datasets.py 115 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
7
8
9
10
11
12
13
"""
This module defines a framework for sampling benchmark requests from various
datasets. Each dataset subclass of BenchmarkDataset must implement sample
generation. Supported dataset types include:
  - ShareGPT
  - Random (synthetic)
  - Sonnet
  - BurstGPT
  - HuggingFace
  - VisionArena
"""
14

15
import argparse
16
import ast
17
18
19
20
import base64
import io
import json
import logging
21
import math
22
23
import random
from abc import ABC, abstractmethod
24
from collections.abc import Callable, Iterator, Mapping
25
from contextlib import suppress
26
from copy import deepcopy
27
28
29
from dataclasses import dataclass
from functools import cache
from io import BytesIO
30
from tempfile import NamedTemporaryFile
31
from typing import Any, cast
32
33
34
35

import numpy as np
from PIL import Image
from transformers import PreTrainedTokenizerBase
36
from typing_extensions import deprecated
37
38
39
40

from vllm.lora.request import LoRARequest
from vllm.lora.utils import get_adapter_absolute_path
from vllm.multimodal import MultiModalDataDict
41
from vllm.multimodal.image import convert_image_mode
42
from vllm.transformers_utils.tokenizer import AnyTokenizer
43
from vllm.utils.import_utils import PlaceholderModule
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

try:
    from datasets import load_dataset
except ImportError:
    datasets = PlaceholderModule("datasets")
    load_dataset = datasets.placeholder_attr("load_dataset")

try:
    import pandas as pd
except ImportError:
    pd = PlaceholderModule("pandas")

try:
    import librosa
except ImportError:
    librosa = PlaceholderModule("librosa")
60

61
try:
62
    from vllm.utils.argparse_utils import FlexibleArgumentParser
63
64
65
except ImportError:
    from argparse import ArgumentParser as FlexibleArgumentParser

66
67
68
69
70
71
72
73
74
75
76
77
78
logger = logging.getLogger(__name__)

# -----------------------------------------------------------------------------
# Data Classes
# -----------------------------------------------------------------------------


@dataclass
class SampleRequest:
    """
    Represents a single inference request for benchmarking.
    """

79
    prompt: str | list[str]
80
81
    prompt_len: int
    expected_output_len: int
82
83
84
    multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None
    lora_request: LoRARequest | None = None
    request_id: str | None = None
85
86
87
88
89
90
91
92
93


# -----------------------------------------------------------------------------
# Benchmark Dataset Base Class
# -----------------------------------------------------------------------------


class BenchmarkDataset(ABC):
    DEFAULT_SEED = 0
94
    IS_MULTIMODAL = False
95
96
97

    def __init__(
        self,
98
        dataset_path: str | None = None,
99
        random_seed: int = DEFAULT_SEED,
100
101
        disable_shuffle: bool = False,
        **kwargs,
102
103
104
    ) -> None:
        """
        Initialize the BenchmarkDataset with an optional dataset path and random
105
106
        seed.

107
108
        Args:
            dataset_path (Optional[str]): Path to the dataset. If None, it
109
                indicates that a default or random dataset might be used.
110
            random_seed (int): Seed value for reproducible shuffling or
111
                sampling. Defaults to DEFAULT_SEED.
112
113
114
115
        """
        self.dataset_path = dataset_path
        # Set the random seed, ensuring that a None value is replaced with the
        # default seed.
116
        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
117
        self.disable_shuffle = disable_shuffle
118
119
120
        self.data = None

    def apply_multimodal_chat_transformation(
121
122
        self,
        prompt: str,
123
        mm_content: MultiModalDataDict | dict | list[dict] | None = None,
124
    ) -> list[dict]:
125
126
127
128
129
130
131
        """
        Transform a prompt and optional multimodal content into a chat format.
        This method is used for chat models that expect a specific conversation
        format.
        """
        content = [{"text": prompt, "type": "text"}]
        if mm_content is not None:
132
133
134
135
136
            if isinstance(mm_content, list):
                content.extend(cast(list[dict[str, Any]], mm_content))
            elif isinstance(mm_content, dict):
                content.append(mm_content)
            else:
137
                raise TypeError(
138
139
                    "Could not process multimodal content of type: "
                    + f"{type(mm_content)}"
140
                )
141
142
143
144
145
146
147
148
149
150
151
152
153
        return [{"role": "user", "content": content}]

    def load_data(self) -> None:
        """
        Load data from the dataset path into self.data.

        This method must be overridden by subclasses since the method to load
        data will vary depending on the dataset format and source.

        Raises:
            NotImplementedError: If a subclass does not implement this method.
        """
        # TODO (jenniferzhao): add support for downloading data
154
        raise NotImplementedError("load_data must be implemented in subclasses.")
155
156
157

    def get_random_lora_request(
        self,
158
159
160
        max_loras: int | None = None,
        lora_path: str | None = None,
    ) -> LoRARequest | None:
161
        """
162
        Optionally select a random LoRA request.
163
164

        This method is used when LoRA parameters are provided.  It randomly
165
        selects a LoRA based on max_loras.
166
167

        Args:
168
169
170
171
            max_loras (Optional[int]): The maximum number of LoRAs available.
                If `None`, LoRA is not used.
            lora_path (Optional[str]): Path to the LoRA parameters on disk.
                If `None`, LoRA is not used.
172
173

        Returns:
174
175
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
176
177
        """
        if max_loras is None or lora_path is None:
178
            return None
179
180
181
182
183
184
185
186

        # Generate a random LoRA ID in the range [1, max_loras].
        lora_id = random.randint(1, max_loras)
        lora_request = LoRARequest(
            lora_name=str(lora_id),
            lora_int_id=lora_id,
            lora_path=lora_path_on_disk(lora_path),
        )
187
        return lora_request
188
189

    @abstractmethod
190
191
192
193
194
195
196
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
    ) -> list[SampleRequest]:
197
198
199
200
201
202
203
204
        """
        Abstract method to generate sample requests from the dataset.

        Subclasses must override this method to implement dataset-specific logic
        for generating a list of SampleRequest objects.

        Args:
            tokenizer (PreTrainedTokenizerBase): The tokenizer to be used
205
                for processing the dataset's text.
206
            num_requests (int): The number of sample requests to generate.
207
            request_id_prefix (str): The prefix of request_id.
208
209
210
211
212
213
214

        Returns:
            list[SampleRequest]: A list of sample requests generated from the
            dataset.
        """
        raise NotImplementedError("sample must be implemented in subclasses.")

215
216
217
218
219
    def maybe_oversample_requests(
        self,
        requests: list[SampleRequest],
        num_requests: int,
        request_id_prefix: str = "",
220
        no_oversample: bool = False,
221
    ) -> None:
222
223
224
225
226
227
        """
        Oversamples the list of requests if its size is less than the desired
        number.

        Args:
            requests (List[SampleRequest]): The current list of sampled
228
229
                requests.
            num_requests (int): The target number of requests.
230
231
            request_id_prefix (str): The prefix applied to generated request
                identifiers.
232

233
        """
234
        if no_oversample:
235
            logger.info("Skipping oversampling. Total samples: %d.", len(requests))
236
237
            return

238
239
        if len(requests) < num_requests:
            random.seed(self.random_seed)
240
241
242
243
            needed = num_requests - len(requests)
            additional = []
            for i in range(needed):
                req = deepcopy(random.choice(requests))
244
                req.request_id = request_id_prefix + str(len(requests) + i)
245
                additional.append(req)
246
            requests.extend(additional)
247
            logger.info("Oversampled requests to reach %d total samples.", num_requests)
248

249
250
        ids = [req.request_id for req in requests]
        if len(ids) != len(set(ids)):
251
252
253
254
255
            raise ValueError(
                "Duplicate request_id found in the sampled "
                "requests. Please ensure that each request_id "
                "is unique."
            )
256

257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279

# -----------------------------------------------------------------------------
# Utility Functions and Global Caches
# -----------------------------------------------------------------------------


def is_valid_sequence(
    prompt_len: int,
    output_len: int,
    min_len: int = 4,
    max_prompt_len: int = 1024,
    max_total_len: int = 2048,
    skip_min_output_len_check: bool = False,
) -> bool:
    """
    Validate a sequence based on prompt and output lengths.

    Default pruning criteria are copied from the original `sample_hf_requests`
    and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as
    from `sample_requests` in benchmark_throughput.py.
    """
    # Check for invalid conditions
    prompt_too_short = prompt_len < min_len
280
    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
281
282
283
284
    prompt_too_long = prompt_len > max_prompt_len
    combined_too_long = (prompt_len + output_len) > max_total_len

    # Return True if none of the invalid conditions are met
285
286
287
    return not (
        prompt_too_short or output_too_short or prompt_too_long or combined_too_long
    )
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302


@cache
def lora_path_on_disk(lora_path: str) -> str:
    return get_adapter_absolute_path(lora_path)


# Global cache for LoRA tokenizers.
lora_tokenizer_cache: dict[int, AnyTokenizer] = {}


def process_image(image: Any) -> Mapping[str, Any]:
    """
    Process a single image input and return a multimedia content dictionary.

303
    Supports the following input types:
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318

    1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key
       containing raw image data.  - Loads the bytes as a PIL.Image.Image.

    2. PIL.Image.Image input: - Converts the image to RGB.  - Saves the image as
       a JPEG in memory.  - Encodes the JPEG data as a base64 string.  - Returns
       a dictionary with the image as a base64 data URL.

    3. String input: - Treats the string as a URL or local file path.  -
       Prepends "file://" if the string doesn't start with "http://" or
       "file://".  - Returns a dictionary with the image URL.

    Raises:
        ValueError: If the input is not a supported type.
    """
319
320
    if isinstance(image, dict) and "bytes" in image:
        image = Image.open(BytesIO(image["bytes"]))
321
    if isinstance(image, Image.Image):
322
        image = convert_image_mode(image, "RGB")
323
324
        with io.BytesIO() as image_data:
            image.save(image_data, format="JPEG")
325
            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
326
327
        return {
            "type": "image_url",
328
            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
329
330
331
        }

    if isinstance(image, str):
332
333
334
335
336
        image_url = (
            image
            if image.startswith(("http://", "https://", "file://"))
            else f"file://{image}"
        )
337
338
        return {"type": "image_url", "image_url": {"url": image_url}}

339
340
341
342
    raise ValueError(
        f"Invalid image input {image}. Must be a PIL.Image.Image"
        " or str or dictionary with raw image bytes."
    )
343
344


345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
def process_video(video: Any) -> Mapping[str, Any]:
    """
    Process a single video input and return a multimedia content dictionary.

    Supports the following input types:

    1. Dictionary with raw video bytes: - Expects a dict with a 'bytes' key
       containing raw video data.

    2. String input: - Treats the string as a URL or local file path.  -
       Prepends "file://" if the string doesn't start with "http://" or
       "file://".  - Returns a dictionary with the image URL.

    Raises:
        ValueError: If the input is not a supported type.
    """
361
362
    if isinstance(video, dict) and "bytes" in video:
        video_bytes = video["bytes"]
363
364
365
        video_base64 = base64.b64encode(video_bytes).decode("utf-8")
        return {
            "type": "video_url",
366
            "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
367
368
369
        }

    if isinstance(video, str):
370
371
372
373
374
        video_url = (
            video
            if video.startswith(("http://", "https://", "file://"))
            else f"file://{video}"
        )
375
376
377
378
379
380
        return {"type": "video_url", "video_url": {"url": video_url}}

    raise ValueError(
        f"Invalid video input {video}. Must be a string of local path/remote url, or a dictionary with raw video bytes in the form of `{{'bytes': raw_video_bytes}}`."  # noqa: E501
    )

381
382
383
384
385
386
387

def gen_prompt_decode_to_target_len(
    tokenizer: PreTrainedTokenizerBase,
    token_sequence: list[int],
    target_token_len: int,
    max_retry: int = 10,
    add_special_tokens: bool = False,
388
    rng: np.random.Generator | None = None,
389
390
391
392
393
) -> tuple[str, list[int]]:
    """
    Ensure decoded-then-encoded prompt length matches the target token length.

    This function decodes an initial token sequence to text and re-encodes it
394
395
    , iteratively adjusting the token sequence length to match a target.
    This is necessary because some tokenizers do not guarantee a 1:1 mapping
396
397
398
399
400
401
402
403
404
405
406
    between consecutive tokens and the decoded-then-encoded sequence length.
    For example, for GPT2Tokenizer:
    [6880, 6881] -> ['Ġcalls', 'here'] ->
    [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']

    Returns a tuple of the final prompt string and the adjusted token sequence.
    """
    remain_num_try = max_retry
    token_mismatch = 0
    while True:
        prompt = tokenizer.decode(token_sequence)
407
        token_sequence = tokenizer.encode(prompt, add_special_tokens=add_special_tokens)
408
409
410
411
        if remain_num_try <= 0:
            if len(token_sequence) != target_token_len:
                token_mismatch = len(token_sequence) - target_token_len
            break
412

413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
        if len(token_sequence) == target_token_len:
            break
        elif len(token_sequence) < target_token_len:
            if rng is not None:
                extra_tokens = rng.integers(
                    0,
                    tokenizer.vocab_size,
                    size=target_token_len - len(token_sequence),
                ).tolist()
            else:
                extra_tokens = np.random.randint(
                    0,
                    tokenizer.vocab_size,
                    size=target_token_len - len(token_sequence),
                ).tolist()
            token_sequence.extend(extra_tokens)
        elif len(token_sequence) > target_token_len:
            token_sequence = token_sequence[:target_token_len]

        remain_num_try -= 1

    return prompt, token_sequence, token_mismatch

436

437
438
439
440
# -----------------------------------------------------------------------------
# Random Dataset Implementation (Synthetic Data)
# -----------------------------------------------------------------------------

441

442
class RandomDataset(BenchmarkDataset):
443
444
445
446
447
448
449
450
451
452
453
454
    """
    Synthetic text-only dataset for serving/throughput benchmarks.

    Strategy:
    - Sample input/output token lengths per request from integer-uniform ranges
      around configured means (controlled by range_ratio).
    - Prepend a fixed random prefix of length prefix_len.
    - Generate the remaining tokens as a reproducible sequence:
      (offset + index + arange(input_len)) % vocab_size.
    - Decode then re-encode/truncate to ensure prompt token counts match.
    - Uses numpy.default_rng seeded with random_seed for reproducible sampling.
    """
455

456
457
458
459
460
461
    # Default values copied from benchmark_serving.py for the random dataset.
    DEFAULT_PREFIX_LEN = 0
    DEFAULT_RANGE_RATIO = 0.0
    DEFAULT_INPUT_LEN = 1024
    DEFAULT_OUTPUT_LEN = 128

462
    def __init__(self, **kwargs) -> None:
463
        super().__init__(**kwargs)
464
465
466
467
        # Use numpy's default_rng for deterministic sampling
        # Do not use random.seed() or np.random.seed() elsewhere in this class.
        # This ensures that the RNG is isolated from global RNG state.
        self._rng = np.random.default_rng(self.random_seed)
468
469
470
471
472

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
473
        request_id_prefix: str = "",
474
        no_oversample: bool = False,
475
476
477
478
        prefix_len: int = DEFAULT_PREFIX_LEN,
        range_ratio: float = DEFAULT_RANGE_RATIO,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
479
        batchsize: int = 1,
480
481
        **kwargs,
    ) -> list[SampleRequest]:
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
        # validate total input tokens (prefix + sampled) is at least 1.
        num_special = int(tokenizer.num_special_tokens_to_add())
        real_input_len = max(0, int(input_len) - num_special)
        min_sampled_input = math.floor(real_input_len * (1.0 - float(range_ratio)))
        min_total_input = int(prefix_len) + min_sampled_input
        if min_total_input < 1:
            raise ValueError(
                "--random-input-len is too small: with tokenizer special "
                f"tokens {num_special} and --random-range-ratio {range_ratio}, "
                "the minimum possible total input tokens (prefix + sampled) is "
                f"{min_total_input}. Increase --random-input-len and/or "
                "--random-prefix-len, or decrease --random-range-ratio so that "
                "prefix_len + floor(max(0, random_input_len - num_special)) "
                "* (1 - range_ratio) >= 1."
            )

498
499
        input_lens, output_lens, offsets = self.get_sampling_params(
            num_requests, range_ratio, input_len, output_len, tokenizer
500
501
502
        )

        vocab_size = tokenizer.vocab_size
503
504
505
506
507
508
        prohibited_tokens = tokenizer.all_special_ids
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))

        # Generate prefix once
        prefix_token_ids = self.get_prefix(allowed_tokens, prefix_len)
509

510
        requests = []
511
        token_mismatch_total = 0
512
        for i in range(num_requests):
513
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
514
515
516
517
518
519
520
                tokenizer=tokenizer,
                prefix_token_ids=prefix_token_ids,
                prefix_len=prefix_len,
                vocab_size=vocab_size,
                input_len=int(input_lens[i]),
                offset=int(offsets[i]),
                index=i,
521
                allowed_tokens=allowed_tokens,
522
            )
523
            token_mismatch_total += token_mismatch
524
525
526
527
528
529
530
531
            requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
                    request_id=request_id_prefix + str(i),
                )
            )
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
        # only used for embeddings benchmark.
        if batchsize > 1:
            batch_requests = []
            # Create batched requests
            for i in range(0, num_requests, batchsize):
                batch = requests[i : i + batchsize]
                batch_requests.append(
                    SampleRequest(
                        prompt=[req.prompt for req in batch],
                        prompt_len=sum(req.prompt_len for req in batch),
                        expected_output_len=0,
                        request_id=request_id_prefix + str(i // batchsize),
                    )
                )
            requests = batch_requests
547

548
549
550
551
552
553
554
555
556
557
558
        if token_mismatch_total != 0:
            sign = "more" if token_mismatch_total > 0 else "fewer"
            logger.warning(
                "Across all generated prompts, there were %d %s tokens "
                "than expected after decoding and re-encoding. This is "
                "expected due to the imperfect nature of the sampling "
                "procedure.",
                abs(token_mismatch_total),
                sign,
            )

559
560
561
        return requests

    def get_prefix(
562
563
564
        self,
        allowed_tokens: np.ndarray,
        prefix_len: int,
565
566
567
568
569
    ) -> list[int]:
        """
        Get the prefix for the dataset.
        """
        return (
570
571
572
            allowed_tokens[
                self._rng.integers(0, len(allowed_tokens), size=prefix_len)
            ].tolist()
573
574
575
            if prefix_len > 0
            else []
        )
576

577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
    def get_sampling_params(
        self,
        num_requests: int,
        range_ratio: float,
        input_len: int,
        output_len: int,
        tokenizer: PreTrainedTokenizerBase,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        """
        Get the sampling parameters for the dataset.
        """
        # Enforce range_ratio < 1
        if not (0.0 <= range_ratio < 1.0):
            raise ValueError("range_ratio must be in [0, 1).")
        num_special_tokens = int(tokenizer.num_special_tokens_to_add())
        real_input_len = max(0, int(input_len) - num_special_tokens)
        # Bounds use floor for low and ceil for high
        input_low = math.floor(real_input_len * (1 - range_ratio))
        input_high = math.ceil(real_input_len * (1 + range_ratio))
        output_low = math.floor(output_len * (1 - range_ratio))
        output_high = math.ceil(output_len * (1 + range_ratio))
598
599
        # Ensure the lower bound for output length is at least 1 to
        # prevent sampling 0 tokens.
600
        output_low = max(output_low, 1)
601
        output_high = max(output_high, 1)
602
603
604

        if input_low > input_high:
            raise ValueError(
605
                f"Invalid input sampling interval: low={input_low} > high={input_high}"
606
607
608
609
610
611
            )
        if output_low > output_high:
            raise ValueError(
                "Invalid output sampling interval: "
                f"low={output_low} > high={output_high}"
            )
612

613
614
        logger.info(
            "Sampling input_len from [%s, %s] and output_len from [%s, %s]",
615
616
617
618
619
            input_low,
            input_high,
            output_low,
            output_high,
        )
620

621
622
623
        input_lens = self._rng.integers(input_low, input_high + 1, size=num_requests)
        output_lens = self._rng.integers(output_low, output_high + 1, size=num_requests)
        offsets = self._rng.integers(0, tokenizer.vocab_size, size=num_requests)
624
        return input_lens, output_lens, offsets
625

626
627
628
629
630
631
632
633
634
635
    def generate_token_sequence(
        self,
        *,
        tokenizer: PreTrainedTokenizerBase,
        prefix_token_ids: list[int],
        prefix_len: int,
        vocab_size: int,
        input_len: int,
        offset: int,
        index: int,
636
        allowed_tokens: np.ndarray,
637
    ) -> tuple[str, int, int]:
638
639
640
641
642
643
644
645
646
647
        """
        Returns (prompt, total_input_len).

        NOTE: After decoding the prompt we have to encode and decode it again.
        This is done because in some cases N consecutive tokens
        give a string tokenized into != N number of tokens.
        For example for GPT2Tokenizer:
        [6880, 6881] -> ['Ġcalls', 'here'] ->
        [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']
        To avoid uncontrolled change of the prompt length,
648
        the encoded sequence is truncated before being decoded again.
649
        """
650
651
652
653
654
        # Build the inner sequence by sampling
        # sequentially from the allowed tokens
        inner_seq = allowed_tokens[
            (offset + index + np.arange(input_len)) % len(allowed_tokens)
        ].tolist()
655
656
657
658
        token_sequence = prefix_token_ids + inner_seq

        # Decode, then re-encode and truncate to preserve token count invariants
        total_input_len = prefix_len + int(input_len)
659
        prompt, adjusted_token_sequence, token_mismatch = (
660
            gen_prompt_decode_to_target_len(
661
662
663
664
665
666
                tokenizer=tokenizer,
                token_sequence=token_sequence,
                target_token_len=total_input_len,
                add_special_tokens=False,
                rng=self._rng,
            )
667
668
669
        )
        total_input_len = len(adjusted_token_sequence)
        return prompt, total_input_len, token_mismatch
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
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
# -----------------------------------------------------------------------------
# Random Dataset Implementation (Synthetic Data)
# -----------------------------------------------------------------------------


class RandomDatasetForReranking(RandomDataset):
    """
    Random dataset specialized for the needs of scoring:
    - Batches of inputs
    - Inputs composed of pairs
    """

    def __init__(self, **kwargs) -> None:
        super().__init__(**kwargs)

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
        range_ratio: float = RandomDataset.DEFAULT_RANGE_RATIO,
        input_len: int = RandomDataset.DEFAULT_INPUT_LEN,
        batchsize: int = 1,
        is_reranker: bool = True,
        **kwargs,
    ) -> list[SampleRequest]:
        n_sep_tokens = int(is_reranker)

        query_len_param = (input_len // 2) - n_sep_tokens if is_reranker else input_len

        query_lens, _, query_offsets = self.get_sampling_params(
            1, range_ratio, query_len_param, 0, tokenizer
        )

        query_len = int(query_lens[0])

        if not is_reranker:
            assert num_requests > 1 and batchsize > 1
            num_requests -= 1
            batchsize -= 1
            doc_len_param = input_len
        else:
            doc_len_param = input_len - query_len - n_sep_tokens

        doc_lens, _, doc_offsets = self.get_sampling_params(
            num_requests, range_ratio, doc_len_param, 0, tokenizer
        )
        vocab_size = tokenizer.vocab_size

        query_prompt, query_input_len, token_mismatch_total = (
            self.generate_token_sequence(
                tokenizer=tokenizer,
                prefix_token_ids=[],
                prefix_len=0,
                vocab_size=vocab_size,
                input_len=query_len,
                offset=int(query_offsets[0]),
                index=0,
            )
        )

        requests = []
        for i in range(num_requests):
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
                tokenizer=tokenizer,
                prefix_token_ids=[],
                prefix_len=0,
                vocab_size=vocab_size,
                input_len=int(doc_lens[i]),
                offset=int(doc_offsets[i]),
                index=i + 1,
            )
            token_mismatch_total += token_mismatch
            requests.append((prompt, total_input_len))

        batch_requests = []
        # Create batched requests
        for i in range(0, num_requests, batchsize):
            batch = requests[i : i + batchsize]
            query_contrib = (
                (query_input_len + n_sep_tokens) * len(batch)
                if is_reranker
                else query_input_len
            )
            batch_requests.append(
                SampleRequest(
                    prompt=[query_prompt] + [req[0] for req in batch],
                    prompt_len=query_contrib + sum(req[1] for req in batch),
                    expected_output_len=0,
                    request_id=request_id_prefix + str(i // batchsize),
                )
            )

        if token_mismatch_total != 0:
            logger.warning(
                "Across all generated prompts, there were %d %s tokens "
                "than expected after decoding and re-encoding. This is "
                "expected due to the imperfect nature of the sampling "
                "procedure.",
                abs(token_mismatch_total),
                "more" if token_mismatch_total > 0 else "fewer",
            )

        return batch_requests


778
779
780
781
# -----------------------------------------------------------------------------
# MultiModalDataset Implementation
# -----------------------------------------------------------------------------

782

783
784
785
786
787
788
class RandomMultiModalDataset(RandomDataset):
    """
    Synthetic multimodal dataset (text + images) that extends RandomDataset.

    Status:
    - Images: supported via synthetic RGB data.
789
    - Video: supported via synthetic RGB data.
790
791
792
793
794
795
796
797
    - Audio: not yet supported.

    Sampling overview:
    1) Number of items per request is sampled uniformly from the integer range
       [floor(n·(1−r)), ceil(n·(1+r))], where n is the base count and r is
       `num_mm_items_range_ratio` in [0, 1]. r=0 keeps it fixed; r=1 allows 0.
       The maximum is further clamped to the sum of per-modality limits.
    2) Each item’s modality and shape is sampled from `bucket_config`, a dict
798
       mapping (height, width, num_frames) → probability. We treat
799
       `num_frames`=1 as image and `num_frames` > 1 as video.
800
       Entries with zero probability are removed and the rest are renormalized
801
802
803
804
805
806
807
       to sum to 1.
    3) Per-modality hard caps are enforced via `limit_mm_per_prompt`.
       When a modality reaches its cap, all of its buckets are excluded and the
       remaining probabilities are renormalized.

    Example bucket configuration:
    {(256, 256, 1): 0.5, (720, 1280, 1): 0.4, (720, 1280, 16): 0.1}
808
809
      - Two image buckets (`num_frames`=1) and one video bucket
      (`num_frames`=16).
810
811
812
813
    OBS.: Only image sampling is supported for now.
    """

    IS_MULTIMODAL = True
814
    DEFAULT_LIMIT_MM_PER_PROMPT = {"image": 255, "video": 1}
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829

    DEFAULT_BASE_ITEMS_PER_REQUEST = 1
    DEFAULT_NUM_MM_ITEMS_RANGE_RATIO = 0.0
    DEFAULT_MM_ITEM_BUCKET_CONFIG = {
        (256, 256, 1): 0.5,
        (720, 1280, 1): 0.5,
        (720, 1280, 16): 0.0,
    }
    DEFAULT_ENABLE_MULTIMODAL_CHAT = False

    def __init__(self, **kwargs) -> None:
        super().__init__(**kwargs)

    def generate_synthetic_image(self, width: int, height: int) -> Image.Image:
        """Generate synthetic PIL image with random RGB values.
830
831
832

        NOTE: iid pixel sampling results in worst-case compression
        (good for stressing I/O), but very unlike real photos.
833
834
835
836
837
838
839
840
841
842
843
        We could consider a “low-freq” mode (e.g., noise blur)
        to emulate network realism instead of max stress.
        """
        random_pixels = self._rng.integers(
            0,
            256,
            (height, width, 3),
            dtype=np.uint8,
        )
        return Image.fromarray(random_pixels)

844
845
846
    def generate_synthetic_video(
        self, width: int, height: int, num_frames: int
    ) -> dict:
847
        """Generate synthetic video with random values.
848

849
850
        Creates a video with random pixel values, encodes it to MP4 format,
        and returns the content as bytes.
851
        """
852
853
        import cv2

854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
        random_pixels = self._rng.integers(
            0,
            256,
            (num_frames, height, width, 3),
            dtype=np.uint8,
        )

        # Create a temporary video file in memory
        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
        fps = 30  # frames per second

        with NamedTemporaryFile(suffix=".mp4", delete_on_close=False) as temp_file:
            temp_path = temp_file.name

            # Create video writer
            video_writer = cv2.VideoWriter(
                temp_path, fourcc=fourcc, fps=fps, frameSize=(width, height)
            )

            if not video_writer.isOpened():
                raise RuntimeError("Failed to create video writer")

            for frame in random_pixels:
                video_writer.write(frame)

            video_writer.release()
            temp_file.close()

            # Read the video file content
            with open(temp_path, "rb") as f:
                video_content = f.read()

            return {"bytes": video_content}
887
888
889
890
891
892
893
894
895
896

    def map_config_to_modality(self, config: tuple[int, int, int]) -> str:
        """Map the configuration to the modality."""
        if config[-1] == 1:
            return "image"
        elif config[-1] > 1:
            return "video"
        else:
            raise ValueError(f"Invalid multimodal item configuration: {config}")

897
898
899
    def normalize_bucket_config(
        self, bucket_config: dict[tuple[int, int, int], float]
    ) -> dict[tuple[int, int, int], float]:
900
901
902
903
904
905
906
907
908
909
910
        """
        Remove zero probability entries
        and normalize the bucket config to sum to 1.
        """
        # Raise error if value is negative
        if any(v < 0 for v in bucket_config.values()):
            raise ValueError("Bucket config values must be non-negative.")
        # Remove zero probability entries
        bucket_config = {k: v for k, v in bucket_config.items() if v > 0}
        # if bucket config is empty, raise error
        if not bucket_config:
911
912
913
            raise ValueError(
                "Got invalid bucket config. Bucket config values must be non-zero."
            )
914
915
916
917
        # Normalize the remaining bucket config to sum to 1
        total = sum(bucket_config.values())
        return {k: v / total for k, v in bucket_config.items()}

918
919
920
921
    def generate_mm_item(
        self,
        mm_item_config: tuple[int, int, int],
    ) -> Mapping[str, Any]:
922
        """
923
        Create synthetic images and videos and
924
925
926
927
        apply process_image/process_video respectively.
        This follows the OpenAI API chat completions
        https://github.com/openai/openai-python
        """
928

929
        if self.map_config_to_modality(mm_item_config) == "image":
930
931
932
            return process_image(
                self.generate_synthetic_image(mm_item_config[1], mm_item_config[0])
            )
933
        elif self.map_config_to_modality(mm_item_config) == "video":
934
935
936
937
938
            return process_video(
                self.generate_synthetic_video(
                    mm_item_config[1], mm_item_config[0], mm_item_config[2]
                )
            )
939
        else:
940
            raise ValueError(f"Invalid multimodal item configuration: {mm_item_config}")
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960

    def get_mm_item_sampling_params(
        self,
        base_items_per_request: int,
        num_mm_items_range_ratio: float,
        limit_mm_per_prompt: dict[str, int],
        bucket_config: dict[tuple[int, int, int], float],
    ) -> tuple[int, int, dict[str, int], dict[tuple[int, int, int], float]]:
        """
        Get the sampling parameters for the multimodal items.
        """
        # Enforce num_mm_items_range_ratio <= 1
        if not (0.0 <= num_mm_items_range_ratio <= 1.0):
            raise ValueError("num_mm_items_range_ratio must be in [0, 1].")

        # Ensure modalities to sample are in limit_mm_per_prompt
        for k, v in bucket_config.items():
            # get modality from bucket config
            modality = self.map_config_to_modality(k)
            if modality not in limit_mm_per_prompt:
961
962
963
964
965
                raise ValueError(
                    f"Modality {modality} is not in "
                    f"limit_mm_per_prompt: "
                    f"{limit_mm_per_prompt.keys()}"
                )
966

967
        # Remove zero probability entries
968
969
970
        # and normalize bucket config to sum to 1
        bucket_config = self.normalize_bucket_config(bucket_config)
        logger.info(
971
972
            "Normalized bucket config: %s",
            bucket_config,
973
974
        )
        # Only consider limit per prompt for modalities in bucket config
975
        allowed_modalities = {self.map_config_to_modality(cfg) for cfg in bucket_config}
976
        limit_mm_per_prompt = {
977
978
            k: v for k, v in limit_mm_per_prompt.items() if k in allowed_modalities
        }
979
        if not limit_mm_per_prompt:
980
            raise ValueError("No valid limits for modalities present in bucket_config.")
981
982

        logger.info(
983
984
            "Updated mm-limit-per-prompt: %s",
            limit_mm_per_prompt,
985
986
987
988
989
        )

        # Get max and min num mm items and ensure
        # it is at most the sum of limit_mm_per_prompt for all modalities
        max_num_mm_items = min(
990
            sum(limit_mm_per_prompt.values()),
991
            math.ceil(base_items_per_request * (1 + num_mm_items_range_ratio)),
992
993
994
        )
        # Ensure min num mm items is at least 0
        min_num_mm_items = max(
995
            0, math.floor(base_items_per_request * (1 - num_mm_items_range_ratio))
996
997
998
        )
        # Raise error if min num mm items is greater than max num mm items
        if min_num_mm_items > max_num_mm_items:
999
1000
1001
1002
            raise ValueError(
                f"Min num mm items is greater than max mm items: "
                f"{min_num_mm_items} > {max_num_mm_items}"
            )
1003

1004
1005
        logger.info(
            "Sampling number of multimodal items from [%s, %s]",
1006
1007
            min_num_mm_items,
            max_num_mm_items,
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
        )

        return (
            min_num_mm_items,
            max_num_mm_items,
            limit_mm_per_prompt,
            bucket_config,
        )

    def get_mm_item_iterator(
        self,
        min_num_mm_items: int,
        max_num_mm_items: int,
        bucket_config: dict[tuple[int, int, int], float],
        limit_mm_per_prompt: dict[str, int],
1023
    ) -> Iterator[tuple[int, int, int]]:
1024
1025
1026
1027
1028
        """
        Iterator over the multimodal items for each request
        whose size is between min_num_mm_items and max_num_mm_items.

        Loop over the bucket config and sample a multimodal item.
1029
1030
        Loop until the number of multimodal items sampled is equal to
        request_num_mm_items or limit of multimodal items per prompt
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
        for all modalities is reached.

        Note:
        - This function operates on a per-request shallow copy of
          `bucket_config` (tuple->float). The original dict passed to
          `sample` is not mutated. If this ever changes, a test
          is implemented and will fail.
        """
        # Get the number of multimodal items to sample
        request_num_mm_items = int(
            self._rng.integers(min_num_mm_items, max_num_mm_items + 1)
1042
        )
1043
1044
1045
1046
        # If request_num_mm_items is 0, yield an empty iterator
        if request_num_mm_items == 0:
            return
        # Initialize modality counters
1047
        modality_counter = {self.map_config_to_modality(k): 0 for k in bucket_config}
1048
1049
1050
1051
1052
        # Copy the bucket config to avoid modifying the original
        bucket_config_copy = bucket_config.copy()
        # Loop over the number of multimodal items to sample
        while sum(modality_counter.values()) < request_num_mm_items:
            # Sample a multimodal item config
1053
1054
1055
            mm_item_config = self._rng.choice(
                list(bucket_config_copy.keys()), p=list(bucket_config_copy.values())
            )
1056
1057
1058
1059
            modality = self.map_config_to_modality(mm_item_config)
            # Check that modality count is less than limit per prompt
            if modality_counter[modality] < limit_mm_per_prompt[modality]:
                modality_counter[modality] += 1
1060
                yield (mm_item_config)
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
            else:
                # If the counter is greater than the limit per prompt
                # set all multimodal items of this modality to 0
                for k, v in bucket_config_copy.items():
                    if self.map_config_to_modality(k) == modality:
                        bucket_config_copy[k] = 0
                # If all configs are 0, break the loop
                # This should not happen as request_num_mm_items is at most
                # the sum of limit_mm_per_prompt for all modalities
                if all(v == 0 for v in bucket_config_copy.values()):
1071
1072
1073
                    logger.warning(
                        "Exhausted all multimodal items of modality %s", modality
                    )
1074
1075
                    break
                # Renormalize the bucket config
1076
                bucket_config_copy = self.normalize_bucket_config(bucket_config_copy)
1077
1078
1079
1080
1081
1082

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
1083
        no_oversample: bool = False,
1084
1085
1086
1087
1088
1089
1090
        prefix_len: int = RandomDataset.DEFAULT_PREFIX_LEN,
        range_ratio: float = RandomDataset.DEFAULT_RANGE_RATIO,
        input_len: int = RandomDataset.DEFAULT_INPUT_LEN,
        output_len: int = RandomDataset.DEFAULT_OUTPUT_LEN,
        limit_mm_per_prompt: dict[str, int] = DEFAULT_LIMIT_MM_PER_PROMPT,
        base_items_per_request: int = DEFAULT_BASE_ITEMS_PER_REQUEST,
        num_mm_items_range_ratio: float = DEFAULT_NUM_MM_ITEMS_RANGE_RATIO,
1091
1092
1093
        bucket_config: dict[
            tuple[int, int, int], float
        ] = DEFAULT_MM_ITEM_BUCKET_CONFIG,
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
        enable_multimodal_chat: bool = DEFAULT_ENABLE_MULTIMODAL_CHAT,
        **kwargs,
    ) -> list[SampleRequest]:
        # Get the sampling parameters for the dataset
        input_lens, output_lens, offsets = self.get_sampling_params(
            num_requests, range_ratio, input_len, output_len, tokenizer
        )

        (
            min_num_mm_items,
            max_num_mm_items,
            limit_mm_per_prompt,
            bucket_config,
        ) = self.get_mm_item_sampling_params(
            base_items_per_request,
            num_mm_items_range_ratio,
            limit_mm_per_prompt,
            bucket_config,
        )

        vocab_size = tokenizer.vocab_size
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
        # Can't use tokenizer.all_special_ids since
        # it returns ONLY ids from special_tokens_map.json
        # We want to exclude placeholder tokens and all
        # tokens that indicate start/end of image as it
        # may break prompt replacement logic.
        prohibited_tokens = list(
            tok_id
            for tok_id, token in tokenizer.added_tokens_decoder.items()
            if token.special
        )
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))
        logger.debug(
            "Sampling from %d out of %d (vocab size)", len(allowed_tokens), vocab_size
        )
        # Generate prefix once
        prefix_token_ids = self.get_prefix(allowed_tokens, prefix_len)
1132
1133
        # Add synthetic multimodal items to each request
        mm_requests = []
1134
        token_mismatch_total = 0
1135
        for i in range(num_requests):
1136
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
1137
1138
1139
1140
1141
1142
1143
                tokenizer=tokenizer,
                prefix_token_ids=prefix_token_ids,
                prefix_len=prefix_len,
                vocab_size=vocab_size,
                input_len=int(input_lens[i]),
                offset=int(offsets[i]),
                index=i,
1144
                allowed_tokens=allowed_tokens,
1145
            )
1146
            token_mismatch_total += token_mismatch
1147
1148
1149
1150
1151
1152
1153
1154
            # Get multimodal item iterator for a given request
            mm_item_iterator = self.get_mm_item_iterator(
                min_num_mm_items,
                max_num_mm_items,
                bucket_config,
                limit_mm_per_prompt,
            )

1155
1156
1157
1158
1159
1160
1161
            mm_content = cast(
                list[dict[str, Any]],
                [
                    self.generate_mm_item(mm_item_config)
                    for mm_item_config in mm_item_iterator
                ],
            )
1162
1163

            if enable_multimodal_chat:
1164
                # NOTE: For now this option is only provided for completeness
1165
1166
1167
                # given that the serve.py benchmark currently does not use it.
                mm_chat_prompt: Any = prompt
                mm_chat_prompt = self.apply_multimodal_chat_transformation(
1168
1169
                    prompt, mm_content
                )
1170
1171
1172
1173
1174
1175
1176
1177
1178
                sample_request = SampleRequest(
                    prompt=mm_chat_prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
                    multi_modal_data=None,
                    request_id=request_id_prefix + str(i),
                )
            else:
                sample_request = SampleRequest(
1179
1180
1181
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
1182
                    multi_modal_data=mm_content,
1183
                    request_id=request_id_prefix + str(i),
1184
1185
                )
            mm_requests.append(sample_request)
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197

        if token_mismatch_total != 0:
            sign = "more" if token_mismatch_total > 0 else "fewer"
            logger.warning(
                "Across all generated prompts, there were %d %s tokens "
                "than expected after decoding and re-encoding. This is "
                "expected due to the imperfect nature of the sampling "
                "procedure.",
                abs(token_mismatch_total),
                sign,
            )

1198
        return mm_requests
1199

1200

1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
# -----------------------------------------------------------------------------
# ShareGPT Dataset Implementation
# -----------------------------------------------------------------------------


class ShareGPTDataset(BenchmarkDataset):
    """
    Implements the ShareGPT dataset.  Loads data from a JSON file and generates
    sample requests based on conversation turns.
    """

    def __init__(self, **kwargs) -> None:
        super().__init__(**kwargs)
        self.load_data()

    def load_data(self) -> None:
        if self.dataset_path is None:
            raise ValueError("dataset_path must be provided for loading data.")

        with open(self.dataset_path, encoding="utf-8") as f:
            self.data = json.load(f)
        # Filter entries with at least two conversation turns.
        self.data = [
1224
1225
            entry
            for entry in self.data
1226
1227
1228
            if "conversations" in entry and len(entry["conversations"]) >= 2
        ]
        random.seed(self.random_seed)
1229
1230
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1231
1232
1233
1234
1235

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
1236
1237
1238
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1239
        enable_multimodal_chat: bool = False,
1240
        request_id_prefix: str = "",
1241
        no_oversample: bool = False,
1242
1243
1244
        **kwargs,
    ) -> list:
        samples: list = []
1245
        ind = 0
1246
1247
1248
1249
1250
1251
1252
1253
        for entry in self.data:
            if len(samples) >= num_requests:
                break
            prompt, completion = (
                entry["conversations"][0]["value"],
                entry["conversations"][1]["value"],
            )

1254
            lora_request = self.get_random_lora_request(
1255
1256
                max_loras=max_loras, lora_path=lora_path
            )
1257
1258
1259
            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
1260
1261
1262
1263
1264
1265
            new_output_len = len(completion_ids) if output_len is None else output_len
            if not is_valid_sequence(
                prompt_len,
                new_output_len,
                skip_min_output_len_check=output_len is not None,
            ):
1266
                continue
1267
1268
1269
            if image_path := entry.get("image"):
                mm_content = process_image(image_path)
            elif video_path := entry.get("video"):
1270
                mm_content = process_video(video_path)
1271
            else:
1272
                mm_content = None
1273
            if enable_multimodal_chat:
1274
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
1275
1276
1277
1278
1279
1280
            samples.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=new_output_len,
                    lora_request=lora_request,
1281
                    multi_modal_data=mm_content,
1282
                    request_id=request_id_prefix + str(ind),
1283
1284
                )
            )
1285
            ind += 1
1286
1287
1288
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
1289
1290
1291
        return samples


1292
1293
class _ValidateDatasetArgs(argparse.Action):
    """Argparse action to validate dataset name and path compatibility."""
1294

1295
1296
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, values)
1297

1298
        # Get current values of both dataset_name and dataset_path
1299
1300
        dataset_name = getattr(namespace, "dataset_name", "random")
        dataset_path = getattr(namespace, "dataset_path", None)
1301

1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
        # Validate the combination
        if dataset_name == "random" and dataset_path is not None:
            parser.error(
                "Cannot use 'random' dataset with --dataset-path. "
                "Please specify the appropriate --dataset-name (e.g., "
                "'sharegpt', 'custom', 'sonnet') for your dataset file: "
                f"{dataset_path}"
            )


1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
def add_dataset_parser(parser: FlexibleArgumentParser):
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
        "--num-prompts",
        type=int,
        default=1000,
        help="Number of prompts to process.",
    )
    parser.add_argument(
        "--dataset-name",
        type=str,
        default="random",
1324
        action=_ValidateDatasetArgs,
1325
        choices=[
1326
1327
1328
1329
1330
            "sharegpt",
            "burstgpt",
            "sonnet",
            "random",
            "random-mm",
1331
            "random-rerank",
1332
1333
1334
1335
            "hf",
            "custom",
            "prefix_repetition",
            "spec_bench",
1336
        ],
1337
1338
        help="Name of the dataset to benchmark on.",
    )
1339
1340
1341
1342
1343
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
1344
1345
1346
1347
    parser.add_argument(
        "--dataset-path",
        type=str,
        default=None,
1348
        action=_ValidateDatasetArgs,
1349
1350
1351
        help="Path to the sharegpt/sonnet dataset. "
        "Or the huggingface dataset ID if using HF dataset.",
    )
1352
1353
1354
    parser.add_argument(
        "--no-oversample",
        action="store_true",
1355
        help="Do not oversample if the dataset has fewer samples than num-prompts.",
1356
    )
1357
1358
1359
    parser.add_argument(
        "--skip-chat-template",
        action="store_true",
1360
        help="Skip applying chat template to prompt for datasets that support it.",
1361
    )
1362
1363
1364
1365
1366
    parser.add_argument(
        "--disable-shuffle",
        action="store_true",
        help="Disable shuffling of dataset samples for deterministic ordering.",
    )
1367
1368
1369
1370
1371
1372
1373

    # group for dataset specific arguments
    custom_group = parser.add_argument_group("custom dataset options")
    custom_group.add_argument(
        "--custom-output-len",
        type=int,
        default=256,
1374
        help="Number of output tokens per request, used only for custom dataset.",
1375
1376
    )

1377
1378
1379
1380
1381
    spec_bench_group = parser.add_argument_group("spec bench dataset options")
    spec_bench_group.add_argument(
        "--spec-bench-output-len",
        type=int,
        default=256,
1382
        help="Num of output tokens per request, used only for spec bench dataset.",
1383
1384
1385
1386
1387
    )
    spec_bench_group.add_argument(
        "--spec-bench-category",
        type=str,
        default=None,
1388
        help="Category for spec bench dataset. If None, use all categories.",
1389
1390
    )

1391
1392
1393
1394
1395
    sonnet_group = parser.add_argument_group("sonnet dataset options")
    sonnet_group.add_argument(
        "--sonnet-input-len",
        type=int,
        default=550,
1396
        help="Number of input tokens per request, used only for sonnet dataset.",
1397
1398
1399
1400
1401
    )
    sonnet_group.add_argument(
        "--sonnet-output-len",
        type=int,
        default=150,
1402
        help="Number of output tokens per request, used only for sonnet dataset.",
1403
1404
1405
1406
1407
    )
    sonnet_group.add_argument(
        "--sonnet-prefix-len",
        type=int,
        default=200,
1408
        help="Number of prefix tokens per request, used only for sonnet dataset.",
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
    )

    sharegpt_group = parser.add_argument_group("sharegpt dataset options")
    sharegpt_group.add_argument(
        "--sharegpt-output-len",
        type=int,
        default=None,
        help="Output length for each request. Overrides the output length "
        "from the ShareGPT dataset.",
    )

1420
1421
1422
1423
1424
    blazedit_group = parser.add_argument_group("blazedit dataset options")
    blazedit_group.add_argument(
        "--blazedit-min-distance",
        type=float,
        default=0.0,
1425
        help="Minimum distance for blazedit dataset. Min: 0, Max: 1.0",
1426
1427
1428
1429
1430
    )
    blazedit_group.add_argument(
        "--blazedit-max-distance",
        type=float,
        default=1.0,
1431
        help="Maximum distance for blazedit dataset. Min: 0, Max: 1.0",
1432
1433
    )

1434
1435
1436
1437
1438
    random_group = parser.add_argument_group("random dataset options")
    random_group.add_argument(
        "--random-input-len",
        type=int,
        default=1024,
1439
        help="Number of input tokens per request, used only for random sampling.",
1440
1441
1442
1443
1444
    )
    random_group.add_argument(
        "--random-output-len",
        type=int,
        default=128,
1445
        help="Number of output tokens per request, used only for random sampling.",
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
    )
    random_group.add_argument(
        "--random-range-ratio",
        type=float,
        default=0.0,
        help="Range ratio for sampling input/output length, "
        "used only for random sampling. Must be in the range [0, 1) to define "
        "a symmetric sampling range"
        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
    )
    random_group.add_argument(
        "--random-prefix-len",
        type=int,
        default=0,
1460
1461
1462
1463
1464
1465
1466
1467
        help=(
            "Number of fixed prefix tokens before the random context "
            "in a request. "
            "The total input length is the sum of `random-prefix-len` and "
            "a random "
            "context length sampled from [input_len * (1 - range_ratio), "
            "input_len * (1 + range_ratio)]."
        ),
1468
    )
1469
1470
1471
1472
    random_group.add_argument(
        "--random-batch-size",
        type=int,
        default=1,
1473
        help=("Batch size for random sampling. Only used for embeddings benchmark."),
1474
    )
1475
1476
1477
1478
1479
1480
1481
1482
    random_group.add_argument(
        "--no-reranker",
        action="store_true",
        help=(
            "Whether the model supports reranking natively."
            " Only used for reranker benchmark."
        ),
    )
1483

1484
1485
    # random multimodal dataset options
    random_mm_group = parser.add_argument_group(
1486
1487
        "random multimodal dataset options extended from random dataset"
    )
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
    random_mm_group.add_argument(
        "--random-mm-base-items-per-request",
        type=int,
        default=RandomMultiModalDataset.DEFAULT_BASE_ITEMS_PER_REQUEST,
        help=(
            "Base number of multimodal items per request for random-mm. "
            "Actual per-request count is sampled around this base using "
            "--random-mm-num-mm-items-range-ratio."
        ),
    )
    random_mm_group.add_argument(
        "--random-mm-num-mm-items-range-ratio",
        type=float,
        default=RandomMultiModalDataset.DEFAULT_NUM_MM_ITEMS_RANGE_RATIO,
        help=(
            "Range ratio r in [0, 1] for sampling items per request. "
            "We sample uniformly from the closed integer range "
            "[floor(n*(1-r)), ceil(n*(1+r))] "
            "where n is the base items per request. "
            "r=0 keeps it fixed; r=1 allows 0 items. The maximum is clamped "
            "to the sum of per-modality limits from "
            "--random-mm-limit-mm-per-prompt. "
            "An error is raised if the computed min exceeds the max."
        ),
    )
    random_mm_group.add_argument(
        "--random-mm-limit-mm-per-prompt",
        type=json.loads,
        default=RandomMultiModalDataset.DEFAULT_LIMIT_MM_PER_PROMPT,
        help=(
            "Per-modality hard caps for items attached per request, e.g. "
1519
            '\'{"image": 3, "video": 0}\'. The sampled per-request item '
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
            "count is clamped to the sum of these limits. When a modality "
            "reaches its cap, its buckets are excluded and probabilities are "
            "renormalized."
            "OBS.: Only image sampling is supported for now."
        ),
    )

    def _parse_mm_bucket_config(v: object) -> dict[tuple[int, int, int], float]:
        # If already a dict (e.g., programmatic call), normalize keys
        def normalize(d: dict) -> dict[tuple[int, int, int], float]:
            out: dict[tuple[int, int, int], float] = {}
            for k, val in d.items():
                key = k
                if isinstance(key, str):
                    with suppress(Exception):
                        key = ast.literal_eval(key)
1536
1537
1538
1539
1540
                if not (
                    isinstance(key, tuple)
                    and len(key) == 3
                    and all(isinstance(x, int) for x in key)
                ):
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
                    raise ValueError(
                        f"Invalid bucket key {k!r}. Expected tuple (H, W, T)."
                    )
                out[(int(key[0]), int(key[1]), int(key[2]))] = float(val)
            return out

        if isinstance(v, dict):
            return normalize(v)
        if isinstance(v, str):
            # Python literal (supports tuple keys)
            parsed = ast.literal_eval(v)
            if not isinstance(parsed, dict):
                raise ValueError("Bucket config must parse to a dict.")
            return normalize(parsed)
        raise ValueError("Unsupported value for --random-mm-bucket-config.")

    random_mm_group.add_argument(
        "--random-mm-bucket-config",
        type=_parse_mm_bucket_config,
        default=RandomMultiModalDataset.DEFAULT_MM_ITEM_BUCKET_CONFIG,
        help=(
            "The bucket config is a dictionary mapping a multimodal item"
            "sampling configuration to a probability."
            "Currently allows for 2 modalities: images and videos. "
            "An bucket key is a tuple of (height, width, num_frames)"
            "The value is the probability of sampling that specific item. "
            "Example: "
            "--random-mm-bucket-config "
            "{(256, 256, 1): 0.5, (720, 1280, 1): 0.4, (720, 1280, 16): 0.10} "
            "First item: images with resolution 256x256 w.p. 0.5"
            "Second item: images with resolution 720x1280 w.p. 0.4 "
            "Third item: videos with resolution 720x1280 and 16 frames w.p. 0.1"
            "OBS.: If the probabilities do not sum to 1, they are normalized."
            "OBS bis.: Only image sampling is supported for now."
        ),
    )

1578
    hf_group = parser.add_argument_group("hf dataset options")
1579
1580
1581
1582
1583
1584
    hf_group.add_argument(
        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
    )
    hf_group.add_argument(
        "--hf-split", type=str, default=None, help="Split of the HF dataset."
    )
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
    hf_group.add_argument(
        "--hf-name",
        type=str,
        default=None,
        help=(
            "Name of the dataset on HuggingFace "
            "(e.g., 'lmarena-ai/VisionArena-Chat'). "
            "Specify this if your dataset-path is a local path."
        ),
    )
1595
1596
1597
1598
1599
1600
1601
1602
    hf_group.add_argument(
        "--hf-output-len",
        type=int,
        default=None,
        help="Output length for each request. Overrides the output lengths "
        "from the sampled HF dataset.",
    )

1603
    prefix_repetition_group = parser.add_argument_group(
1604
1605
        "prefix repetition dataset options"
    )
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
    prefix_repetition_group.add_argument(
        "--prefix-repetition-prefix-len",
        type=int,
        default=256,
        help="Number of prefix tokens per request, used only for prefix "
        "repetition dataset.",
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-suffix-len",
        type=int,
        default=256,
        help="Number of suffix tokens per request, used only for prefix "
        "repetition dataset. Total input length is prefix_len + suffix_len.",
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-num-prefixes",
        type=int,
        default=10,
        help="Number of prefixes to generate, used only for prefix repetition "
        "dataset. Prompts per prefix is num_requests // num_prefixes.",
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-output-len",
        type=int,
        default=128,
        help="Number of output tokens per request, used only for prefix "
        "repetition dataset.",
    )

1635
1636

def get_samples(args, tokenizer) -> list[SampleRequest]:
1637
1638
1639
    if not hasattr(args, "request_id_prefix"):
        args.request_id_prefix = ""

1640
    if args.dataset_name == "custom":
1641
1642
1643
        dataset = CustomDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1644
1645
1646
1647
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
1648
            skip_chat_template=args.skip_chat_template,
1649
            request_id_prefix=args.request_id_prefix,
1650
            no_oversample=args.no_oversample,
1651
1652
1653
        )

    elif args.dataset_name == "sonnet":
1654
1655
1656
        dataset = SonnetDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1657
        # For the "sonnet" dataset, formatting depends on the backend.
1658
        if args.backend == "openai-chat":
1659
1660
1661
1662
1663
1664
1665
            input_requests = dataset.sample(
                num_requests=args.num_prompts,
                input_len=args.sonnet_input_len,
                output_len=args.sonnet_output_len,
                prefix_len=args.sonnet_prefix_len,
                tokenizer=tokenizer,
                return_prompt_formatted=False,
1666
                request_id_prefix=args.request_id_prefix,
1667
                no_oversample=args.no_oversample,
1668
1669
1670
            )
        else:
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
1671
1672
                "Tokenizer/model must have chat template for sonnet dataset."
            )
1673
1674
1675
1676
1677
1678
1679
            input_requests = dataset.sample(
                num_requests=args.num_prompts,
                input_len=args.sonnet_input_len,
                output_len=args.sonnet_output_len,
                prefix_len=args.sonnet_prefix_len,
                tokenizer=tokenizer,
                return_prompt_formatted=True,
1680
                request_id_prefix=args.request_id_prefix,
1681
                no_oversample=args.no_oversample,
1682
1683
1684
1685
1686
            )

    elif args.dataset_name == "hf":
        # all following datasets are implemented from the
        # HuggingFaceDataset base class
1687
        hf_kwargs = {}
1688
1689
1690
1691
        if (
            args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS
        ):
1692
1693
1694
            dataset_class = VisionArenaDataset
            args.hf_split = "train"
            args.hf_subset = None
1695
1696
1697
1698
1699
1700
1701
        elif (
            args.dataset_path in MMVUDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMVUDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMVUDataset
            args.hf_split = "validation"
            args.hf_subset = None
1702
1703
1704
1705
        elif (
            args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS
        ):
1706
1707
            dataset_class = InstructCoderDataset
            args.hf_split = "train"
1708
1709
1710
1711
        elif (
            args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MTBenchDataset.SUPPORTED_DATASET_PATHS
        ):
1712
1713
            dataset_class = MTBenchDataset
            args.hf_split = "train"
1714
1715
1716
1717
        elif (
            args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS
        ):
1718
            dataset_class = ConversationDataset
1719
1720
1721
1722
        elif (
            args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS
        ):
1723
1724
            dataset_class = AIMODataset
            args.hf_split = "train"
1725
        elif (
1726
            args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS  # noqa: E501
1727
1728
            or args.hf_name in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS
        ):
1729
1730
            dataset_class = NextEditPredictionDataset
            args.hf_split = "train"
1731
1732
1733
1734
        elif (
            args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS
        ):
1735
1736
            dataset_class = ASRDataset
            args.hf_split = "train"
1737
1738
1739
1740
1741
1742
1743
        elif args.dataset_path in BlazeditDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = BlazeditDataset
            args.hf_split = "train"
            hf_kwargs = {
                "min_distance": args.blazedit_min_distance,
                "max_distance": args.blazedit_max_distance,
            }
1744
1745
1746
1747
        elif (
            args.dataset_path in MLPerfDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MLPerfDataset.SUPPORTED_DATASET_PATHS
        ):
1748
1749
            dataset_class = MLPerfDataset
            args.hf_split = "train"
1750
1751
1752
1753
1754
1755
1756
        elif (
            args.dataset_path in MMStarDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMStarDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMStarDataset
            args.hf_split = "val"
            args.hf_subset = None
1757
        else:
1758
1759
1760
1761
1762
1763
1764
            supported_datasets = set(
                [
                    dataset_name
                    for cls in HuggingFaceDataset.__subclasses__()
                    for dataset_name in cls.SUPPORTED_DATASET_PATHS
                ]
            )
1765
1766
1767
1768
1769
            raise ValueError(
                f"Unsupported dataset path: {args.dataset_path}. "
                "Huggingface dataset only supports dataset_path"
                f" from one of following: {supported_datasets}. "
                "Please consider contributing if you would "
1770
1771
                "like to add support for additional dataset formats."
            )
1772

1773
1774
        if dataset_class.IS_MULTIMODAL and not (
            args.backend in ("openai-chat", "openai-audio")
1775
            or "embeddings-" in args.backend
1776
        ):
1777
1778
            # multi-modal benchmark is only available on OpenAI Chat
            # endpoint-type.
1779
1780
            raise ValueError(
                "Multi-modal content is only supported on 'openai-chat' and "
1781
1782
                "'openai-audio' backends."
            )
1783
1784
1785
1786
1787
        input_requests = dataset_class(
            dataset_path=args.dataset_path,
            dataset_subset=args.hf_subset,
            dataset_split=args.hf_split,
            random_seed=args.seed,
1788
            no_stream=args.no_stream,
1789
            hf_name=args.hf_name,
1790
            disable_shuffle=args.disable_shuffle,
1791
1792
1793
1794
        ).sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.hf_output_len,
1795
            request_id_prefix=args.request_id_prefix,
1796
            no_oversample=args.no_oversample,
1797
            skip_chat_template=args.skip_chat_template,
1798
            **hf_kwargs,
1799
1800
1801
1802
1803
        )

    else:
        # For datasets that follow a similar structure, use a mapping.
        dataset_mapping = {
1804
            "spec_bench": lambda: SpecBench(
1805
1806
1807
                dataset_path=args.dataset_path,
                category=args.spec_bench_category,
                disable_shuffle=args.disable_shuffle,
1808
            ).sample(
1809
1810
1811
1812
                num_requests=args.num_prompts,
                tokenizer=tokenizer,
                output_len=args.spec_bench_output_len,
                request_id_prefix=args.request_id_prefix,
1813
                no_oversample=args.no_oversample,
1814
            ),
1815
            "sharegpt": lambda: ShareGPTDataset(
1816
1817
1818
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1819
1820
1821
1822
1823
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
                request_id_prefix=args.request_id_prefix,
1824
                no_oversample=args.no_oversample,
1825
1826
            ),
            "burstgpt": lambda: BurstGPTDataset(
1827
1828
1829
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1830
1831
1832
1833
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                request_id_prefix=args.request_id_prefix,
1834
                no_oversample=args.no_oversample,
1835
1836
            ),
            "random": lambda: RandomDataset(
1837
1838
1839
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1840
            ).sample(
1841
1842
1843
1844
1845
1846
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                prefix_len=args.random_prefix_len,
                input_len=args.random_input_len,
                output_len=args.random_output_len,
                range_ratio=args.random_range_ratio,
1847
                request_id_prefix=args.request_id_prefix,
1848
                batchsize=args.random_batch_size,
1849
                no_oversample=args.no_oversample,
1850
            ),
1851
            "random-mm": lambda: RandomMultiModalDataset(
1852
1853
1854
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                prefix_len=args.random_prefix_len,
                range_ratio=args.random_range_ratio,
                input_len=args.random_input_len,
                output_len=args.random_output_len,
                base_items_per_request=args.random_mm_base_items_per_request,
                limit_mm_per_prompt=args.random_mm_limit_mm_per_prompt,
                num_mm_items_range_ratio=args.random_mm_num_mm_items_range_ratio,
                bucket_config=args.random_mm_bucket_config,
                request_id_prefix=args.request_id_prefix,
1867
                no_oversample=args.no_oversample,
1868
            ),
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
            "random-rerank": lambda: RandomDatasetForReranking(
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                input_len=args.random_input_len,
                range_ratio=args.random_range_ratio,
                request_id_prefix=args.request_id_prefix,
                batchsize=args.random_batch_size,
                is_reranker=not args.no_reranker,
            ),
1882
            "prefix_repetition": lambda: PrefixRepetitionRandomDataset(
1883
1884
1885
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1886
1887
1888
1889
1890
1891
1892
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                prefix_len=args.prefix_repetition_prefix_len,
                suffix_len=args.prefix_repetition_suffix_len,
                num_prefixes=args.prefix_repetition_num_prefixes,
                output_len=args.prefix_repetition_output_len,
1893
                request_id_prefix=args.request_id_prefix,
1894
                no_oversample=args.no_oversample,
1895
            ),
1896
1897
1898
        }

        try:
1899
            # Enforce endpoint compatibility for multimodal datasets.
1900
            if args.dataset_name == "random-mm" and args.backend not in ["openai-chat"]:
1901
1902
1903
1904
                raise ValueError(
                    "Multi-modal content (images) is only supported on "
                    "'openai-chat' backend."
                )
1905
1906
1907
1908
1909
1910
1911
            input_requests = dataset_mapping[args.dataset_name]()
        except KeyError as err:
            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err

    return input_requests


1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
# -----------------------------------------------------------------------------
# Custom Dataset Implementation
# -----------------------------------------------------------------------------


class CustomDataset(BenchmarkDataset):
    """
    Implements the Custom dataset.  Loads data from a JSONL file and generates
    sample requests based on conversation turns. E.g.,
    ```
    {"prompt": "What is the capital of India?"}
    {"prompt": "What is the capital of Iran?"}
    {"prompt": "What is the capital of China?"}
    ```
    """

    def __init__(self, **kwargs) -> None:
        super().__init__(**kwargs)
        self.load_data()

    def load_data(self) -> None:
        if self.dataset_path is None:
            raise ValueError("dataset_path must be provided for loading data.")

        # self.data will be a list of dictionaries
        # e.g., [{"prompt": "What is the capital of India?"}, ...]
        # This will be the standardized format which load_data()
        # has to convert into depending on the filetype of dataset_path.
        # sample() will assume this standardized format of self.data
        self.data = []

        # Load the JSONL file
        if self.dataset_path.endswith(".jsonl"):
1945
            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958

            # check if the JSONL file has a 'prompt' column
            if "prompt" not in jsonl_data.columns:
                raise ValueError("JSONL file must contain a 'prompt' column.")

            # Convert each row to a dictionary and append to self.data
            # This will convert the DataFrame to a list of dictionaries
            # where each dictionary corresponds to a row in the DataFrame.
            # This is the standardized format we want for self.data
            for _, row in jsonl_data.iterrows():
                self.data.append(row.to_dict())
        else:
            raise NotImplementedError(
1959
1960
                "Only JSONL format is supported for CustomDataset."
            )
1961
1962

        random.seed(self.random_seed)
1963
1964
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1965
1966
1967
1968
1969

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
1970
1971
1972
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1973
1974
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
1975
        request_id_prefix: str = "",
1976
        no_oversample: bool = False,
1977
1978
        **kwargs,
    ) -> list:
1979
1980
1981
1982
        # load all data if needed
        self.num_available_samples = len(self.data)
        if num_requests <= 0:
            num_requests = self.num_available_samples
1983
1984
1985
1986
1987
            logger.info(
                "num_requests is set to 0 or negative, "
                "so using all available samples: %d",
                num_requests,
            )
1988

1989
        sampled_requests = []
1990
        for i, item in enumerate(self.data):
1991
1992
1993
1994
1995
1996
1997
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["prompt"]

            # apply template
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
1998
                    [{"role": "user", "content": prompt}],
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
                    add_generation_prompt=True,
                    tokenize=False,
                )

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2009
                    request_id=request_id_prefix + str(i),
2010
2011
2012
2013
2014
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2015
2016
2017
2018

        return sampled_requests


2019
2020
2021
2022
2023
2024
2025
2026
# -----------------------------------------------------------------------------
# Spec Bench Dataset Implementation
# -----------------------------------------------------------------------------


class SpecBench(CustomDataset):
    """
    Implements the SpecBench dataset: https://github.com/hemingkx/Spec-Bench
2027
    Download the dataset using:
2028
    wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
2029
    """  # noqa: E501
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042

    def __init__(self, **kwargs) -> None:
        self.category = kwargs.pop("category", None)
        super().__init__(**kwargs)
        self.load_data()

    def load_data(self) -> None:
        if self.dataset_path is None:
            raise ValueError("dataset_path must be provided for loading data.")

        self.data = []

        # Load the JSONL file
2043
        jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2044
2045
2046
2047
2048
2049
2050

        # check if the JSONL file has a 'turns' column
        if "turns" not in jsonl_data.columns:
            raise ValueError("JSONL file must contain a 'turns' column.")

        for _, row in jsonl_data.iterrows():
            # sample only from a specific category if specified
2051
            if (not self.category) or (self.category == row["category"]):
2052
2053
2054
2055
                prompt = row["turns"][0]
                self.data.append({"prompt": prompt})

        random.seed(self.random_seed)
2056
2057
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2058
2059
2060
2061

    def sample(self, **kwargs) -> list:
        # leverage CustomDataset sample
        return super().sample(**kwargs)
2062
2063


2064
2065
2066
2067
# -----------------------------------------------------------------------------
# Sonnet Dataset Implementation
# -----------------------------------------------------------------------------

2068

2069
2070
2071
@deprecated(
    "SonnetDataset is deprecated and will be removed in a future version.",
)
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
class SonnetDataset(BenchmarkDataset):
    """
    Simplified implementation of the Sonnet dataset.  Loads poem lines from a
    text file and generates sample requests.  Default values here copied from
    `benchmark_serving.py` for the sonnet dataset.
    """

    DEFAULT_PREFIX_LEN = 200
    DEFAULT_INPUT_LEN = 550
    DEFAULT_OUTPUT_LEN = 150

    def __init__(
        self,
        **kwargs,
    ) -> None:
        super().__init__(**kwargs)
        self.load_data()

    def load_data(self) -> None:
        if not self.dataset_path:
            raise ValueError("dataset_path must be provided.")
        with open(self.dataset_path, encoding="utf-8") as f:
            self.data = f.readlines()

    def sample(
        self,
        tokenizer,
        num_requests: int,
        prefix_len: int = DEFAULT_PREFIX_LEN,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
        return_prompt_formatted: bool = False,
2104
        request_id_prefix: str = "",
2105
        no_oversample: bool = False,
2106
2107
2108
2109
        **kwargs,
    ) -> list:
        # Calculate average token length for a poem line.
        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
2110
        avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines)
2111
2112
2113
2114

        # Build the base prompt.
        base_prompt = "Pick as many lines as you can from these poem lines:\n"
        base_msg = [{"role": "user", "content": base_prompt}]
2115
2116
2117
        base_fmt = tokenizer.apply_chat_template(
            base_msg, add_generation_prompt=True, tokenize=False
        )
2118
2119
2120
2121
        base_offset = len(tokenizer(base_fmt).input_ids)
        if input_len <= base_offset:
            raise ValueError(
                f"'input_len' must be higher than the base prompt length "
2122
2123
                f"({base_offset})."
            )
2124
2125
2126
2127
2128
2129
2130

        # Determine how many poem lines to use.
        num_input_lines = round((input_len - base_offset) / avg_len)
        num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0)
        prefix_lines = self.data[:num_prefix_lines]

        samples = []
2131
        ind = 0
2132
        while len(samples) < num_requests:
2133
2134
2135
            extra_lines = random.choices(
                self.data, k=num_input_lines - num_prefix_lines
            )
2136
2137
2138
            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
            msg = [{"role": "user", "content": prompt}]
            prompt_formatted = tokenizer.apply_chat_template(
2139
2140
                msg, add_generation_prompt=True, tokenize=False
            )
2141
2142
2143
2144
            prompt_len = len(tokenizer(prompt_formatted).input_ids)
            if prompt_len <= input_len:
                samples.append(
                    SampleRequest(
2145
                        prompt=prompt_formatted if return_prompt_formatted else prompt,
2146
2147
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
2148
2149
2150
                        request_id=request_id_prefix + str(ind),
                    )
                )
2151
                ind += 1
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
        return samples


# -----------------------------------------------------------------------------
# BurstGPT Dataset Implementation
# -----------------------------------------------------------------------------


class BurstGPTDataset(BenchmarkDataset):
    """
    Implements the BurstGPT dataset.  Loads data from a CSV file and generates
    sample requests based on synthetic prompt generation. Only rows with Model
    "GPT-4" and positive response tokens are used.
    """

    def __init__(self, **kwargs) -> None:
        super().__init__(**kwargs)
        self.load_data()

2171
2172
2173
    def load_data(
        self,
    ):
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
        if self.dataset_path is None:
            raise ValueError("dataset_path must be provided for loading data.")

        df = pd.read_csv(self.dataset_path)
        # Filter to keep only GPT-4 rows.
        gpt4_df = df[df["Model"] == "GPT-4"]
        # Remove failed requests (where Response tokens is 0 or less).
        gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
        # Sample the desired number of rows.
        self.data = gpt4_df

    def _sample_loaded_data(self, num_requests: int) -> list:
        if num_requests <= len(self.data):
2187
            data = self.data.sample(n=num_requests, random_state=self.random_seed)
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
        else:
            data = self.data.sample(
                n=num_requests,
                random_state=self.random_seed,
                replace=True,
            )
        # Convert the dataframe to a list of lists.
        return data.values.tolist()

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2201
2202
        max_loras: int | None = None,
        lora_path: str | None = None,
2203
        request_id_prefix: str = "",
2204
        no_oversample: bool = False,
2205
2206
2207
2208
2209
2210
2211
        **kwargs,
    ) -> list[SampleRequest]:
        samples = []
        data = self._sample_loaded_data(num_requests=num_requests)
        for i in range(num_requests):
            input_len = int(data[i][2])
            output_len = int(data[i][3])
2212
            lora_req = self.get_random_lora_request(
2213
2214
                max_loras=max_loras, lora_path=lora_path
            )
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
            vocab_size = tokenizer.vocab_size
            # Generate a synthetic prompt: a list of token IDs computed as (i +
            # j) modulo vocab_size.
            token_ids = [(i + j) % vocab_size for j in range(input_len)]
            prompt = tokenizer.decode(token_ids)
            samples.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=input_len,
                    expected_output_len=output_len,
                    lora_request=lora_req,
2226
                    request_id=request_id_prefix + str(i),
2227
2228
                )
            )
2229
2230
2231
2232
2233
2234
2235
2236
2237
        return samples


# -----------------------------------------------------------------------------
# HuggingFace Dataset Base Implementation
# -----------------------------------------------------------------------------
class HuggingFaceDataset(BenchmarkDataset):
    """Base class for datasets hosted on HuggingFace."""

2238
    SUPPORTED_DATASET_PATHS: set[str] | dict[str, Callable] = set()
2239
2240
2241
2242
2243

    def __init__(
        self,
        dataset_path: str,
        dataset_split: str,
2244
        no_stream: bool = False,
2245
2246
        dataset_subset: str | None = None,
        hf_name: str | None = None,
2247
2248
2249
2250
2251
2252
        **kwargs,
    ) -> None:
        super().__init__(dataset_path=dataset_path, **kwargs)

        self.dataset_split = dataset_split
        self.dataset_subset = dataset_subset
2253
        self.load_stream = not no_stream
2254
        self.hf_name = hf_name or dataset_path
2255
2256
2257
2258
2259
2260
2261
2262
        self.load_data()

    def load_data(self) -> None:
        """Load data from HuggingFace datasets."""
        self.data = load_dataset(
            self.dataset_path,
            name=self.dataset_subset,
            split=self.dataset_split,
2263
            streaming=self.load_stream,
2264
        )
2265
2266
        if not getattr(self, "disable_shuffle", False):
            self.data = self.data.shuffle(seed=self.random_seed)
2267
2268
2269
2270
2271
2272
2273
2274
2275


# -----------------------------------------------------------------------------
# Conversation Dataset Implementation
# -----------------------------------------------------------------------------


class ConversationDataset(HuggingFaceDataset):
    """Dataset for conversation data with multimodal support."""
2276

2277
    SUPPORTED_DATASET_PATHS = {
2278
2279
        "lmms-lab/LLaVA-OneVision-Data",
        "Aeala/ShareGPT_Vicuna_unfiltered",
2280
    }
2281
    IS_MULTIMODAL = True
2282

2283
2284
2285
2286
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2287
        output_len: int | None = None,
2288
2289
2290
2291
2292
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2293
        # Filter examples with at least 2 conversations
2294
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
2295
        sampled_requests = []
2296
        ind = 0
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
        dynamic_output = output_len is None

        for item in filtered_data:
            if len(sampled_requests) >= num_requests:
                break
            conv = item["conversations"]
            prompt, completion = conv[0]["value"], conv[1]["value"]

            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
            completion_len = len(completion_ids)
            output_len = completion_len if dynamic_output else output_len
            assert isinstance(output_len, int) and output_len > 0
2311
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
2312
                continue
2313
            mm_content = process_image(item["image"]) if "image" in item else None
2314
2315
2316
2317
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len and output len
2318
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2319
2320
2321
2322
2323
2324
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2325
                    request_id=request_id_prefix + str(ind),
2326
2327
                )
            )
2328
            ind += 1
2329
2330
2331
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
        return sampled_requests


# -----------------------------------------------------------------------------
# Vision Arena Dataset Implementation
# -----------------------------------------------------------------------------


class VisionArenaDataset(HuggingFaceDataset):
    """
    Vision Arena Dataset.
    """

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2347
2348
        "lmarena-ai/VisionArena-Chat": lambda x: x["conversation"][0][0]["content"],
        "lmarena-ai/vision-arena-bench-v0.1": lambda x: x["turns"][0][0]["content"],
2349
    }
2350
    IS_MULTIMODAL = True
2351
2352
2353
2354
2355

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2356
        output_len: int | None = None,
2357
        enable_multimodal_chat: bool = False,
2358
        request_id_prefix: str = "",
2359
        no_oversample: bool = False,
2360
2361
        **kwargs,
    ) -> list:
2362
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2363
        sampled_requests = []
2364
        for i, item in enumerate(self.data):
2365
2366
            if len(sampled_requests) >= num_requests:
                break
2367
            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
2368
            if parser_fn is None:
2369
                raise ValueError(f"Unsupported dataset path: {self.hf_name}")
2370
2371
2372
2373
2374
2375
2376
            prompt = parser_fn(item)
            mm_content = process_image(item["images"][0])
            prompt_len = len(tokenizer(prompt).input_ids)
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len
2377
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2378
2379
2380
2381
2382
2383
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2384
                    request_id=request_id_prefix + str(i),
2385
2386
2387
2388
2389
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2390
2391
2392
        return sampled_requests


2393
2394
2395
2396
2397
2398
2399
2400
class MMVUDataset(HuggingFaceDataset):
    """
    MMVU Dataset.
    https://huggingface.co/datasets/yale-nlp/MMVU
    """

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2401
2402
2403
        "yale-nlp/MMVU": lambda x: x["question"]
        + " "
        + (" ".join(f"{k}.{v}" for k, v in x["choices"].items())),
2404
2405
2406
2407
2408
2409
    }

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2410
        output_len: int | None = None,
2411
2412
2413
2414
2415
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2416
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
        sampled_requests = []
        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
            if parser_fn is None:
                raise ValueError(f"Unsupported dataset path: {self.hf_name}")
            prompt = parser_fn(item)
            mm_content = process_video(item["video"])
            prompt_len = len(tokenizer(prompt).input_ids)
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len
2431
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2432
2433
2434
2435
2436
2437
2438
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
                    request_id=request_id_prefix + str(i),
2439
2440
2441
2442
2443
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2444
2445
2446
        return sampled_requests


2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
# -----------------------------------------------------------------------------
# Instruct Coder Dataset Implementation
# -----------------------------------------------------------------------------


class InstructCoderDataset(HuggingFaceDataset):
    """
    InstructCoder Dataset.
    https://huggingface.co/datasets/likaixin/InstructCoder

    InstructCoder is the dataset designed for general code editing.  It consists
    of 114,239 instruction-input-output triplets, and covers multiple distinct
    code editing scenario.
    """

    DEFAULT_OUTPUT_LEN = 200  # this is the average default output length
    SUPPORTED_DATASET_PATHS = {
        "likaixin/InstructCoder",
    }

2467
2468
2469
2470
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2471
        output_len: int | None = None,
2472
2473
2474
2475
2476
2477
2478
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2479
        sampled_requests = []
2480
        for i, item in enumerate(self.data):
2481
2482
            if len(sampled_requests) >= num_requests:
                break
2483
2484
2485
2486
            prompt = (
                f"{item['input']}\n\n{item['instruction']} Just output "
                "the code, do not include any explanation."
            )
2487
2488

            # apply template
2489
2490
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2491
                    [{"role": "user", "content": prompt}],
2492
2493
2494
                    add_generation_prompt=True,
                    tokenize=False,
                )
2495

2496
2497
2498
2499
2500
2501
            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2502
                    request_id=request_id_prefix + str(i),
2503
2504
2505
2506
2507
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2508
2509
2510
        return sampled_requests


2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
# -----------------------------------------------------------------------------
# MT-Bench Dataset Implementation
# -----------------------------------------------------------------------------


class MTBenchDataset(HuggingFaceDataset):
    """
    MT-Bench Dataset.
    https://huggingface.co/datasets/philschmid/mt-bench

    We create a single turn dataset for MT-Bench.
    This is similar to Spec decoding benchmark setup in vLLM
    https://github.com/vllm-project/vllm/blob/9d98ab5ec/examples/offline_inference/eagle.py#L14-L18
    """  # noqa: E501

    DEFAULT_OUTPUT_LEN = 256  # avg len used in SD bench in vLLM
    SUPPORTED_DATASET_PATHS = {
        "philschmid/mt-bench",
    }

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2535
        output_len: int | None = None,
2536
        enable_multimodal_chat: bool = False,
2537
        skip_chat_template: bool = False,
2538
        request_id_prefix: str = "",
2539
        no_oversample: bool = False,
2540
2541
        **kwargs,
    ) -> list:
2542
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2543
2544
        sampled_requests = []

2545
        for i, item in enumerate(self.data):
2546
2547
2548
2549
2550
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["turns"][0]

            # apply template
2551
2552
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2553
                    [{"role": "user", "content": prompt}],
2554
2555
2556
                    add_generation_prompt=True,
                    tokenize=False,
                )
2557
2558
2559
2560
2561
2562
2563

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2564
                    request_id=request_id_prefix + str(i),
2565
2566
2567
2568
2569
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2570
2571
2572
        return sampled_requests


2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
# -----------------------------------------------------------------------------
# Blazedit Dataset Implementation
# -----------------------------------------------------------------------------


class BlazeditDataset(HuggingFaceDataset):
    """
    Blazedit Dataset.
    https://github.com/ise-uiuc/blazedit

    5k char version: vdaita/edit_5k_char
    10k char version: vdaita/edit_10k_char
    """  # noqa: E501

    # 5k char version will have output as ~5k chars
    # 10k char version will have output as ~10k chars
    # Assuming 3 char per token, 10k chars will be 3333 tokens
    # We set default to 4000 to be safe
    DEFAULT_OUTPUT_LEN = 4000
    SUPPORTED_DATASET_PATHS = {
        "vdaita/edit_5k_char",
        "vdaita/edit_10k_char",
    }

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2601
        output_len: int | None = None,
2602
        skip_chat_template: bool = False,
2603
        request_id_prefix: str = "",
2604
        no_oversample: bool = False,
2605
2606
2607
2608
        min_distance: float = 0.0,
        max_distance: float = 1.0,
        **kwargs,
    ) -> list:
2609
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
        sampled_requests = []

        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            code = item["code"]
            change_request = item["change_request"]
            norm_distance = item["norm_distance"]

            # compare the levenshtein distance normalized by code length
            if norm_distance < min_distance or norm_distance > max_distance:
                continue
2622
2623

            # template copied from
2624
            # https://github.com/ise-uiuc/blazedit/blob/7765137e656fd62de877422d2e4cf8de51228054/dataset/create_refined_dataset.py#L94-L105 # noqa: E501
2625
            prompt = f"""Given a code file, please apply the change requests and generate the new file.
2626
2627
2628
2629
2630
2631
2632
2633
2634

Original file:
```python
{code}
```

Change request:
{change_request}

2635
Please generate the new code file in the "New file" section below."""  # noqa: E501
2636
2637

            # apply template
2638
2639
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2640
                    [{"role": "user", "content": prompt}],
2641
2642
2643
                    add_generation_prompt=True,
                    tokenize=False,
                )
2644
2645
2646
2647
2648
2649
2650
2651
2652

            prompt_len = len(tokenizer(prompt).input_ids)

            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    request_id=request_id_prefix + str(i),
2653
2654
2655
2656
2657
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2658

2659
2660
2661
        return sampled_requests


2662
2663
2664
2665
2666
2667
2668
2669
2670
# -----------------------------------------------------------------------------
# AIMO Dataset Implementation
# -----------------------------------------------------------------------------


class AIMODataset(HuggingFaceDataset):
    """
    Dataset class for processing a AIMO dataset with reasoning questions.
    """
2671

2672
    SUPPORTED_DATASET_PATHS = {
2673
2674
2675
        "AI-MO/aimo-validation-aime",
        "AI-MO/NuminaMath-1.5",
        "AI-MO/NuminaMath-CoT",
2676
2677
    }

2678
2679
2680
2681
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2682
        output_len: int | None = None,
2683
2684
2685
2686
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2687
        sampled_requests = []
2688
        ind = 0
2689
2690
2691
2692
2693
        dynamic_output = output_len is None

        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
2694
            prompt, completion = item["problem"], item["solution"]
2695
2696
2697
2698
2699
2700
2701

            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
            completion_len = len(completion_ids)
            output_len = completion_len if dynamic_output else output_len
            assert isinstance(output_len, int) and output_len > 0
2702
2703
2704
            if dynamic_output and not is_valid_sequence(
                prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000
            ):
2705
2706
2707
2708
2709
2710
2711
                continue
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=None,
2712
                    request_id=request_id_prefix + str(ind),
2713
2714
                )
            )
2715
            ind += 1
2716
2717
2718
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2719
        return sampled_requests
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739


# -----------------------------------------------------------------------------
# Next Edit Prediction Dataset Implementation
# -----------------------------------------------------------------------------


zeta_prompt = """### Instruction:
You are a code completion assistant and your task is to analyze user edits and then rewrite an excerpt that the user provides, suggesting the appropriate edits within the excerpt, taking into account the cursor location.

### User Edits:

{}

### User Excerpt:

{}

### Response:

2740
"""  # noqa: E501
2741
2742
2743


def _format_zeta_prompt(
2744
2745
    sample: dict, original_start_marker: str = "<|editable_region_start|>"
) -> dict:
2746
    """Format the zeta prompt for the Next Edit Prediction (NEP) dataset.
2747
2748
2749

    This function formats examples from the NEP dataset
    into prompts and expected outputs. It could be
2750
    further extended to support more NEP datasets.
2751

2752
    Args:
2753
        sample: The dataset sample containing events,
2754
            inputs, and outputs.
2755
2756
        original_start_marker: The marker indicating the
            start of the editable region. Defaults to
2757
            "<|editable_region_start|>".
2758

2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
    Returns:
        A dictionary with the formatted prompts and expected outputs.
    """
    events = sample["events"]
    input = sample["input"]
    output = sample["output"]
    prompt = zeta_prompt.format(events, input)

    # following the original implementation, extract the focused region
    # from the raw output
    output_start_index = output.find(original_start_marker)
    output_focused_region = output[output_start_index:]
    expected_output = output_focused_region

    return {"prompt": prompt, "expected_output": expected_output}


class NextEditPredictionDataset(HuggingFaceDataset):
    """
    Dataset class for processing a Next Edit Prediction dataset.
    """

    SUPPORTED_DATASET_PATHS = {
        "zed-industries/zeta",
    }
    MAPPING_PROMPT_FUNCS = {
        "zed-industries/zeta": _format_zeta_prompt,
    }

2788
2789
2790
2791
2792
2793
2794
2795
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ):
2796
        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.hf_name)
2797
        if formatting_prompt_func is None:
2798
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")
2799
        samples = []
2800
        for i, sample in enumerate(self.data):
2801
2802
2803
2804
2805
2806
            sample = formatting_prompt_func(sample)
            samples.append(
                SampleRequest(
                    prompt=sample["prompt"],
                    prompt_len=len(tokenizer(sample["prompt"]).input_ids),
                    expected_output_len=len(
2807
2808
                        tokenizer(sample["expected_output"]).input_ids
                    ),
2809
                    request_id=request_id_prefix + str(i),
2810
2811
                )
            )
2812
2813
            if len(samples) >= num_requests:
                break
2814
2815
2816
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
2817
        return samples
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856


# -----------------------------------------------------------------------------
# ASR Dataset Implementation
# -----------------------------------------------------------------------------


class ASRDataset(HuggingFaceDataset):
    """
    Dataset class for processing a ASR dataset for transcription.
    Tested on the following set:

    +----------------+----------------------------------------+--------------------------+-----------------------------+
    | Dataset        | Domain                                 | Speaking Style           | hf-subset                   |
    +----------------+----------------------------------------+--------------------------+-----------------------------+
    | TED-LIUM       | TED talks                              | Oratory                  | release1, release2, release3|
    |                |                                        |                          | release3-speaker-adaptation |
    | VoxPopuli      | European Parliament                    | Oratory                  | en, de, it, fr,  ...        |
    | LibriSpeech    | Audiobook                              | Narrated                 | "LIUM/tedlium"              |
    | GigaSpeech     | Audiobook, podcast, YouTube            | Narrated, spontaneous    | xs, s, m, l, xl, dev, test  |
    | SPGISpeech     | Financial meetings                     | Oratory, spontaneous     | S, M, L, dev, test          |
    | AMI            | Meetings                               | Spontaneous              | ihm, sdm                    |
    +----------------+----------------------------------------+--------------------------+-----------------------------+

    """  # noqa: E501

    SUPPORTED_DATASET_PATHS = {
        "openslr/librispeech_asr",
        "facebook/voxpopuli",
        "LIUM/tedlium",
        "edinburghcstr/ami",
        "speechcolab/gigaspeech",
        "kensho/spgispeech",
    }

    DEFAULT_OUTPUT_LEN = 128
    IS_MULTIMODAL = True

    # TODO Whisper-specific. Abstract interface when more models are supported.
2857
    TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>"
2858
2859
2860
2861
2862
2863
    skip_long_audios: bool = True

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2864
        output_len: int | None = None,
2865
        request_id_prefix: str = "",
2866
        no_oversample: bool = False,
2867
2868
        **kwargs,
    ) -> list:
2869
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2870
2871
2872
        prompt = ASRDataset.TRANSCRIPTION_PREAMBLE
        prompt_len = len(tokenizer(prompt).input_ids)
        sampled_requests = []
2873
        ind = 0
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
        skipped = 0
        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
            audio = item["audio"]
            y, sr = audio["array"], audio["sampling_rate"]
            duration_s = librosa.get_duration(y=y, sr=sr)
            # Whisper max supported duration
            if self.skip_long_audios and duration_s > 30:
                skipped += 1
                continue

            mm_content = {"audio": (y, sr)}
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2893
                    request_id=request_id_prefix + str(ind),
2894
2895
                )
            )
2896
            ind += 1
2897
2898
2899
2900
2901
2902
2903
        if skipped:
            logger.warning(
                "%d samples discarded from dataset due to"
                " their length being greater than"
                " what Whisper supports.",
                skipped,
            )
2904
2905
2906
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2907
        return sampled_requests
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941


# -----------------------------------------------------------------------------
# MLPerf Dataset Implementation
# -----------------------------------------------------------------------------


class MLPerfDataset(HuggingFaceDataset):
    """
    MLPerf Inference Dataset.

    Dataset on HF:
    https://huggingface.co/datasets/mgoin/mlperf-inference-llama2-data
    https://huggingface.co/datasets/mgoin/mlperf-inference-llama3.1-data

    Each record contains:
      - "system_prompt": system role instruction.
      - "question": user question.
      - "output": reference answer.

    We combine the system prompt and question into a chat-formatted prompt
    (using the tokenizer's chat template) and set the expected output length to
    the tokenized length of the provided reference answer.
    """

    SUPPORTED_DATASET_PATHS = {
        "mgoin/mlperf-inference-llama2-data",
        "mgoin/mlperf-inference-llama3.1-data",
    }

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2942
        output_len: int | None = None,
2943
        request_id_prefix: str = "",
2944
        no_oversample: bool = False,
2945
2946
2947
2948
2949
        **kwargs,
    ) -> list[SampleRequest]:
        # Force dynamic output length based on reference completion.
        dynamic_output = output_len is None
        sampled_requests: list[SampleRequest] = []
2950
        ind = 0
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984

        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break

            system_prompt = item["system_prompt"]
            question = item["question"]
            reference_answer = item["output"]

            # Build chat-style prompt using tokenizer template, if available.
            messages = [
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": question},
            ]
            prompt_formatted = tokenizer.apply_chat_template(
                messages, add_generation_prompt=True, tokenize=False
            )
            prompt_len = len(tokenizer(prompt_formatted).input_ids)

            # Determine output length from reference answer tokens.
            ref_out_len = len(
                tokenizer(reference_answer, add_special_tokens=False).input_ids
            )
            expected_output_len = ref_out_len if dynamic_output else output_len

            # Validate sequence lengths.
            if not is_valid_sequence(prompt_len, expected_output_len):
                continue

            sampled_requests.append(
                SampleRequest(
                    prompt=prompt_formatted,
                    prompt_len=prompt_len,
                    expected_output_len=expected_output_len,
2985
                    request_id=request_id_prefix + str(ind),
2986
2987
                )
            )
2988
            ind += 1
2989

2990
2991
2992
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2993
        return sampled_requests
2994
2995
2996
2997
2998
2999
3000
3001


# -----------------------------------------------------------------------------
# Prefix Repetition Dataset Implementation
# -----------------------------------------------------------------------------


class PrefixRepetitionRandomDataset(BenchmarkDataset):
3002
    # Default values copied from benchmark_serving.py for the repeated prefix
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
    # dataset.
    DEFAULT_PREFIX_LEN = 256
    DEFAULT_SUFFIX_LEN = 256
    DEFAULT_NUM_PREFIXES = 10
    DEFAULT_OUTPUT_LEN = 128

    def __init__(
        self,
        **kwargs,
    ) -> None:
        super().__init__(**kwargs)
        random.seed(self.random_seed)
        np.random.seed(self.random_seed)

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        prefix_len: int = DEFAULT_PREFIX_LEN,
        suffix_len: int = DEFAULT_SUFFIX_LEN,
        num_prefixes: int = DEFAULT_NUM_PREFIXES,
        output_len: int = DEFAULT_OUTPUT_LEN,
3025
        request_id_prefix: str = "",
3026
        no_oversample: bool = False,
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
        **kwargs,
    ) -> list[SampleRequest]:
        vocab_size = tokenizer.vocab_size
        prompts_per_prefix = num_requests // num_prefixes
        if prompts_per_prefix == 0:
            raise ValueError(
                f"num_requests ({num_requests}) must be greater than or equal "
                f"to num_prefixes ({num_prefixes})"
            )

        def _generate_exact_length_tokens(target_length: int) -> list[int]:
            """Generate tokens that decode and re-encode to exactly
            target_length."""
            # Generate random tokens
3041
            tokens = np.random.randint(0, vocab_size, size=target_length).tolist()
3042

3043
            _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(  # noqa: E501
3044
3045
3046
3047
3048
3049
                tokenizer=tokenizer,
                token_sequence=tokens,
                target_token_len=target_length,
                add_special_tokens=False,
            )
            return adjusted_tokens, token_mismatch
3050
3051

        requests = []
3052
        token_mismatch_total = 0
3053
        for _ in range(num_prefixes):
3054
3055
            prefix_tokens, prefix_mismatch = _generate_exact_length_tokens(prefix_len)
            token_mismatch_total += prefix_mismatch
3056
3057

            for _ in range(prompts_per_prefix):
3058
                suffix_tokens, suffix_mismatch = _generate_exact_length_tokens(
3059
                    suffix_len
3060
                )
3061
                token_mismatch_total += suffix_mismatch
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
                combined_tokens = prefix_tokens + suffix_tokens
                prompt = tokenizer.decode(combined_tokens)
                prompt_len = len(combined_tokens)
                requests.append(
                    SampleRequest(
                        prompt=prompt,
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
                    )
                )

3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
        if token_mismatch_total != 0:
            sign = "more" if token_mismatch_total > 0 else "fewer"
            logger.warning(
                "Across all generated prompts, there were %d %s tokens "
                "than expected after decoding and re-encoding. This is "
                "expected due to the imperfect nature of the sampling "
                "procedure.",
                abs(token_mismatch_total),
                sign,
            )
3083
3084
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(requests)
3085
        return requests
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097


# -----------------------------------------------------------------------------
# MMStar Dataset Implementation
# -----------------------------------------------------------------------------


class MMStarDataset(HuggingFaceDataset):
    """
    Lin-Chen/MMStar: https://huggingface.co/datasets/Lin-Chen/MMStar
    refer to: https://github.com/sgl-project/SpecForge/pull/106
    """
3098

3099
3100
3101
3102
3103
3104
3105
3106
    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {"Lin-Chen/MMStar"}
    IS_MULTIMODAL = True

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
3107
        output_len: int | None = None,
3108
3109
3110
3111
3112
3113
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list[SampleRequest]:
        # If --hf-output-len is not set, use the default output length.
3114
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
        sampled_requests: list[SampleRequest] = []

        for ind, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            # Split the question text from options
            # (keep only the part before "Options:").
            full_q: str = item.get("question", "")
            question_text = full_q.split("Options:", 1)[0].strip()

            # Multimodal image content.
            mm_content = process_image(item["image"])

            # Compute prompt token length (note: this is plain text length
            # if enable_multimodal_chat is False).
            prompt_len = len(tokenizer(question_text).input_ids)

            if enable_multimodal_chat:
                # If multimodal content should be embedded in the chat message,
                # convert to [{"role":"user","content":[...]}]
                prompt = self.apply_multimodal_chat_transformation(
                    question_text, mm_content
                )
                mm_for_request = None  # Already embedded in chat content.
            else:
                # Default: prompt is plain text,
                # image is in mm_content for the bench to assemble.
                prompt = question_text
                mm_for_request = mm_content

            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_for_request,
                    request_id=request_id_prefix + str(ind),
                )
            )

        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
        return sampled_requests