datasets.py 126 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

import numpy as np
from PIL import Image
35
from typing_extensions import deprecated
36
37
38
39

from vllm.lora.request import LoRARequest
from vllm.lora.utils import get_adapter_absolute_path
from vllm.multimodal import MultiModalDataDict
40
from vllm.multimodal.image import convert_image_mode
41
from vllm.tokenizers import TokenizerLike
42
from vllm.utils.argparse_utils import FlexibleArgumentParser
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
62

logger = logging.getLogger(__name__)

63
64
DEFAULT_NUM_PROMPTS = 1000

65
66
67
68
69
70
71
72
73
74
75
# -----------------------------------------------------------------------------
# Data Classes
# -----------------------------------------------------------------------------


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

76
    prompt: str | list[str]
77
78
    prompt_len: int
    expected_output_len: int
79
80
81
    multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None
    lora_request: LoRARequest | None = None
    request_id: str | None = None
82
83
84
85
86
87
88
89
90


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


class BenchmarkDataset(ABC):
    DEFAULT_SEED = 0
91
    IS_MULTIMODAL = False
92
93
94

    def __init__(
        self,
95
        dataset_path: str | None = None,
96
        random_seed: int = DEFAULT_SEED,
97
98
        disable_shuffle: bool = False,
        **kwargs,
99
100
101
    ) -> None:
        """
        Initialize the BenchmarkDataset with an optional dataset path and random
102
103
        seed.

104
105
        Args:
            dataset_path (Optional[str]): Path to the dataset. If None, it
106
                indicates that a default or random dataset might be used.
107
            random_seed (int): Seed value for reproducible shuffling or
108
                sampling. Defaults to DEFAULT_SEED.
109
110
111
112
        """
        self.dataset_path = dataset_path
        # Set the random seed, ensuring that a None value is replaced with the
        # default seed.
113
        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
114
        self.disable_shuffle = disable_shuffle
115
116
117
        self.data = None

    def apply_multimodal_chat_transformation(
118
119
        self,
        prompt: str,
120
        mm_content: MultiModalDataDict | dict | list[dict] | None = None,
121
    ) -> list[dict]:
122
123
124
125
126
127
128
        """
        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:
129
130
131
132
133
            if isinstance(mm_content, list):
                content.extend(cast(list[dict[str, Any]], mm_content))
            elif isinstance(mm_content, dict):
                content.append(mm_content)
            else:
134
                raise TypeError(
135
                    f"Could not process multimodal content of type: {type(mm_content)}"
136
                )
137
138
139
140
141
142
143
144
145
146
147
148
149
        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
150
        raise NotImplementedError("load_data must be implemented in subclasses.")
151
152
153

    def get_random_lora_request(
        self,
154
155
156
        max_loras: int | None = None,
        lora_path: str | None = None,
    ) -> LoRARequest | None:
157
        """
158
        Optionally select a random LoRA request.
159
160

        This method is used when LoRA parameters are provided.  It randomly
161
        selects a LoRA based on max_loras.
162
163

        Args:
164
165
166
167
            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.
168
169

        Returns:
170
171
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
172
173
        """
        if max_loras is None or lora_path is None:
174
            return None
175
176
177
178
179
180
181
182

        # 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),
        )
183
        return lora_request
184
185

    @abstractmethod
186
187
    def sample(
        self,
188
        tokenizer: TokenizerLike,
189
190
191
192
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
    ) -> list[SampleRequest]:
193
194
195
196
197
198
199
        """
        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:
200
            tokenizer (TokenizerLike): The tokenizer to be used
201
                for processing the dataset's text.
202
            num_requests (int): The number of sample requests to generate.
203
            request_id_prefix (str): The prefix of request_id.
204
205
206
207
208
209
210

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

211
212
213
214
215
    def maybe_oversample_requests(
        self,
        requests: list[SampleRequest],
        num_requests: int,
        request_id_prefix: str = "",
216
        no_oversample: bool = False,
217
    ) -> None:
218
219
220
221
222
223
        """
        Oversamples the list of requests if its size is less than the desired
        number.

        Args:
            requests (List[SampleRequest]): The current list of sampled
224
225
                requests.
            num_requests (int): The target number of requests.
226
227
            request_id_prefix (str): The prefix applied to generated request
                identifiers.
228

229
        """
230
        if no_oversample:
231
            logger.info("Skipping oversampling. Total samples: %d.", len(requests))
232
233
            return

234
235
        if len(requests) < num_requests:
            random.seed(self.random_seed)
236
237
238
239
            needed = num_requests - len(requests)
            additional = []
            for i in range(needed):
                req = deepcopy(random.choice(requests))
240
                req.request_id = request_id_prefix + str(len(requests) + i)
241
                additional.append(req)
242
            requests.extend(additional)
243
            logger.info("Oversampled requests to reach %d total samples.", num_requests)
244

245
246
        ids = [req.request_id for req in requests]
        if len(ids) != len(set(ids)):
247
248
249
250
251
            raise ValueError(
                "Duplicate request_id found in the sampled "
                "requests. Please ensure that each request_id "
                "is unique."
            )
252

253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275

# -----------------------------------------------------------------------------
# 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
276
    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
277
278
279
280
    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
281
282
283
    return not (
        prompt_too_short or output_too_short or prompt_too_long or combined_too_long
    )
284
285
286
287
288
289
290
291


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


# Global cache for LoRA tokenizers.
292
lora_tokenizer_cache: dict[int, TokenizerLike] = {}
293
294
295
296
297
298


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

299
    Supports the following input types:
300
301
302
303
304
305
306
307

    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.

308
309
310
311
312
    3. String input: - Treats the string as a URL, local file path, or base64
       encoded data.  - If string starts with "data:image/", treats as base64.
       - If string starts with "http://", "https://", or "file://", treats as URL.
       - Otherwise treats as local file path and prepends "file://".
       - Returns a dictionary with the image URL or base64 data.
313
314
315
316

    Raises:
        ValueError: If the input is not a supported type.
    """
317
318
    if isinstance(image, dict) and "bytes" in image:
        image = Image.open(BytesIO(image["bytes"]))
319
    if isinstance(image, Image.Image):
320
        image = convert_image_mode(image, "RGB")
321
322
        with io.BytesIO() as image_data:
            image.save(image_data, format="JPEG")
323
            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
324
325
        return {
            "type": "image_url",
326
            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
327
328
329
        }

    if isinstance(image, str):
330
331
        image_url = (
            image
332
            if image.startswith(("http://", "https://", "file://", "data:image/"))
333
334
            else f"file://{image}"
        )
335
336
        return {"type": "image_url", "image_url": {"url": image_url}}

337
    raise ValueError(
338
339
        f"Invalid image input {image}. Must be a PIL.Image.Image, "
        "str (URL, file path, or base64 data URL), or dictionary with raw image bytes."
340
    )
341
342


343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
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.
    """
359
360
    if isinstance(video, dict) and "bytes" in video:
        video_bytes = video["bytes"]
361
362
363
        video_base64 = base64.b64encode(video_bytes).decode("utf-8")
        return {
            "type": "video_url",
364
            "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
365
366
367
        }

    if isinstance(video, str):
368
369
370
371
372
        video_url = (
            video
            if video.startswith(("http://", "https://", "file://"))
            else f"file://{video}"
        )
373
374
375
376
377
378
        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
    )

379
380

def gen_prompt_decode_to_target_len(
381
    tokenizer: TokenizerLike,
382
383
384
385
    token_sequence: list[int],
    target_token_len: int,
    max_retry: int = 10,
    add_special_tokens: bool = False,
386
    rng: np.random.Generator | None = None,
387
) -> tuple[str, list[int], int]:
388
389
390
391
    """
    Ensure decoded-then-encoded prompt length matches the target token length.

    This function decodes an initial token sequence to text and re-encodes it
392
393
    , iteratively adjusting the token sequence length to match a target.
    This is necessary because some tokenizers do not guarantee a 1:1 mapping
394
395
396
397
398
    between consecutive tokens and the decoded-then-encoded sequence length.
    For example, for GPT2Tokenizer:
    [6880, 6881] -> ['Ġcalls', 'here'] ->
    [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']

399
400
401
    Returns a tuple of the final prompt string, the adjusted token sequence,
    and the token mismatch (final_len - target_token_len) if the retry budget
    is exhausted.
402
403
404
405
406
    """
    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

    def sample(
        self,
471
        tokenizer: TokenizerLike,
472
        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
        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
508
        prefix_token_ids = self.get_prefix(tokenizer, 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
        self,
563
        tokenizer: TokenizerLike,
564
565
        allowed_tokens: np.ndarray,
        prefix_len: int,
566
567
568
569
    ) -> list[int]:
        """
        Get the prefix for the dataset.
        """
570
571
572
573
574
575
576
577
578
579
580
581
        if prefix_len <= 0:
            return []

        prefix_tokens = allowed_tokens[
            self._rng.integers(0, len(allowed_tokens), size=prefix_len)
        ].tolist()
        _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(
            tokenizer=tokenizer,
            token_sequence=prefix_tokens,
            target_token_len=prefix_len,
            add_special_tokens=False,
            rng=self._rng,
582
        )
583
584
585
586
587
588
589
590
591
592
        if token_mismatch != 0:
            sign = "more" if token_mismatch > 0 else "fewer"
            logger.warning(
                "Prefix tokenization produced %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),
                sign,
            )
        return adjusted_tokens
593

594
595
596
597
598
599
    def get_sampling_params(
        self,
        num_requests: int,
        range_ratio: float,
        input_len: int,
        output_len: int,
600
        tokenizer: TokenizerLike,
601
602
603
604
605
606
607
608
609
610
611
612
613
614
    ) -> 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))
615
616
        # Ensure the lower bound for output length is at least 1 to
        # prevent sampling 0 tokens.
617
        output_low = max(output_low, 1)
618
        output_high = max(output_high, 1)
619
620
621

        if input_low > input_high:
            raise ValueError(
622
                f"Invalid input sampling interval: low={input_low} > high={input_high}"
623
624
625
626
627
628
            )
        if output_low > output_high:
            raise ValueError(
                "Invalid output sampling interval: "
                f"low={output_low} > high={output_high}"
            )
629

630
631
        logger.info(
            "Sampling input_len from [%s, %s] and output_len from [%s, %s]",
632
633
634
635
636
            input_low,
            input_high,
            output_low,
            output_high,
        )
637

638
639
640
        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)
641
        return input_lens, output_lens, offsets
642

643
644
645
    def generate_token_sequence(
        self,
        *,
646
        tokenizer: TokenizerLike,
647
648
649
650
651
652
        prefix_token_ids: list[int],
        prefix_len: int,
        vocab_size: int,
        input_len: int,
        offset: int,
        index: int,
653
        allowed_tokens: np.ndarray,
654
    ) -> tuple[str, int, int]:
655
656
657
658
659
660
661
662
663
664
        """
        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,
665
        the encoded sequence is truncated before being decoded again.
666
        """
667
668
669
670
671
        # 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()
672
673
674
675
        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)
676
        prompt, adjusted_token_sequence, token_mismatch = (
677
            gen_prompt_decode_to_target_len(
678
679
680
681
682
683
                tokenizer=tokenizer,
                token_sequence=token_sequence,
                target_token_len=total_input_len,
                add_special_tokens=False,
                rng=self._rng,
            )
684
685
686
        )
        total_input_len = len(adjusted_token_sequence)
        return prompt, total_input_len, token_mismatch
687
688


689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
# -----------------------------------------------------------------------------
# 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,
706
        tokenizer: TokenizerLike,
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
        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
        )
736

737
        vocab_size = tokenizer.vocab_size
738
739
740
        prohibited_tokens = tokenizer.all_special_ids
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))
741
742
743
744
745
746
747
748
749
750

        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,
751
                allowed_tokens=allowed_tokens,
752
753
754
755
756
757
758
759
760
761
762
763
764
            )
        )

        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,
765
                allowed_tokens=allowed_tokens,
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
            )
            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


801
802
803
804
# -----------------------------------------------------------------------------
# MultiModalDataset Implementation
# -----------------------------------------------------------------------------

805

806
807
808
809
810
811
class RandomMultiModalDataset(RandomDataset):
    """
    Synthetic multimodal dataset (text + images) that extends RandomDataset.

    Status:
    - Images: supported via synthetic RGB data.
812
    - Video: supported via synthetic RGB data.
813
814
815
816
817
818
819
820
    - 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
821
       mapping (height, width, num_frames) → probability. We treat
822
       `num_frames`=1 as image and `num_frames` > 1 as video.
823
       Entries with zero probability are removed and the rest are renormalized
824
825
826
827
828
829
830
       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}
831
832
      - Two image buckets (`num_frames`=1) and one video bucket
      (`num_frames`=16).
833
834
835
836
    OBS.: Only image sampling is supported for now.
    """

    IS_MULTIMODAL = True
837
    DEFAULT_LIMIT_MM_PER_PROMPT = {"image": 255, "video": 1}
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852

    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.
853
854
855

        NOTE: iid pixel sampling results in worst-case compression
        (good for stressing I/O), but very unlike real photos.
856
857
858
859
860
861
862
863
864
865
866
        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)

867
868
869
    def generate_synthetic_video(
        self, width: int, height: int, num_frames: int
    ) -> dict:
870
        """Generate synthetic video with random values.
871

872
873
        Creates a video with random pixel values, encodes it to MP4 format,
        and returns the content as bytes.
874
        """
875
876
        import cv2

877
878
879
880
881
882
883
884
885
886
887
        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

888
        with NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
            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}
910
911
912
913
914
915
916
917
918
919

    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}")

920
921
922
    def normalize_bucket_config(
        self, bucket_config: dict[tuple[int, int, int], float]
    ) -> dict[tuple[int, int, int], float]:
923
924
925
926
927
928
929
930
931
932
933
        """
        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:
934
935
936
            raise ValueError(
                "Got invalid bucket config. Bucket config values must be non-zero."
            )
937
938
939
940
        # 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()}

941
942
943
944
    def generate_mm_item(
        self,
        mm_item_config: tuple[int, int, int],
    ) -> Mapping[str, Any]:
945
        """
946
        Create synthetic images and videos and
947
948
949
950
        apply process_image/process_video respectively.
        This follows the OpenAI API chat completions
        https://github.com/openai/openai-python
        """
951

952
        if self.map_config_to_modality(mm_item_config) == "image":
953
954
955
            return process_image(
                self.generate_synthetic_image(mm_item_config[1], mm_item_config[0])
            )
956
        elif self.map_config_to_modality(mm_item_config) == "video":
957
958
959
960
961
            return process_video(
                self.generate_synthetic_video(
                    mm_item_config[1], mm_item_config[0], mm_item_config[2]
                )
            )
962
        else:
963
            raise ValueError(f"Invalid multimodal item configuration: {mm_item_config}")
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983

    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:
984
985
986
987
988
                raise ValueError(
                    f"Modality {modality} is not in "
                    f"limit_mm_per_prompt: "
                    f"{limit_mm_per_prompt.keys()}"
                )
989

990
        # Remove zero probability entries
991
992
993
        # and normalize bucket config to sum to 1
        bucket_config = self.normalize_bucket_config(bucket_config)
        logger.info(
994
995
            "Normalized bucket config: %s",
            bucket_config,
996
997
        )
        # Only consider limit per prompt for modalities in bucket config
998
        allowed_modalities = {self.map_config_to_modality(cfg) for cfg in bucket_config}
999
        limit_mm_per_prompt = {
1000
1001
            k: v for k, v in limit_mm_per_prompt.items() if k in allowed_modalities
        }
1002
        if not limit_mm_per_prompt:
1003
            raise ValueError("No valid limits for modalities present in bucket_config.")
1004
1005

        logger.info(
1006
1007
            "Updated mm-limit-per-prompt: %s",
            limit_mm_per_prompt,
1008
1009
1010
1011
1012
        )

        # 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(
1013
            sum(limit_mm_per_prompt.values()),
1014
            math.ceil(base_items_per_request * (1 + num_mm_items_range_ratio)),
1015
1016
1017
        )
        # Ensure min num mm items is at least 0
        min_num_mm_items = max(
1018
            0, math.floor(base_items_per_request * (1 - num_mm_items_range_ratio))
1019
1020
1021
        )
        # Raise error if min num mm items is greater than max num mm items
        if min_num_mm_items > max_num_mm_items:
1022
1023
1024
1025
            raise ValueError(
                f"Min num mm items is greater than max mm items: "
                f"{min_num_mm_items} > {max_num_mm_items}"
            )
1026

1027
1028
        logger.info(
            "Sampling number of multimodal items from [%s, %s]",
1029
1030
            min_num_mm_items,
            max_num_mm_items,
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
        )

        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],
1046
    ) -> Iterator[tuple[int, int, int]]:
1047
1048
1049
1050
1051
        """
        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.
1052
1053
        Loop until the number of multimodal items sampled is equal to
        request_num_mm_items or limit of multimodal items per prompt
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
        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)
1065
        )
1066
1067
1068
1069
        # If request_num_mm_items is 0, yield an empty iterator
        if request_num_mm_items == 0:
            return
        # Initialize modality counters
1070
        modality_counter = {self.map_config_to_modality(k): 0 for k in bucket_config}
1071
1072
1073
1074
1075
        # 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
1076
1077
1078
            mm_item_config = self._rng.choice(
                list(bucket_config_copy.keys()), p=list(bucket_config_copy.values())
            )
1079
1080
1081
1082
            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
1083
                yield (mm_item_config)
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
            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()):
1094
1095
1096
                    logger.warning(
                        "Exhausted all multimodal items of modality %s", modality
                    )
1097
1098
                    break
                # Renormalize the bucket config
1099
                bucket_config_copy = self.normalize_bucket_config(bucket_config_copy)
1100
1101
1102

    def sample(
        self,
1103
        tokenizer: TokenizerLike,
1104
1105
        num_requests: int,
        request_id_prefix: str = "",
1106
        no_oversample: bool = False,
1107
1108
1109
1110
1111
1112
1113
        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,
1114
1115
1116
        bucket_config: dict[
            tuple[int, int, int], float
        ] = DEFAULT_MM_ITEM_BUCKET_CONFIG,
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
        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
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
        # 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
1154
        prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len)
1155
1156
        # Add synthetic multimodal items to each request
        mm_requests = []
1157
        token_mismatch_total = 0
1158
        for i in range(num_requests):
1159
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
1160
1161
1162
1163
1164
1165
1166
                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,
1167
                allowed_tokens=allowed_tokens,
1168
            )
1169
            token_mismatch_total += token_mismatch
1170
1171
1172
1173
1174
1175
1176
1177
            # 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,
            )

1178
1179
1180
1181
1182
1183
1184
            mm_content = cast(
                list[dict[str, Any]],
                [
                    self.generate_mm_item(mm_item_config)
                    for mm_item_config in mm_item_iterator
                ],
            )
1185
1186

            if enable_multimodal_chat:
1187
                # NOTE: For now this option is only provided for completeness
1188
1189
1190
                # given that the serve.py benchmark currently does not use it.
                mm_chat_prompt: Any = prompt
                mm_chat_prompt = self.apply_multimodal_chat_transformation(
1191
1192
                    prompt, mm_content
                )
1193
1194
1195
1196
1197
1198
1199
1200
1201
                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(
1202
1203
1204
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
1205
                    multi_modal_data=mm_content,
1206
                    request_id=request_id_prefix + str(i),
1207
1208
                )
            mm_requests.append(sample_request)
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220

        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,
            )

1221
        return mm_requests
1222

1223

1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
# -----------------------------------------------------------------------------
# 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 = [
1247
1248
            entry
            for entry in self.data
1249
1250
1251
            if "conversations" in entry and len(entry["conversations"]) >= 2
        ]
        random.seed(self.random_seed)
1252
1253
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1254
1255
1256

    def sample(
        self,
1257
        tokenizer: TokenizerLike,
1258
        num_requests: int,
1259
1260
1261
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1262
        enable_multimodal_chat: bool = False,
1263
        request_id_prefix: str = "",
1264
        no_oversample: bool = False,
1265
1266
1267
        **kwargs,
    ) -> list:
        samples: list = []
1268
        ind = 0
1269
1270
1271
1272
1273
1274
1275
1276
        for entry in self.data:
            if len(samples) >= num_requests:
                break
            prompt, completion = (
                entry["conversations"][0]["value"],
                entry["conversations"][1]["value"],
            )

1277
            lora_request = self.get_random_lora_request(
1278
1279
                max_loras=max_loras, lora_path=lora_path
            )
1280
1281
1282
            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
1283
1284
1285
1286
1287
1288
            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,
            ):
1289
                continue
1290
1291
1292
            if image_path := entry.get("image"):
                mm_content = process_image(image_path)
            elif video_path := entry.get("video"):
1293
                mm_content = process_video(video_path)
1294
            else:
1295
                mm_content = None
1296
            if enable_multimodal_chat:
1297
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
1298
1299
1300
1301
1302
1303
            samples.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=new_output_len,
                    lora_request=lora_request,
1304
                    multi_modal_data=mm_content,
1305
                    request_id=request_id_prefix + str(ind),
1306
1307
                )
            )
1308
            ind += 1
1309
1310
1311
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
1312
1313
1314
        return samples


1315
1316
class _ValidateDatasetArgs(argparse.Action):
    """Argparse action to validate dataset name and path compatibility."""
1317

1318
1319
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, values)
1320

1321
        # Get current values of both dataset_name and dataset_path
1322
1323
        dataset_name = getattr(namespace, "dataset_name", "random")
        dataset_path = getattr(namespace, "dataset_path", None)
1324

1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
        # 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}"
            )


1335
def add_dataset_parser(parser: FlexibleArgumentParser):
1336
1337
1338
1339
1340
    parser.add_argument(
        "--trust-remote-code",
        action="store_true",
        help="Trust remote code from huggingface",
    )
1341
1342
1343
1344
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
        "--num-prompts",
        type=int,
1345
        default=DEFAULT_NUM_PROMPTS,
1346
1347
1348
1349
1350
1351
        help="Number of prompts to process.",
    )
    parser.add_argument(
        "--dataset-name",
        type=str,
        default="random",
1352
        action=_ValidateDatasetArgs,
1353
        choices=[
1354
1355
1356
1357
1358
            "sharegpt",
            "burstgpt",
            "sonnet",
            "random",
            "random-mm",
1359
            "random-rerank",
1360
1361
            "hf",
            "custom",
1362
            "custom_mm",
1363
1364
            "prefix_repetition",
            "spec_bench",
1365
        ],
1366
1367
        help="Name of the dataset to benchmark on.",
    )
1368
1369
1370
1371
1372
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
1373
1374
1375
1376
    parser.add_argument(
        "--dataset-path",
        type=str,
        default=None,
1377
        action=_ValidateDatasetArgs,
1378
1379
1380
        help="Path to the sharegpt/sonnet dataset. "
        "Or the huggingface dataset ID if using HF dataset.",
    )
1381
1382
1383
    parser.add_argument(
        "--no-oversample",
        action="store_true",
1384
        help="Do not oversample if the dataset has fewer samples than num-prompts.",
1385
    )
1386
1387
1388
    parser.add_argument(
        "--skip-chat-template",
        action="store_true",
1389
        help="Skip applying chat template to prompt for datasets that support it.",
1390
    )
1391
1392
1393
1394
1395
    parser.add_argument(
        "--enable-multimodal-chat",
        action="store_true",
        help="Enable multimodal chat transformation for datasets that support it.",
    )
1396
1397
1398
1399
1400
    parser.add_argument(
        "--disable-shuffle",
        action="store_true",
        help="Disable shuffling of dataset samples for deterministic ordering.",
    )
1401
1402
1403
1404
1405
1406
1407

    # 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,
1408
1409
1410
        help="Number of output tokens per request. Unless it is set to -1, the "
        "value overrides potential output length loaded from the dataset. It is "
        "used only for custom dataset.",
1411
1412
    )

1413
1414
1415
1416
1417
    spec_bench_group = parser.add_argument_group("spec bench dataset options")
    spec_bench_group.add_argument(
        "--spec-bench-output-len",
        type=int,
        default=256,
1418
        help="Num of output tokens per request, used only for spec bench dataset.",
1419
1420
1421
1422
1423
    )
    spec_bench_group.add_argument(
        "--spec-bench-category",
        type=str,
        default=None,
1424
        help="Category for spec bench dataset. If None, use all categories.",
1425
1426
    )

1427
1428
1429
1430
1431
    sonnet_group = parser.add_argument_group("sonnet dataset options")
    sonnet_group.add_argument(
        "--sonnet-input-len",
        type=int,
        default=550,
1432
        help="Number of input tokens per request, used only for sonnet dataset.",
1433
1434
1435
1436
1437
    )
    sonnet_group.add_argument(
        "--sonnet-output-len",
        type=int,
        default=150,
1438
        help="Number of output tokens per request, used only for sonnet dataset.",
1439
1440
1441
1442
1443
    )
    sonnet_group.add_argument(
        "--sonnet-prefix-len",
        type=int,
        default=200,
1444
        help="Number of prefix tokens per request, used only for sonnet dataset.",
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
    )

    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.",
    )

1456
1457
1458
1459
1460
    blazedit_group = parser.add_argument_group("blazedit dataset options")
    blazedit_group.add_argument(
        "--blazedit-min-distance",
        type=float,
        default=0.0,
1461
        help="Minimum distance for blazedit dataset. Min: 0, Max: 1.0",
1462
1463
1464
1465
1466
    )
    blazedit_group.add_argument(
        "--blazedit-max-distance",
        type=float,
        default=1.0,
1467
        help="Maximum distance for blazedit dataset. Min: 0, Max: 1.0",
1468
1469
    )

1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
    asr_group = parser.add_argument_group("asr dataset options")
    asr_group.add_argument(
        "--asr-max-audio-len-sec",
        type=float,
        default=float("inf"),
        help="Maximum audio length in seconds for ASR dataset.",
    )
    asr_group.add_argument(
        "--asr-min-audio-len-sec",
        type=float,
        default=0.0,
        help="Minimum audio length in seconds for ASR dataset.",
    )

1484
    random_group = parser.add_argument_group("random dataset options")
1485
1486
1487
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
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
    add_random_dataset_base_args(random_group)

    random_mm_group = parser.add_argument_group(
        "random multimodal dataset options extended from random dataset"
    )
    add_random_multimodal_dataset_args(random_mm_group)

    hf_group = parser.add_argument_group("hf dataset options")
    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."
    )
    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."
        ),
    )
    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.",
    )

    prefix_repetition_group = parser.add_argument_group(
        "prefix repetition dataset options"
    )
    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.",
    )


def add_random_dataset_base_args(
    parser_or_group: FlexibleArgumentParser | argparse._ArgumentGroup,
) -> None:
    """Add CLI arguments for base random dataset options.

    This function adds arguments needed for:
    - random (random dataset)
    - random-mm (random multimodal dataset)
    - random-rerank (random dataset for reranking)

    Args:
        parser_or_group: Either a parser or an argument group to add arguments to.
    """
    parser_or_group.add_argument(
1564
1565
1566
        "--random-input-len",
        type=int,
        default=1024,
1567
        help="Number of input tokens per request, used only for random sampling.",
1568
    )
1569
    parser_or_group.add_argument(
1570
1571
1572
        "--random-output-len",
        type=int,
        default=128,
1573
        help="Number of output tokens per request, used only for random sampling.",
1574
    )
1575
    parser_or_group.add_argument(
1576
1577
1578
1579
1580
1581
1582
1583
        "--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)].",
    )
1584
    parser_or_group.add_argument(
1585
1586
1587
        "--random-prefix-len",
        type=int,
        default=0,
1588
1589
1590
1591
1592
1593
1594
1595
        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)]."
        ),
1596
    )
1597
    parser_or_group.add_argument(
1598
1599
1600
        "--random-batch-size",
        type=int,
        default=1,
1601
        help=("Batch size for random sampling. Only used for embeddings benchmark."),
1602
    )
1603
    parser_or_group.add_argument(
1604
1605
1606
1607
1608
1609
1610
        "--no-reranker",
        action="store_true",
        help=(
            "Whether the model supports reranking natively."
            " Only used for reranker benchmark."
        ),
    )
1611

1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624

def add_random_multimodal_dataset_args(
    parser_or_group: FlexibleArgumentParser | argparse._ArgumentGroup,
) -> None:
    """Add CLI arguments for random multimodal dataset options.

    This function adds arguments needed for:
    - random-mm (random multimodal dataset)

    Args:
        parser_or_group: Either a parser or an argument group to add arguments to.
    """
    parser_or_group.add_argument(
1625
1626
1627
1628
1629
1630
1631
1632
1633
        "--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."
        ),
    )
1634
    parser_or_group.add_argument(
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
        "--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."
        ),
    )
1649
    parser_or_group.add_argument(
1650
1651
1652
1653
1654
        "--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. "
1655
            '\'{"image": 3, "video": 0}\'. The sampled per-request item '
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
            "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)
1672
1673
1674
1675
1676
                if not (
                    isinstance(key, tuple)
                    and len(key) == 3
                    and all(isinstance(x, int) for x in key)
                ):
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
                    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.")

1693
    parser_or_group.add_argument(
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
        "--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."
        ),
1712
1713
    )

1714

1715
def get_samples(args, tokenizer: TokenizerLike) -> list[SampleRequest]:
1716
1717
1718
    if not hasattr(args, "request_id_prefix"):
        args.request_id_prefix = ""

1719
    if args.dataset_name == "custom":
1720
1721
1722
        dataset = CustomDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1723
1724
1725
1726
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
1727
            skip_chat_template=args.skip_chat_template,
1728
            request_id_prefix=args.request_id_prefix,
1729
            no_oversample=args.no_oversample,
1730
1731
        )

1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
    elif args.dataset_name == "custom_mm":
        dataset = CustomMMDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
            enable_multimodal_chat=args.enable_multimodal_chat,
            request_id_prefix=args.request_id_prefix,
            no_oversample=args.no_oversample,
        )

1745
    elif args.dataset_name == "sonnet":
1746
1747
1748
        dataset = SonnetDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1749
        # For the "sonnet" dataset, formatting depends on the backend.
1750
        if args.backend == "openai-chat":
1751
1752
1753
1754
1755
1756
1757
            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,
1758
                request_id_prefix=args.request_id_prefix,
1759
                no_oversample=args.no_oversample,
1760
1761
1762
            )
        else:
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
1763
1764
                "Tokenizer/model must have chat template for sonnet dataset."
            )
1765
1766
1767
1768
1769
1770
1771
            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,
1772
                request_id_prefix=args.request_id_prefix,
1773
                no_oversample=args.no_oversample,
1774
1775
1776
1777
1778
            )

    elif args.dataset_name == "hf":
        # all following datasets are implemented from the
        # HuggingFaceDataset base class
1779
        hf_kwargs = {}
1780
1781
1782
1783
        if (
            args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS
        ):
1784
            dataset_class = VisionArenaDataset
1785
            args.hf_split = args.hf_split if args.hf_split else "train"
1786
            args.hf_subset = None
1787
1788
1789
1790
1791
        elif (
            args.dataset_path in MMVUDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMVUDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMVUDataset
1792
            args.hf_split = args.hf_split if args.hf_split else "validation"
1793
            args.hf_subset = None
1794
1795
1796
1797
        elif (
            args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS
        ):
1798
            dataset_class = InstructCoderDataset
1799
            args.hf_split = args.hf_split if args.hf_split else "train"
1800
1801
1802
1803
        elif (
            args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MTBenchDataset.SUPPORTED_DATASET_PATHS
        ):
1804
            dataset_class = MTBenchDataset
1805
            args.hf_split = args.hf_split if args.hf_split else "train"
1806
1807
1808
1809
1810
        elif (
            args.dataset_path in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MultiModalConversationDataset
1811
1812
1813
1814
        elif (
            args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS
        ):
1815
            dataset_class = ConversationDataset
1816
1817
1818
1819
        elif (
            args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS
        ):
1820
            dataset_class = AIMODataset
1821
            args.hf_split = args.hf_split if args.hf_split else "train"
1822
        elif (
1823
            args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS  # noqa: E501
1824
1825
            or args.hf_name in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS
        ):
1826
            dataset_class = NextEditPredictionDataset
1827
            args.hf_split = args.hf_split if args.hf_split else "train"
1828
1829
1830
1831
        elif (
            args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS
        ):
1832
            dataset_class = ASRDataset
1833
1834
1835
1836
1837
            args.hf_split = args.hf_split if args.hf_split else "train"
            hf_kwargs = {
                "asr_min_audio_len_sec": args.asr_min_audio_len_sec,
                "asr_max_audio_len_sec": args.asr_max_audio_len_sec,
            }
1838
1839
        elif args.dataset_path in BlazeditDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = BlazeditDataset
1840
            args.hf_split = args.hf_split if args.hf_split else "train"
1841
1842
1843
1844
            hf_kwargs = {
                "min_distance": args.blazedit_min_distance,
                "max_distance": args.blazedit_max_distance,
            }
1845
1846
1847
1848
        elif (
            args.dataset_path in MLPerfDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MLPerfDataset.SUPPORTED_DATASET_PATHS
        ):
1849
            dataset_class = MLPerfDataset
1850
            args.hf_split = args.hf_split if args.hf_split else "train"
1851
1852
1853
1854
1855
        elif (
            args.dataset_path in MMStarDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMStarDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMStarDataset
1856
            args.hf_split = args.hf_split if args.hf_split else "val"
1857
            args.hf_subset = None
1858
        else:
1859
1860
1861
1862
1863
1864
1865
            supported_datasets = set(
                [
                    dataset_name
                    for cls in HuggingFaceDataset.__subclasses__()
                    for dataset_name in cls.SUPPORTED_DATASET_PATHS
                ]
            )
1866
1867
1868
1869
1870
            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 "
1871
1872
                "like to add support for additional dataset formats."
            )
1873

1874
1875
        if dataset_class.IS_MULTIMODAL and not (
            args.backend in ("openai-chat", "openai-audio")
1876
            or "embeddings-" in args.backend
1877
        ):
1878
1879
            # multi-modal benchmark is only available on OpenAI Chat
            # endpoint-type.
1880
1881
            raise ValueError(
                "Multi-modal content is only supported on 'openai-chat' and "
1882
1883
                "'openai-audio' backends."
            )
1884
1885
1886
1887
1888
        input_requests = dataset_class(
            dataset_path=args.dataset_path,
            dataset_subset=args.hf_subset,
            dataset_split=args.hf_split,
            random_seed=args.seed,
1889
            no_stream=args.no_stream,
1890
            hf_name=args.hf_name,
1891
            disable_shuffle=args.disable_shuffle,
1892
            trust_remote_code=args.trust_remote_code,
1893
1894
1895
1896
        ).sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.hf_output_len,
1897
            enable_multimodal_chat=args.enable_multimodal_chat,
1898
            request_id_prefix=args.request_id_prefix,
1899
            no_oversample=args.no_oversample,
1900
            skip_chat_template=args.skip_chat_template,
1901
            **hf_kwargs,
1902
1903
1904
1905
1906
        )

    else:
        # For datasets that follow a similar structure, use a mapping.
        dataset_mapping = {
1907
            "spec_bench": lambda: SpecBench(
1908
1909
1910
                dataset_path=args.dataset_path,
                category=args.spec_bench_category,
                disable_shuffle=args.disable_shuffle,
1911
            ).sample(
1912
1913
1914
                num_requests=args.num_prompts,
                tokenizer=tokenizer,
                output_len=args.spec_bench_output_len,
1915
                enable_multimodal_chat=args.enable_multimodal_chat,
1916
                request_id_prefix=args.request_id_prefix,
1917
                no_oversample=args.no_oversample,
1918
            ),
1919
            "sharegpt": lambda: ShareGPTDataset(
1920
1921
1922
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1923
1924
1925
1926
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
1927
                enable_multimodal_chat=args.enable_multimodal_chat,
1928
                request_id_prefix=args.request_id_prefix,
1929
                no_oversample=args.no_oversample,
1930
1931
            ),
            "burstgpt": lambda: BurstGPTDataset(
1932
1933
1934
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1935
1936
1937
1938
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                request_id_prefix=args.request_id_prefix,
1939
                no_oversample=args.no_oversample,
1940
1941
            ),
            "random": lambda: RandomDataset(
1942
1943
1944
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1945
            ).sample(
1946
1947
1948
1949
1950
1951
                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,
1952
                request_id_prefix=args.request_id_prefix,
1953
                batchsize=args.random_batch_size,
1954
                no_oversample=args.no_oversample,
1955
            ),
1956
            "random-mm": lambda: RandomMultiModalDataset(
1957
1958
1959
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
            ).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,
1971
                enable_multimodal_chat=args.enable_multimodal_chat,
1972
                request_id_prefix=args.request_id_prefix,
1973
                no_oversample=args.no_oversample,
1974
            ),
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
            "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,
            ),
1988
            "prefix_repetition": lambda: PrefixRepetitionRandomDataset(
1989
1990
1991
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1992
1993
1994
1995
1996
1997
1998
            ).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,
1999
                request_id_prefix=args.request_id_prefix,
2000
                no_oversample=args.no_oversample,
2001
            ),
2002
2003
2004
        }

        try:
2005
            # Enforce endpoint compatibility for multimodal datasets.
2006
            if args.dataset_name == "random-mm" and args.backend not in ["openai-chat"]:
2007
2008
2009
2010
                raise ValueError(
                    "Multi-modal content (images) is only supported on "
                    "'openai-chat' backend."
                )
2011
2012
2013
2014
2015
2016
2017
            input_requests = dataset_mapping[args.dataset_name]()
        except KeyError as err:
            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err

    return input_requests


2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
# -----------------------------------------------------------------------------
# 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.,
    ```
2028
2029
2030
    {"prompt": "What is the capital of India?", "output_tokens": 10}
    {"prompt": "What is the capital of Iran?", "output_tokens": 1520}
    {"prompt": "What is the capital of China?", "output_tokens": 819}
2031
    ```
2032
2033
    Note that 'output_tokens' column is optional and has to be provided only if
    'custom-output-len' argument is None or -1.
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
    """

    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"):
2053
            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066

            # 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(
2067
2068
                "Only JSONL format is supported for CustomDataset."
            )
2069
2070

        random.seed(self.random_seed)
2071
2072
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2073
2074
2075

    def sample(
        self,
2076
        tokenizer: TokenizerLike,
2077
        num_requests: int,
2078
2079
2080
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
2081
2082
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
2083
        request_id_prefix: str = "",
2084
        no_oversample: bool = False,
2085
2086
        **kwargs,
    ) -> list:
2087
2088
2089
2090
        # load all data if needed
        self.num_available_samples = len(self.data)
        if num_requests <= 0:
            num_requests = self.num_available_samples
2091
2092
2093
2094
2095
            logger.info(
                "num_requests is set to 0 or negative, "
                "so using all available samples: %d",
                num_requests,
            )
2096

2097
        sampled_requests = []
2098
        for i, item in enumerate(self.data):
2099
2100
2101
2102
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["prompt"]

2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
            if tokenizer is None:
                new_output_len = 1
            else:
                new_output_len = output_len
                if output_len is None or output_len == -1:
                    # check that the request has an 'output_tokens' field
                    if "output_tokens" not in item:
                        raise ValueError(
                            "If no output length is provided the "
                            "custom dataset must contain an 'output_tokens' field."
                        )
                    # Use number of output tokens from the request data
                    try:
                        new_output_len = int(item["output_tokens"])
                    except (ValueError, TypeError) as e:
                        raise ValueError(
                            f"Invalid value for 'output_tokens' in custom dataset: "
                            f"'{item['output_tokens']}'. Must be an integer."
                        ) from e

            if tokenizer is None:
                prompt_len = 1
            else:
                # apply template
                if not skip_chat_template:
                    prompt = tokenizer.apply_chat_template(
                        [{"role": "user", "content": prompt}],
                        add_generation_prompt=True,
                        tokenize=False,
2132
                    )
2133

2134
                prompt_len = len(tokenizer(prompt).input_ids)
2135
2136
2137
2138
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
2139
                    expected_output_len=new_output_len,
2140
                    request_id=request_id_prefix + str(i),
2141
2142
2143
2144
2145
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2146
2147
2148
2149

        return sampled_requests


2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
class CustomMMDataset(CustomDataset):
    """
    Implements the Custom MultiModal dataset. Loads data from a JSONL file and generates
    sample requests based on conversation turns. E.g.,
    ```
    {
        "prompt": "How many red blocks in the given images?",
        "image_files": ["path/to/image1.png", "path/to/image2.png"],
    }
    {
        "prompt": "Which country has the most pokemons based on the given graphs?",
        "image_files": ["path/to/image.png"],
    }
    ```

    NOTE: Only the first image file in "image_files" is used for each sample request.

    This is used to benchmark multimodal LLMs on arbitrary datasets.
    """

    IS_MULTIMODAL = True

    def sample(
        self,
        tokenizer: TokenizerLike,
        num_requests: int,
        output_len: int | None = None,
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
        # load all data if needed
        self.num_available_samples = len(self.data)
        if num_requests <= 0:
            num_requests = self.num_available_samples
            logger.info(
                "num_requests is set to 0 or negative, "
                "so using all available samples: %d",
                num_requests,
            )

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

            prompt_len = len(tokenizer(prompt).input_ids)
            images = item["image_files"]
            if len(images) > 1:
                logger.warning(
                    "Multiple image files found for sample %d. "
                    "Only the first image will be used.",
                    i,
                )
            mm_content = process_image(images[0])
            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
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)

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

        return sampled_requests


2229
2230
2231
2232
2233
2234
2235
2236
# -----------------------------------------------------------------------------
# Spec Bench Dataset Implementation
# -----------------------------------------------------------------------------


class SpecBench(CustomDataset):
    """
    Implements the SpecBench dataset: https://github.com/hemingkx/Spec-Bench
2237
    Download the dataset using:
2238
    wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
2239
    """  # noqa: E501
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252

    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
2253
        jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2254
2255
2256
2257
2258
2259
2260

        # 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
2261
            if (not self.category) or (self.category == row["category"]):
2262
2263
2264
2265
                prompt = row["turns"][0]
                self.data.append({"prompt": prompt})

        random.seed(self.random_seed)
2266
2267
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2268
2269
2270
2271

    def sample(self, **kwargs) -> list:
        # leverage CustomDataset sample
        return super().sample(**kwargs)
2272
2273


2274
2275
2276
2277
# -----------------------------------------------------------------------------
# Sonnet Dataset Implementation
# -----------------------------------------------------------------------------

2278

2279
2280
2281
@deprecated(
    "SonnetDataset is deprecated and will be removed in a future version.",
)
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
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,
2308
        tokenizer: TokenizerLike,
2309
2310
2311
2312
2313
        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,
2314
        request_id_prefix: str = "",
2315
        no_oversample: bool = False,
2316
2317
2318
2319
        **kwargs,
    ) -> list:
        # Calculate average token length for a poem line.
        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
2320
        avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines)
2321
2322
2323
2324

        # 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}]
2325
2326
2327
        base_fmt = tokenizer.apply_chat_template(
            base_msg, add_generation_prompt=True, tokenize=False
        )
2328
2329
2330
2331
        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 "
2332
2333
                f"({base_offset})."
            )
2334
2335
2336
2337
2338
2339
2340

        # 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 = []
2341
        ind = 0
2342
        while len(samples) < num_requests:
2343
2344
2345
            extra_lines = random.choices(
                self.data, k=num_input_lines - num_prefix_lines
            )
2346
2347
2348
            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
            msg = [{"role": "user", "content": prompt}]
            prompt_formatted = tokenizer.apply_chat_template(
2349
2350
                msg, add_generation_prompt=True, tokenize=False
            )
2351
2352
2353
2354
            prompt_len = len(tokenizer(prompt_formatted).input_ids)
            if prompt_len <= input_len:
                samples.append(
                    SampleRequest(
2355
                        prompt=prompt_formatted if return_prompt_formatted else prompt,
2356
2357
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
2358
2359
2360
                        request_id=request_id_prefix + str(ind),
                    )
                )
2361
                ind += 1
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
        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()

2381
2382
2383
    def load_data(
        self,
    ):
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
        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):
2397
            data = self.data.sample(n=num_requests, random_state=self.random_seed)
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
        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,
2409
        tokenizer: TokenizerLike,
2410
        num_requests: int,
2411
2412
        max_loras: int | None = None,
        lora_path: str | None = None,
2413
        request_id_prefix: str = "",
2414
        no_oversample: bool = False,
2415
2416
2417
2418
2419
2420
2421
        **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])
2422
            lora_req = self.get_random_lora_request(
2423
2424
                max_loras=max_loras, lora_path=lora_path
            )
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
            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,
2436
                    request_id=request_id_prefix + str(i),
2437
2438
                )
            )
2439
2440
2441
2442
2443
2444
2445
2446
2447
        return samples


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

2448
    SUPPORTED_DATASET_PATHS: set[str] | dict[str, Callable] = set()
2449
2450
2451
2452
2453

    def __init__(
        self,
        dataset_path: str,
        dataset_split: str,
2454
        no_stream: bool = False,
2455
2456
        dataset_subset: str | None = None,
        hf_name: str | None = None,
2457
        trust_remote_code: bool = False,
2458
2459
2460
2461
2462
2463
        **kwargs,
    ) -> None:
        super().__init__(dataset_path=dataset_path, **kwargs)

        self.dataset_split = dataset_split
        self.dataset_subset = dataset_subset
2464
        self.load_stream = not no_stream
2465
        self.hf_name = hf_name or dataset_path
2466
        self.trust_remote_code = trust_remote_code
2467
2468
2469
2470
2471
2472
2473
2474
        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,
2475
            streaming=self.load_stream,
2476
            trust_remote_code=self.trust_remote_code,
2477
        )
2478
2479
        if not getattr(self, "disable_shuffle", False):
            self.data = self.data.shuffle(seed=self.random_seed)
2480
2481
2482
2483
2484
2485
2486
2487


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


class ConversationDataset(HuggingFaceDataset):
2488
    """Dataset for text-only conversation data."""
2489

2490
    SUPPORTED_DATASET_PATHS = {
2491
        "Aeala/ShareGPT_Vicuna_unfiltered",
2492
    }
2493
2494
2495
2496
    IS_MULTIMODAL = False

    def sample(
        self,
2497
        tokenizer: TokenizerLike,
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
        num_requests: int,
        output_len: int | None = None,
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
        # Filter examples with at least 2 conversations
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
        sampled_requests = []
        ind = 0
        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
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
                continue
            mm_content = process_image(item["image"]) if "image" in item else None
            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
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
            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(ind),
                )
            )
            ind += 1
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
        return sampled_requests


class MultiModalConversationDataset(HuggingFaceDataset):
    """Dataset for multimodal conversation data."""

    SUPPORTED_DATASET_PATHS = {
        "lmms-lab/LLaVA-OneVision-Data",
    }
2553
    IS_MULTIMODAL = True
2554

2555
2556
    def sample(
        self,
2557
        tokenizer: TokenizerLike,
2558
        num_requests: int,
2559
        output_len: int | None = None,
2560
2561
2562
2563
2564
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2565
        # Filter examples with at least 2 conversations
2566
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
2567
        sampled_requests = []
2568
        ind = 0
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
        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
2583
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
2584
                continue
2585
            mm_content = process_image(item["image"]) if "image" in item else None
2586
2587
2588
2589
            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
2590
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2591
2592
2593
2594
2595
2596
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2597
                    request_id=request_id_prefix + str(ind),
2598
2599
                )
            )
2600
            ind += 1
2601
2602
2603
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
        return sampled_requests


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


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

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2619
2620
        "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"],
2621
    }
2622
    IS_MULTIMODAL = True
2623
2624
2625

    def sample(
        self,
2626
        tokenizer: TokenizerLike,
2627
        num_requests: int,
2628
        output_len: int | None = None,
2629
        enable_multimodal_chat: bool = False,
2630
        request_id_prefix: str = "",
2631
        no_oversample: bool = False,
2632
2633
        **kwargs,
    ) -> list:
2634
2635
2636
2637
        parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
        if parser_fn is None:
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")

2638
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2639

2640
        sampled_requests = []
2641
        for i, item in enumerate(self.data):
2642
2643
            if len(sampled_requests) >= num_requests:
                break
2644

2645
2646
            prompt = parser_fn(item)
            mm_content = process_image(item["images"][0])
2647
            prompt_len = len(tokenizer.encode(prompt))
2648
2649
2650
2651
            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
2652
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2653

2654
2655
2656
2657
2658
2659
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2660
                    request_id=request_id_prefix + str(i),
2661
2662
                )
            )
2663

2664
2665
2666
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2667
2668
2669
        return sampled_requests


2670
2671
2672
2673
2674
2675
2676
2677
class MMVUDataset(HuggingFaceDataset):
    """
    MMVU Dataset.
    https://huggingface.co/datasets/yale-nlp/MMVU
    """

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2678
2679
2680
        "yale-nlp/MMVU": lambda x: x["question"]
        + " "
        + (" ".join(f"{k}.{v}" for k, v in x["choices"].items())),
2681
2682
2683
2684
    }

    def sample(
        self,
2685
        tokenizer: TokenizerLike,
2686
        num_requests: int,
2687
        output_len: int | None = None,
2688
2689
2690
2691
2692
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2693
2694
2695
2696
        parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
        if parser_fn is None:
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")

2697
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2698

2699
2700
2701
2702
        sampled_requests = []
        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
2703

2704
2705
            prompt = parser_fn(item)
            mm_content = process_video(item["video"])
2706
            prompt_len = len(tokenizer.encode(prompt))
2707
2708
2709
2710
            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
2711
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2712

2713
2714
2715
2716
2717
2718
2719
            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),
2720
2721
                )
            )
2722

2723
2724
2725
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2726
2727
2728
        return sampled_requests


2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
# -----------------------------------------------------------------------------
# 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",
    }

2749
2750
    def sample(
        self,
2751
        tokenizer: TokenizerLike,
2752
        num_requests: int,
2753
        output_len: int | None = None,
2754
2755
2756
2757
2758
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
2759
    ) -> list[SampleRequest]:
2760
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2761
        sampled_requests = []
2762
        for i, prompt in enumerate(self.sample_prompts(n=num_requests)):
2763
            # apply template
2764
2765
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2766
                    [{"role": "user", "content": prompt}],
2767
2768
2769
                    add_generation_prompt=True,
                    tokenize=False,
                )
2770

2771
2772
2773
2774
2775
2776
            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2777
                    request_id=request_id_prefix + str(i),
2778
2779
2780
2781
2782
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2783
2784
        return sampled_requests

2785
2786
2787
2788
2789
2790
2791
2792
    def sample_prompts(self, n: int) -> Iterator[str]:
        for item in self.data.take(n):
            prompt = (
                f"{item['input']}\n\n{item['instruction']} Just output "
                "the code, do not include any explanation."
            )
            yield prompt

2793

2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
# -----------------------------------------------------------------------------
# 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,
2816
        tokenizer: TokenizerLike,
2817
        num_requests: int,
2818
        output_len: int | None = None,
2819
        enable_multimodal_chat: bool = False,
2820
        skip_chat_template: bool = False,
2821
        request_id_prefix: str = "",
2822
        no_oversample: bool = False,
2823
2824
        **kwargs,
    ) -> list:
2825
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2826
2827
        sampled_requests = []

2828
        for i, item in enumerate(self.data):
2829
2830
2831
2832
2833
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["turns"][0]

            # apply template
2834
2835
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2836
                    [{"role": "user", "content": prompt}],
2837
2838
2839
                    add_generation_prompt=True,
                    tokenize=False,
                )
2840
2841
2842
2843
2844
2845
2846

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2847
                    request_id=request_id_prefix + str(i),
2848
2849
2850
2851
2852
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2853
2854
2855
        return sampled_requests


2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
# -----------------------------------------------------------------------------
# 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,
2882
        tokenizer: TokenizerLike,
2883
        num_requests: int,
2884
        output_len: int | None = None,
2885
        skip_chat_template: bool = False,
2886
        request_id_prefix: str = "",
2887
        no_oversample: bool = False,
2888
2889
2890
2891
        min_distance: float = 0.0,
        max_distance: float = 1.0,
        **kwargs,
    ) -> list:
2892
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
        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
2905
2906

            # template copied from
2907
            # https://github.com/ise-uiuc/blazedit/blob/7765137e656fd62de877422d2e4cf8de51228054/dataset/create_refined_dataset.py#L94-L105 # noqa: E501
2908
            prompt = f"""Given a code file, please apply the change requests and generate the new file.
2909
2910
2911
2912
2913
2914
2915
2916
2917

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

Change request:
{change_request}

2918
Please generate the new code file in the "New file" section below."""  # noqa: E501
2919
2920

            # apply template
2921
2922
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2923
                    [{"role": "user", "content": prompt}],
2924
2925
2926
                    add_generation_prompt=True,
                    tokenize=False,
                )
2927
2928
2929
2930
2931
2932
2933
2934
2935

            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),
2936
2937
2938
2939
2940
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2941

2942
2943
2944
        return sampled_requests


2945
2946
2947
2948
2949
2950
2951
2952
2953
# -----------------------------------------------------------------------------
# AIMO Dataset Implementation
# -----------------------------------------------------------------------------


class AIMODataset(HuggingFaceDataset):
    """
    Dataset class for processing a AIMO dataset with reasoning questions.
    """
2954

2955
    SUPPORTED_DATASET_PATHS = {
2956
2957
2958
        "AI-MO/aimo-validation-aime",
        "AI-MO/NuminaMath-1.5",
        "AI-MO/NuminaMath-CoT",
2959
2960
    }

2961
2962
    def sample(
        self,
2963
        tokenizer: TokenizerLike,
2964
        num_requests: int,
2965
        output_len: int | None = None,
2966
2967
2968
2969
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2970
        sampled_requests = []
2971
        ind = 0
2972
2973
2974
2975
2976
        dynamic_output = output_len is None

        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
2977
            prompt, completion = item["problem"], item["solution"]
2978
2979
2980
2981
2982
2983
2984

            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
2985
2986
2987
            if dynamic_output and not is_valid_sequence(
                prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000
            ):
2988
2989
2990
2991
2992
2993
2994
                continue
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=None,
2995
                    request_id=request_id_prefix + str(ind),
2996
2997
                )
            )
2998
            ind += 1
2999
3000
3001
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3002
        return sampled_requests
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022


# -----------------------------------------------------------------------------
# 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:

3023
"""  # noqa: E501
3024
3025
3026


def _format_zeta_prompt(
3027
3028
    sample: dict, original_start_marker: str = "<|editable_region_start|>"
) -> dict:
3029
    """Format the zeta prompt for the Next Edit Prediction (NEP) dataset.
3030
3031
3032

    This function formats examples from the NEP dataset
    into prompts and expected outputs. It could be
3033
    further extended to support more NEP datasets.
3034

3035
    Args:
3036
        sample: The dataset sample containing events,
3037
            inputs, and outputs.
3038
3039
        original_start_marker: The marker indicating the
            start of the editable region. Defaults to
3040
            "<|editable_region_start|>".
3041

3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
    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,
    }

3071
3072
    def sample(
        self,
3073
        tokenizer: TokenizerLike,
3074
3075
3076
3077
3078
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ):
3079
        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.hf_name)
3080
        if formatting_prompt_func is None:
3081
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")
3082
        samples = []
3083
        for i, sample in enumerate(self.data):
3084
3085
3086
3087
3088
3089
            sample = formatting_prompt_func(sample)
            samples.append(
                SampleRequest(
                    prompt=sample["prompt"],
                    prompt_len=len(tokenizer(sample["prompt"]).input_ids),
                    expected_output_len=len(
3090
3091
                        tokenizer(sample["expected_output"]).input_ids
                    ),
3092
                    request_id=request_id_prefix + str(i),
3093
3094
                )
            )
3095
3096
            if len(samples) >= num_requests:
                break
3097
3098
3099
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
3100
        return samples
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135


# -----------------------------------------------------------------------------
# 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",
    }

3136
    DEFAULT_OUTPUT_LEN = 1024
3137
3138
3139
3140
    IS_MULTIMODAL = True

    def sample(
        self,
3141
        tokenizer: TokenizerLike,
3142
        num_requests: int,
3143
        output_len: int | None = None,
3144
        request_id_prefix: str = "",
3145
        no_oversample: bool = False,
3146
3147
        **kwargs,
    ) -> list:
3148
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
3149
3150
3151
3152
        if "openai" in tokenizer.name_or_path:
            prompt = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>"
        else:
            prompt = ""
3153
3154
        prompt_len = len(tokenizer(prompt).input_ids)
        sampled_requests = []
3155
        ind = 0
3156
        skipped = 0
3157
3158
3159
        asr_min_audio_len_sec = kwargs.get("asr_min_audio_len_sec")
        asr_max_audio_len_sec = kwargs.get("asr_max_audio_len_sec")
        durations = []
3160
3161
3162
3163
3164
3165
        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)
3166
            if duration_s < asr_min_audio_len_sec or duration_s > asr_max_audio_len_sec:
3167
3168
3169
                skipped += 1
                continue

3170
            durations.append(duration_s)
3171
3172
3173
3174
3175
3176
3177
            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,
3178
                    request_id=request_id_prefix + str(ind),
3179
3180
                )
            )
3181
            ind += 1
3182
3183
3184
3185
3186
3187
3188
        if skipped:
            logger.warning(
                "%d samples discarded from dataset due to"
                " their length being greater than"
                " what Whisper supports.",
                skipped,
            )
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202

        logger.info("Number of audio samples: %d", len(durations))
        avg_duration = sum(durations) / len(durations) if durations else 0
        min_duration = min(durations) if durations else 0
        max_duration = max(durations) if durations else 0
        median_duration = np.median(durations) if durations else 0
        logger.info(
            "Audio duration statistics (s): avg=%.2f, min=%.2f, max=%.2f, median=%.2f",
            avg_duration,
            min_duration,
            max_duration,
            median_duration,
        )

3203
3204
3205
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3206
        return sampled_requests
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238


# -----------------------------------------------------------------------------
# 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,
3239
        tokenizer: TokenizerLike,
3240
        num_requests: int,
3241
        output_len: int | None = None,
3242
        request_id_prefix: str = "",
3243
        no_oversample: bool = False,
3244
3245
3246
3247
3248
        **kwargs,
    ) -> list[SampleRequest]:
        # Force dynamic output length based on reference completion.
        dynamic_output = output_len is None
        sampled_requests: list[SampleRequest] = []
3249
        ind = 0
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283

        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,
3284
                    request_id=request_id_prefix + str(ind),
3285
3286
                )
            )
3287
            ind += 1
3288

3289
3290
3291
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3292
        return sampled_requests
3293
3294
3295
3296
3297
3298
3299
3300


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


class PrefixRepetitionRandomDataset(BenchmarkDataset):
3301
    # Default values copied from benchmark_serving.py for the repeated prefix
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
    # 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,
3318
        tokenizer: TokenizerLike,
3319
3320
3321
3322
3323
        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,
3324
        request_id_prefix: str = "",
3325
        no_oversample: bool = False,
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
        **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
3340
            tokens = np.random.randint(0, vocab_size, size=target_length).tolist()
3341

3342
            _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(  # noqa: E501
3343
3344
3345
3346
3347
3348
                tokenizer=tokenizer,
                token_sequence=tokens,
                target_token_len=target_length,
                add_special_tokens=False,
            )
            return adjusted_tokens, token_mismatch
3349
3350

        requests = []
3351
        token_mismatch_total = 0
3352
        for _ in range(num_prefixes):
3353
3354
            prefix_tokens, prefix_mismatch = _generate_exact_length_tokens(prefix_len)
            token_mismatch_total += prefix_mismatch
3355
3356

            for _ in range(prompts_per_prefix):
3357
                suffix_tokens, suffix_mismatch = _generate_exact_length_tokens(
3358
                    suffix_len
3359
                )
3360
                token_mismatch_total += suffix_mismatch
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
                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,
                    )
                )

3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
        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,
            )
3382
3383
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(requests)
3384
        return requests
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396


# -----------------------------------------------------------------------------
# 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
    """
3397

3398
3399
3400
3401
3402
3403
    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {"Lin-Chen/MMStar"}
    IS_MULTIMODAL = True

    def sample(
        self,
3404
        tokenizer: TokenizerLike,
3405
        num_requests: int,
3406
        output_len: int | None = None,
3407
3408
3409
3410
3411
3412
        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.
3413
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
        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