        sample_kwargs["prefix_len"] = args.prefix_len
        sample_kwargs["return_prompt_formatted"] = True
    elif args.dataset_name == "burstgpt":
        dataset_cls = BurstGPTDataset
413
    elif args.dataset_name == "hf":
414
        common_kwargs["no_stream"] = args.no_stream
415
416
        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
            dataset_cls = VisionArenaDataset
417
418
            common_kwargs["dataset_subset"] = None
            common_kwargs["dataset_split"] = "train"
419
            sample_kwargs["enable_multimodal_chat"] = True
420
        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
421
            dataset_cls = InstructCoderDataset
422
            common_kwargs["dataset_split"] = "train"
423
424
        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
            dataset_cls = ConversationDataset
425
426
            common_kwargs["dataset_subset"] = args.hf_subset
            common_kwargs["dataset_split"] = args.hf_split
427
            sample_kwargs["enable_multimodal_chat"] = True
428
429
        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
            dataset_cls = AIMODataset
430
431
            common_kwargs["dataset_subset"] = None
            common_kwargs["dataset_split"] = "train"
432
433
434
435
436
    else:
        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
    # Remove None values
    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
437
438


439
440
441
442
@deprecated(
    "benchmark_throughput.py is deprecated and will be removed in a "
    "future version. Please use 'vllm bench throughput' instead.",
)
443
def main(args: argparse.Namespace):
444
445
    if args.seed is None:
        args.seed = 0
446
447
448
    print(args)
    random.seed(args.seed)
    # Sample the requests.
449
    tokenizer = AutoTokenizer.from_pretrained(
450
451
        args.tokenizer, trust_remote_code=args.trust_remote_code
    )
452
    requests = get_requests(args, tokenizer)
453
    is_multi_modal = any(request.multi_modal_data is not None for request in requests)
454
    request_outputs: Optional[list[RequestOutput]] = None
Woosuk Kwon's avatar
Woosuk Kwon committed
455
    if args.backend == "vllm":
456
        if args.async_engine:
457
458
459
460
461
462
            elapsed_time = uvloop.run(
                run_vllm_async(
                    requests,
                    args.n,
                    AsyncEngineArgs.from_cli_args(args),
                    args.disable_frontend_multiprocessing,
463
                    args.disable_detokenize,
464
465
                )
            )
466
        else:
467
            elapsed_time, request_outputs = run_vllm(
468
469
                requests,
                args.n,
zhuwenwen's avatar
zhuwenwen committed
470
                args.num_iters_warmup,
471
472
473
                EngineArgs.from_cli_args(args),
                args.disable_detokenize,
            )
474
475
    elif args.backend == "hf":
        assert args.tensor_parallel_size == 1
476
477
478
479
480
481
482
483
484
        elapsed_time = run_hf(
            requests,
            args.model,
            tokenizer,
            args.n,
            args.hf_max_batch_size,
            args.trust_remote_code,
            args.disable_detokenize,
        )
485
    elif args.backend == "mii":
486
487
488
        elapsed_time = run_mii(
            requests, args.model, args.tensor_parallel_size, args.output_len
        )
489
490
    elif args.backend == "vllm-chat":
        elapsed_time, request_outputs = run_vllm_chat(
491
492
            requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize
        )
493
494
    else:
        raise ValueError(f"Unknown backend: {args.backend}")
495
496
497
498
499
500
501
502
503

    if request_outputs:
        # Note: with the vllm and vllm-chat backends,
        # we have request_outputs, which we use to count tokens.
        total_prompt_tokens = 0
        total_output_tokens = 0
        for ro in request_outputs:
            if not isinstance(ro, RequestOutput):
                continue
504
505
506
507
            total_prompt_tokens += (
                len(ro.prompt_token_ids) if ro.prompt_token_ids else 0
            )
            total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o)
508
509
        total_num_tokens = total_prompt_tokens + total_output_tokens
    else:
510
        total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests)
511
512
513
514
        total_output_tokens = sum(r.expected_output_len for r in requests)
        total_prompt_tokens = total_num_tokens - total_output_tokens

    if is_multi_modal and args.backend != "vllm-chat":
515
516
517
518
519
520
        print(
            "\033[91mWARNING\033[0m: Multi-modal request with "
            f"{args.backend} backend detected. The "
            "following metrics are not accurate because image tokens are not"
            " counted. See vllm-project/vllm/issues/9778 for details."
        )
521
        # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
522
523
        # vllm-chat backend counts the image tokens now

zhuwenwen's avatar
zhuwenwen committed
524
    print(f"Latency: {elapsed_time:.2f} s")
525
526
527
528
529
    print(
        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
        f"{total_output_tokens / elapsed_time:.2f} output tokens/s"
    )
530
531
    print(f"Total num prompt tokens:  {total_prompt_tokens}")
    print(f"Total num output tokens:  {total_output_tokens}")
532

533
534
535
536
537
538
539
540
541
542
543
    # 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)
544
        save_to_pytorch_benchmark_format(args, results)
545

546

547
548
549
550
551
552
553
554
555
556
def validate_args(args):
    """
    Validate command-line arguments.
    """

    # === Deprecation and Defaulting ===
    if args.dataset is not None:
        warnings.warn(
            "The '--dataset' argument will be deprecated in the next release. "
            "Please use '--dataset-name' and '--dataset-path' instead.",
557
558
            stacklevel=2,
        )
559
560
561
562
563
564
565
566
567
568
569
570
        args.dataset_path = args.dataset

    if not getattr(args, "tokenizer", None):
        args.tokenizer = args.model

    # === Backend Validation ===
    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
    if args.backend not in valid_backends:
        raise ValueError(f"Unsupported backend: {args.backend}")

    # === Dataset Configuration ===
    if not args.dataset and not args.dataset_path:
571
572
        print("When dataset path is not set, it will default to random dataset")
        args.dataset_name = "random"
573
574
575
576
577
578
579
        if args.input_len is None:
            raise ValueError("input_len must be provided for a random dataset")

    # === Dataset Name Specific Checks ===
    # --hf-subset and --hf-split: only used
    # when dataset_name is 'hf'
    if args.dataset_name != "hf" and (
580
581
582
583
584
        getattr(args, "hf_subset", None) is not None
        or getattr(args, "hf_split", None) is not None
    ):
        warnings.warn(
            "--hf-subset and --hf-split will be ignored \
585
                since --dataset-name is not 'hf'.",
586
587
            stacklevel=2,
        )
588
    elif args.dataset_name == "hf":
589
        if args.dataset_path in (
590
591
592
593
594
595
596
597
598
599
600
601
602
            VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
            | ConversationDataset.SUPPORTED_DATASET_PATHS
        ):
            assert args.backend == "vllm-chat", (
                f"{args.dataset_path} needs to use vllm-chat as the backend."
            )  # noqa: E501
        elif args.dataset_path in (
            InstructCoderDataset.SUPPORTED_DATASET_PATHS
            | AIMODataset.SUPPORTED_DATASET_PATHS
        ):
            assert args.backend == "vllm", (
                f"{args.dataset_path} needs to use vllm as the backend."
            )  # noqa: E501
603
        else:
604
            raise ValueError(f"{args.dataset_path} is not supported by hf dataset.")
605
606

    # --random-range-ratio: only used when dataset_name is 'random'
607
608
609
    if args.dataset_name != "random" and args.random_range_ratio is not None:
        warnings.warn(
            "--random-range-ratio will be ignored since \
610
                --dataset-name is not 'random'.",
611
612
            stacklevel=2,
        )
613
614
615

    # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
    # set.
616
617
618
619
620
621
    if (
        args.dataset_name not in {"random", "sonnet", None}
        and args.prefix_len is not None
    ):
        warnings.warn(
            "--prefix-len will be ignored since --dataset-name\
622
                 is not 'random', 'sonnet', or not set.",
623
624
            stacklevel=2,
        )
625
626
627

    # === LoRA Settings ===
    if getattr(args, "enable_lora", False) and args.backend != "vllm":
628
        raise ValueError("LoRA benchmarking is only supported for vLLM backend")
629
630
631
632
633
634
635
636
637
    if getattr(args, "enable_lora", False) and args.lora_path is None:
        raise ValueError("LoRA path must be provided when enable_lora is True")

    # === Backend-specific Validations ===
    if args.backend == "hf" and args.hf_max_batch_size is None:
        raise ValueError("HF max batch size is required for HF backend")
    if args.backend != "hf" and args.hf_max_batch_size is not None:
        raise ValueError("HF max batch size is only for HF backend.")

638
639
640
641
    if (
        args.backend in {"hf", "mii"}
        and getattr(args, "quantization", None) is not None
    ):
642
643
644
645
646
647
648
        raise ValueError("Quantization is only for vLLM backend.")

    if args.backend == "mii" and args.dtype != "auto":
        raise ValueError("dtype must be auto for MII backend.")
    if args.backend == "mii" and args.n != 1:
        raise ValueError("n must be 1 for MII backend.")
    if args.backend == "mii" and args.tokenizer != args.model:
649
        raise ValueError("Tokenizer must be the same as the model for MII backend.")
650

651
652
653
654
    # --data-parallel is not supported currently.
    # https://github.com/vllm-project/vllm/issues/16222
    if args.data_parallel_size > 1:
        raise ValueError(
655
656
            "Data parallel is not supported in offline benchmark, "
            "please use benchmark serving instead"
657
        )
658

659

660
def create_argument_parser():
661
    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
662
663
664
665
666
667
    parser.add_argument(
        "--backend",
        type=str,
        choices=["vllm", "hf", "mii", "vllm-chat"],
        default="vllm",
    )
668
669
670
671
672
    parser.add_argument(
        "--dataset-name",
        type=str,
        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
        help="Name of the dataset to benchmark on.",
673
674
        default="sharegpt",
    )
675
676
677
678
679
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
680
681
682
683
684
685
686
    parser.add_argument(
        "--dataset",
        type=str,
        default=None,
        help="Path to the ShareGPT dataset, will be deprecated in\
            the next release. The dataset is expected to "
        "be a json in form of list[dict[..., conversations: "
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
        "list[dict[..., value: <prompt_or_response>]]]]",
    )
    parser.add_argument(
        "--dataset-path", type=str, default=None, help="Path to the dataset"
    )
    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.",
    )
    parser.add_argument(
        "--n", type=int, default=1, help="Number of generated sequences per prompt."
    )
zhuwenwen's avatar
zhuwenwen committed
708
709
710
    parser.add_argument(
        "--num-iters-warmup", type=int, default=1, help="Number of iterations to run for warmup."
    )
711
    parser.add_argument(
712
713
        "--num-prompts", type=int, default=1000, help="Number of prompts to process."
    )
zhuwenwen's avatar
zhuwenwen committed
714
715
716
717
718
719
720
721
722
723
    parser.add_argument(
        '--profile',
        action='store_true',
        help='profile the generation process of a single batch')
    parser.add_argument(
        '--profile-result-dir',
        type=str,
        default=None,
        help=('path to save the pytorch profiler output. Can be visualized '
              'with ui.perfetto.dev or Tensorboard.'))
724
    parser.add_argument(
725
726
727
728
729
730
731
        "--hf-max-batch-size",
        type=int,
        default=None,
        help="Maximum batch size for HF backend.",
    )
    parser.add_argument(
        "--output-json",
732
733
        type=str,
        default=None,
734
735
736
737
738
739
740
741
742
743
744
745
746
747
        help="Path to save the throughput results in JSON format.",
    )
    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.",
    )
748
749
750
    parser.add_argument(
        "--disable-detokenize",
        action="store_true",
751
752
753
754
755
        help=(
            "Do not detokenize the response (i.e. do not include "
            "detokenization time in the measurement)"
        ),
    )
756
757
758
759
760
    # LoRA
    parser.add_argument(
        "--lora-path",
        type=str,
        default=None,
761
        help="Path to the LoRA adapters to use. This can be an absolute path, "
762
763
        "a relative path, or a Hugging Face model identifier.",
    )
764
765
766
    parser.add_argument(
        "--prefix-len",
        type=int,
767
768
769
770
771
772
773
774
775
776
        default=None,
        help=f"Number of prefix tokens to be used in RandomDataset "
        "and SonnetDataset. For RandomDataset, the total input "
        "length is the sum of prefix-len (default: "
        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
        "sampled from [input_len * (1 - range_ratio), "
        "input_len * (1 + range_ratio)]. For SonnetDataset, "
        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
        "controls how much of the input is fixed lines versus "
        "random lines, but the total input length remains approximately "
777
778
        "input_len tokens.",
    )
779
780
781
782
    # random dataset
    parser.add_argument(
        "--random-range-ratio",
        type=float,
783
        default=None,
784
785
786
787
        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
        "for sampling input/output length, "
        "used only for RandomDataset. Must be in the range [0, 1) to "
        "define a symmetric sampling range "
788
        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
789
    )
790

791
    # hf dtaset
792
793
794
795
796
797
    parser.add_argument(
        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
    )
    parser.add_argument(
        "--hf-split", type=str, default=None, help="Split of the HF dataset."
    )
798

799
    parser = AsyncEngineArgs.add_cli_args(parser)
800
801
802
803
804
805

    return parser


if __name__ == "__main__":
    parser = create_argument_parser()
806
    args = parser.parse_args()
807
808
    if args.tokenizer is None:
        args.tokenizer = args.model
809
    validate_args(args)
zhuwenwen's avatar
zhuwenwen committed
810
    main(args)