datasets.py 112 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 typing import Any, cast
31
32
33
34

import numpy as np
from PIL import Image
from transformers import PreTrainedTokenizerBase
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.transformers_utils.tokenizer import AnyTokenizer
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
from vllm.utils import PlaceholderModule

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")
59

60
61
62
63
64
try:
    from vllm.utils import FlexibleArgumentParser
except ImportError:
    from argparse import ArgumentParser as FlexibleArgumentParser

65
66
67
68
69
70
71
72
73
74
75
76
77
logger = logging.getLogger(__name__)

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


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

78
    prompt: str | list[str]
79
80
    prompt_len: int
    expected_output_len: int
81
82
83
    multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None
    lora_request: LoRARequest | None = None
    request_id: str | None = None
84
85
86
87
88
89
90
91
92


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


class BenchmarkDataset(ABC):
    DEFAULT_SEED = 0
93
    IS_MULTIMODAL = False
94
95
96

    def __init__(
        self,
97
        dataset_path: str | None = None,
98
        random_seed: int = DEFAULT_SEED,
99
100
        disable_shuffle: bool = False,
        **kwargs,
101
102
103
    ) -> None:
        """
        Initialize the BenchmarkDataset with an optional dataset path and random
104
105
        seed.

106
107
        Args:
            dataset_path (Optional[str]): Path to the dataset. If None, it
108
                indicates that a default or random dataset might be used.
109
            random_seed (int): Seed value for reproducible shuffling or
110
                sampling. Defaults to DEFAULT_SEED.
111
112
113
114
        """
        self.dataset_path = dataset_path
        # Set the random seed, ensuring that a None value is replaced with the
        # default seed.
115
        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
116
        self.disable_shuffle = disable_shuffle
117
118
119
        self.data = None

    def apply_multimodal_chat_transformation(
120
121
        self,
        prompt: str,
122
        mm_content: MultiModalDataDict | dict | list[dict] | None = None,
123
    ) -> list[dict]:
124
125
126
127
128
129
130
        """
        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:
131
132
133
134
135
            if isinstance(mm_content, list):
                content.extend(cast(list[dict[str, Any]], mm_content))
            elif isinstance(mm_content, dict):
                content.append(mm_content)
            else:
136
                raise TypeError(
137
138
                    "Could not process multimodal content of type: "
                    + f"{type(mm_content)}"
139
                )
140
141
142
143
144
145
146
147
148
149
150
151
152
        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
153
        raise NotImplementedError("load_data must be implemented in subclasses.")
154
155
156

    def get_random_lora_request(
        self,
157
158
159
        max_loras: int | None = None,
        lora_path: str | None = None,
    ) -> LoRARequest | None:
160
        """
161
        Optionally select a random LoRA request.
162
163

        This method is used when LoRA parameters are provided.  It randomly
164
        selects a LoRA based on max_loras.
165
166

        Args:
167
168
169
170
            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.
171
172

        Returns:
173
174
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
175
176
        """
        if max_loras is None or lora_path is None:
177
            return None
178
179
180
181
182
183
184
185

        # 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),
        )
186
        return lora_request
187
188

    @abstractmethod
189
190
191
192
193
194
195
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
    ) -> list[SampleRequest]:
196
197
198
199
200
201
202
203
        """
        Abstract method to generate sample requests from the dataset.

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

        Args:
            tokenizer (PreTrainedTokenizerBase): The tokenizer to be used
204
                for processing the dataset's text.
205
            num_requests (int): The number of sample requests to generate.
206
            request_id_prefix (str): The prefix of request_id.
207
208
209
210
211
212
213

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

214
215
216
217
218
    def maybe_oversample_requests(
        self,
        requests: list[SampleRequest],
        num_requests: int,
        request_id_prefix: str = "",
219
        no_oversample: bool = False,
220
    ) -> None:
221
222
223
224
225
226
        """
        Oversamples the list of requests if its size is less than the desired
        number.

        Args:
            requests (List[SampleRequest]): The current list of sampled
227
228
                requests.
            num_requests (int): The target number of requests.
229
230
            request_id_prefix (str): The prefix applied to generated request
                identifiers.
231

232
        """
233
        if no_oversample:
234
            logger.info("Skipping oversampling. Total samples: %d.", len(requests))
235
236
            return

237
238
        if len(requests) < num_requests:
            random.seed(self.random_seed)
239
240
241
242
            needed = num_requests - len(requests)
            additional = []
            for i in range(needed):
                req = deepcopy(random.choice(requests))
243
                req.request_id = request_id_prefix + str(len(requests) + i)
244
                additional.append(req)
245
            requests.extend(additional)
246
            logger.info("Oversampled requests to reach %d total samples.", num_requests)
247

248
249
        ids = [req.request_id for req in requests]
        if len(ids) != len(set(ids)):
250
251
252
253
254
            raise ValueError(
                "Duplicate request_id found in the sampled "
                "requests. Please ensure that each request_id "
                "is unique."
            )
255

256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278

# -----------------------------------------------------------------------------
# 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
279
    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
280
281
282
283
    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
284
285
286
    return not (
        prompt_too_short or output_too_short or prompt_too_long or combined_too_long
    )
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301


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


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


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

302
    Supports the following input types:
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317

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

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

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

    Raises:
        ValueError: If the input is not a supported type.
    """
318
319
    if isinstance(image, dict) and "bytes" in image:
        image = Image.open(BytesIO(image["bytes"]))
320
    if isinstance(image, Image.Image):
321
        image = convert_image_mode(image, "RGB")
322
323
        with io.BytesIO() as image_data:
            image.save(image_data, format="JPEG")
324
            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
325
326
        return {
            "type": "image_url",
327
            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
328
329
330
        }

    if isinstance(image, str):
331
332
333
334
335
        image_url = (
            image
            if image.startswith(("http://", "https://", "file://"))
            else f"file://{image}"
        )
336
337
        return {"type": "image_url", "image_url": {"url": image_url}}

338
339
340
341
    raise ValueError(
        f"Invalid image input {image}. Must be a PIL.Image.Image"
        " or str or dictionary with raw image bytes."
    )
342
343


344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
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.
    """
360
361
    if isinstance(video, dict) and "bytes" in video:
        video_bytes = video["bytes"]
362
363
364
        video_base64 = base64.b64encode(video_bytes).decode("utf-8")
        return {
            "type": "video_url",
365
            "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
366
367
368
        }

    if isinstance(video, str):
369
370
371
372
373
        video_url = (
            video
            if video.startswith(("http://", "https://", "file://"))
            else f"file://{video}"
        )
374
375
376
377
378
379
        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
    )

380
381
382
383
384
385
386

def gen_prompt_decode_to_target_len(
    tokenizer: PreTrainedTokenizerBase,
    token_sequence: list[int],
    target_token_len: int,
    max_retry: int = 10,
    add_special_tokens: bool = False,
387
    rng: np.random.Generator | None = None,
388
389
390
391
392
) -> tuple[str, list[int]]:
    """
    Ensure decoded-then-encoded prompt length matches the target token length.

    This function decodes an initial token sequence to text and re-encodes it
393
394
    , iteratively adjusting the token sequence length to match a target.
    This is necessary because some tokenizers do not guarantee a 1:1 mapping
395
396
397
398
399
400
401
402
403
404
405
    between consecutive tokens and the decoded-then-encoded sequence length.
    For example, for GPT2Tokenizer:
    [6880, 6881] -> ['Ġcalls', 'here'] ->
    [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']

    Returns a tuple of the final prompt string and the adjusted token sequence.
    """
    remain_num_try = max_retry
    token_mismatch = 0
    while True:
        prompt = tokenizer.decode(token_sequence)
406
        token_sequence = tokenizer.encode(prompt, add_special_tokens=add_special_tokens)
407
408
409
410
        if remain_num_try <= 0:
            if len(token_sequence) != target_token_len:
                token_mismatch = len(token_sequence) - target_token_len
            break
411

412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
        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

435

436
437
438
439
# -----------------------------------------------------------------------------
# Random Dataset Implementation (Synthetic Data)
# -----------------------------------------------------------------------------

440

441
class RandomDataset(BenchmarkDataset):
442
443
444
445
446
447
448
449
450
451
452
453
    """
    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.
    """
454

455
456
457
458
459
460
    # 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

461
    def __init__(self, **kwargs) -> None:
462
        super().__init__(**kwargs)
463
464
465
466
        # 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)
467
468
469
470
471

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
472
        request_id_prefix: str = "",
473
        no_oversample: bool = False,
474
475
476
477
        prefix_len: int = DEFAULT_PREFIX_LEN,
        range_ratio: float = DEFAULT_RANGE_RATIO,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
478
        batchsize: int = 1,
479
480
        **kwargs,
    ) -> list[SampleRequest]:
481
482
        input_lens, output_lens, offsets = self.get_sampling_params(
            num_requests, range_ratio, input_len, output_len, tokenizer
483
484
        )

485
486
        # Generate prefix once
        prefix_token_ids = self.get_prefix(tokenizer, prefix_len)
487
488
        vocab_size = tokenizer.vocab_size

489
        requests = []
490
        token_mismatch_total = 0
491
        for i in range(num_requests):
492
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
493
494
495
496
497
498
499
500
                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,
            )
501
            token_mismatch_total += token_mismatch
502
503
504
505
506
507
508
509
            requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
                    request_id=request_id_prefix + str(i),
                )
            )
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
        # 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
525

526
527
528
529
530
531
532
533
534
535
536
        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,
            )

537
538
539
540
541
542
543
544
545
        return requests

    def get_prefix(
        self, tokenizer: PreTrainedTokenizerBase, prefix_len: int
    ) -> list[int]:
        """
        Get the prefix for the dataset.
        """
        return (
546
            self._rng.integers(0, tokenizer.vocab_size, size=prefix_len).tolist()
547
548
549
            if prefix_len > 0
            else []
        )
550

551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
    def get_sampling_params(
        self,
        num_requests: int,
        range_ratio: float,
        input_len: int,
        output_len: int,
        tokenizer: PreTrainedTokenizerBase,
    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
        """
        Get the sampling parameters for the dataset.
        """
        # Enforce range_ratio < 1
        if not (0.0 <= range_ratio < 1.0):
            raise ValueError("range_ratio must be in [0, 1).")
        num_special_tokens = int(tokenizer.num_special_tokens_to_add())
        real_input_len = max(0, int(input_len) - num_special_tokens)
        # Bounds use floor for low and ceil for high
        input_low = math.floor(real_input_len * (1 - range_ratio))
        input_high = math.ceil(real_input_len * (1 + range_ratio))
        output_low = math.floor(output_len * (1 - range_ratio))
        output_high = math.ceil(output_len * (1 + range_ratio))
572
573
        # Ensure the lower bound for output length is at least 1 to
        # prevent sampling 0 tokens.
574
        output_low = max(output_low, 1)
575
        output_high = max(output_high, 1)
576
577
578

        if input_low > input_high:
            raise ValueError(
579
                f"Invalid input sampling interval: low={input_low} > high={input_high}"
580
581
582
583
584
585
            )
        if output_low > output_high:
            raise ValueError(
                "Invalid output sampling interval: "
                f"low={output_low} > high={output_high}"
            )
586

587
588
        logger.info(
            "Sampling input_len from [%s, %s] and output_len from [%s, %s]",
589
590
591
592
593
            input_low,
            input_high,
            output_low,
            output_high,
        )
594

595
596
597
        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)
598
        return input_lens, output_lens, offsets
599

600
601
602
603
604
605
606
607
608
609
    def generate_token_sequence(
        self,
        *,
        tokenizer: PreTrainedTokenizerBase,
        prefix_token_ids: list[int],
        prefix_len: int,
        vocab_size: int,
        input_len: int,
        offset: int,
        index: int,
610
    ) -> tuple[str, int, int]:
611
612
613
614
615
616
617
618
619
620
        """
        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,
621
        the encoded sequence is truncated before being decoded again.
622
623
        """
        # Build the inner sequence by sampling sequentially from the vocab
624
        inner_seq = ((offset + index + np.arange(input_len)) % vocab_size).tolist()
625
626
627
628
        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)
629
        prompt, adjusted_token_sequence, token_mismatch = (
630
            gen_prompt_decode_to_target_len(
631
632
633
634
635
636
                tokenizer=tokenizer,
                token_sequence=token_sequence,
                target_token_len=total_input_len,
                add_special_tokens=False,
                rng=self._rng,
            )
637
638
639
        )
        total_input_len = len(adjusted_token_sequence)
        return prompt, total_input_len, token_mismatch
640
641


642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
# -----------------------------------------------------------------------------
# Random Dataset Implementation (Synthetic Data)
# -----------------------------------------------------------------------------


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

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

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

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

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

        query_len = int(query_lens[0])

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

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

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

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

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

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

        return batch_requests


748
749
750
751
# -----------------------------------------------------------------------------
# MultiModalDataset Implementation
# -----------------------------------------------------------------------------

752

753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
class RandomMultiModalDataset(RandomDataset):
    """
    Synthetic multimodal dataset (text + images) that extends RandomDataset.

    Status:
    - Images: supported via synthetic RGB data.
    - Video: not yet supported (TODO: implement video generation method).
    - 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
768
769
770
       mapping (height, width, num_frames) → probability. We treat
       `num_frames`=1 as image and and `num_frames` > 1 as video.
       Entries with zero probability are removed and the rest are renormalized
771
772
773
774
775
776
777
       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}
778
779
      - Two image buckets (`num_frames`=1) and one video bucket
      (`num_frames`=16).
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
    OBS.: Only image sampling is supported for now.
    """

    IS_MULTIMODAL = True
    # NOTE: video sampling is WIP. Setting it to 0.
    DEFAULT_LIMIT_MM_PER_PROMPT = {"image": 255, "video": 0}

    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.
801
802
803

        NOTE: iid pixel sampling results in worst-case compression
        (good for stressing I/O), but very unlike real photos.
804
805
806
807
808
809
810
811
812
813
814
        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)

815
    def generate_synthetic_video(self, width: int, height: int, num_frames: int) -> Any:
816
        """Generate synthetic video with random values.
817

818
819
820
821
822
823
824
825
826
827
828
829
830
        TODO: Finish this method.
        """
        raise NotImplementedError("Video sampling is WIP.")

    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}")

831
832
833
    def normalize_bucket_config(
        self, bucket_config: dict[tuple[int, int, int], float]
    ) -> dict[tuple[int, int, int], float]:
834
835
836
837
838
839
840
841
842
843
844
        """
        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:
845
846
847
            raise ValueError(
                "Got invalid bucket config. Bucket config values must be non-zero."
            )
848
849
850
851
        # 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()}

852
853
854
855
    def generate_mm_item(
        self,
        mm_item_config: tuple[int, int, int],
    ) -> Mapping[str, Any]:
856
        """
857
        Create synthetic images and videos and
858
859
860
861
        apply process_image/process_video respectively.
        This follows the OpenAI API chat completions
        https://github.com/openai/openai-python
        """
862

863
        if self.map_config_to_modality(mm_item_config) == "image":
864
865
866
            return process_image(
                self.generate_synthetic_image(mm_item_config[1], mm_item_config[0])
            )
867
        elif self.map_config_to_modality(mm_item_config) == "video":
868
869
870
871
872
            return process_video(
                self.generate_synthetic_video(
                    mm_item_config[1], mm_item_config[0], mm_item_config[2]
                )
            )
873
        else:
874
            raise ValueError(f"Invalid multimodal item configuration: {mm_item_config}")
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894

    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:
895
896
897
898
899
                raise ValueError(
                    f"Modality {modality} is not in "
                    f"limit_mm_per_prompt: "
                    f"{limit_mm_per_prompt.keys()}"
                )
900

901
        # Remove zero probability entries
902
903
904
        # and normalize bucket config to sum to 1
        bucket_config = self.normalize_bucket_config(bucket_config)
        logger.info(
905
906
            "Normalized bucket config: %s",
            bucket_config,
907
908
        )
        # Only consider limit per prompt for modalities in bucket config
909
        allowed_modalities = {self.map_config_to_modality(cfg) for cfg in bucket_config}
910
        limit_mm_per_prompt = {
911
912
            k: v for k, v in limit_mm_per_prompt.items() if k in allowed_modalities
        }
913
        if not limit_mm_per_prompt:
914
            raise ValueError("No valid limits for modalities present in bucket_config.")
915
916

        logger.info(
917
918
            "Updated mm-limit-per-prompt: %s",
            limit_mm_per_prompt,
919
920
921
922
923
        )

        # 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(
924
            sum(limit_mm_per_prompt.values()),
925
            math.ceil(base_items_per_request * (1 + num_mm_items_range_ratio)),
926
927
928
        )
        # Ensure min num mm items is at least 0
        min_num_mm_items = max(
929
            0, math.floor(base_items_per_request * (1 - num_mm_items_range_ratio))
930
931
932
        )
        # Raise error if min num mm items is greater than max num mm items
        if min_num_mm_items > max_num_mm_items:
933
934
935
936
            raise ValueError(
                f"Min num mm items is greater than max mm items: "
                f"{min_num_mm_items} > {max_num_mm_items}"
            )
937

938
939
        logger.info(
            "Sampling number of multimodal items from [%s, %s]",
940
941
            min_num_mm_items,
            max_num_mm_items,
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
        )

        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],
957
    ) -> Iterator[tuple[int, int, int]]:
958
959
960
961
962
        """
        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.
963
964
        Loop until the number of multimodal items sampled is equal to
        request_num_mm_items or limit of multimodal items per prompt
965
966
967
968
969
970
971
972
973
974
975
        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)
976
        )
977
978
979
980
        # If request_num_mm_items is 0, yield an empty iterator
        if request_num_mm_items == 0:
            return
        # Initialize modality counters
981
        modality_counter = {self.map_config_to_modality(k): 0 for k in bucket_config}
982
983
984
985
986
        # 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
987
988
989
            mm_item_config = self._rng.choice(
                list(bucket_config_copy.keys()), p=list(bucket_config_copy.values())
            )
990
991
992
993
            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
994
                yield (mm_item_config)
995
996
997
998
999
1000
1001
1002
1003
1004
            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()):
1005
1006
1007
                    logger.warning(
                        "Exhausted all multimodal items of modality %s", modality
                    )
1008
1009
                    break
                # Renormalize the bucket config
1010
                bucket_config_copy = self.normalize_bucket_config(bucket_config_copy)
1011
1012
1013
1014
1015
1016

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
1017
        no_oversample: bool = False,
1018
1019
1020
1021
1022
1023
1024
        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,
1025
1026
1027
        bucket_config: dict[
            tuple[int, int, int], float
        ] = DEFAULT_MM_ITEM_BUCKET_CONFIG,
1028
1029
1030
1031
1032
        enable_multimodal_chat: bool = DEFAULT_ENABLE_MULTIMODAL_CHAT,
        **kwargs,
    ) -> list[SampleRequest]:
        # NOTE: Video sampling is WIP. Raise error if video is in bucket config
        # and probability is non-zero.
1033
1034
1035
1036
1037
1038
1039
        if any(
            self.map_config_to_modality(cfg) == "video" and p > 0
            for cfg, p in bucket_config.items()
        ):
            raise NotImplementedError(
                "Video sampling not implemented; set its probability to 0."
            )
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062

        # 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,
        )

        # Generate prefix once
        prefix_token_ids = self.get_prefix(tokenizer, prefix_len)
        vocab_size = tokenizer.vocab_size
        # Add synthetic multimodal items to each request
        mm_requests = []
1063
        token_mismatch_total = 0
1064
        for i in range(num_requests):
1065
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
1066
1067
1068
1069
1070
1071
1072
1073
                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,
            )
1074
            token_mismatch_total += token_mismatch
1075
1076
1077
1078
1079
1080
1081
1082
            # 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,
            )

1083
1084
1085
1086
1087
1088
1089
            mm_content = cast(
                list[dict[str, Any]],
                [
                    self.generate_mm_item(mm_item_config)
                    for mm_item_config in mm_item_iterator
                ],
            )
1090
1091

            if enable_multimodal_chat:
1092
                # NOTE: For now this option is only provided for completeness
1093
1094
1095
                # given that the serve.py benchmark currently does not use it.
                mm_chat_prompt: Any = prompt
                mm_chat_prompt = self.apply_multimodal_chat_transformation(
1096
1097
                    prompt, mm_content
                )
1098
1099
1100
1101
1102
1103
1104
1105
1106
                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(
1107
1108
1109
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
1110
                    multi_modal_data=mm_content,
1111
                    request_id=request_id_prefix + str(i),
1112
1113
                )
            mm_requests.append(sample_request)
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125

        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,
            )

1126
        return mm_requests
1127

1128

1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
# -----------------------------------------------------------------------------
# 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 = [
1152
1153
            entry
            for entry in self.data
1154
1155
1156
            if "conversations" in entry and len(entry["conversations"]) >= 2
        ]
        random.seed(self.random_seed)
1157
1158
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1159
1160
1161
1162
1163

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
1164
1165
1166
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1167
        enable_multimodal_chat: bool = False,
1168
        request_id_prefix: str = "",
1169
        no_oversample: bool = False,
1170
1171
1172
        **kwargs,
    ) -> list:
        samples: list = []
1173
        ind = 0
1174
1175
1176
1177
1178
1179
1180
1181
        for entry in self.data:
            if len(samples) >= num_requests:
                break
            prompt, completion = (
                entry["conversations"][0]["value"],
                entry["conversations"][1]["value"],
            )

1182
            lora_request = self.get_random_lora_request(
1183
1184
                max_loras=max_loras, lora_path=lora_path
            )
1185
1186
1187
            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
1188
1189
1190
1191
1192
1193
            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,
            ):
1194
                continue
1195
1196
1197
            if image_path := entry.get("image"):
                mm_content = process_image(image_path)
            elif video_path := entry.get("video"):
1198
                mm_content = process_video(video_path)
1199
            else:
1200
                mm_content = None
1201
            if enable_multimodal_chat:
1202
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
1203
1204
1205
1206
1207
1208
            samples.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=new_output_len,
                    lora_request=lora_request,
1209
                    multi_modal_data=mm_content,
1210
                    request_id=request_id_prefix + str(ind),
1211
1212
                )
            )
1213
            ind += 1
1214
1215
1216
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
1217
1218
1219
        return samples


1220
1221
class _ValidateDatasetArgs(argparse.Action):
    """Argparse action to validate dataset name and path compatibility."""
1222

1223
1224
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, values)
1225

1226
        # Get current values of both dataset_name and dataset_path
1227
1228
        dataset_name = getattr(namespace, "dataset_name", "random")
        dataset_path = getattr(namespace, "dataset_path", None)
1229

1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
        # 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}"
            )


1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
def add_dataset_parser(parser: FlexibleArgumentParser):
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
        "--num-prompts",
        type=int,
        default=1000,
        help="Number of prompts to process.",
    )
    parser.add_argument(
        "--dataset-name",
        type=str,
        default="random",
1252
        action=_ValidateDatasetArgs,
1253
        choices=[
1254
1255
1256
1257
1258
            "sharegpt",
            "burstgpt",
            "sonnet",
            "random",
            "random-mm",
1259
            "random-rerank",
1260
1261
1262
1263
            "hf",
            "custom",
            "prefix_repetition",
            "spec_bench",
1264
        ],
1265
1266
        help="Name of the dataset to benchmark on.",
    )
1267
1268
1269
1270
1271
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
1272
1273
1274
1275
    parser.add_argument(
        "--dataset-path",
        type=str,
        default=None,
1276
        action=_ValidateDatasetArgs,
1277
1278
1279
        help="Path to the sharegpt/sonnet dataset. "
        "Or the huggingface dataset ID if using HF dataset.",
    )
1280
1281
1282
    parser.add_argument(
        "--no-oversample",
        action="store_true",
1283
        help="Do not oversample if the dataset has fewer samples than num-prompts.",
1284
    )
1285
1286
1287
    parser.add_argument(
        "--skip-chat-template",
        action="store_true",
1288
        help="Skip applying chat template to prompt for datasets that support it.",
1289
    )
1290
1291
1292
1293
1294
    parser.add_argument(
        "--disable-shuffle",
        action="store_true",
        help="Disable shuffling of dataset samples for deterministic ordering.",
    )
1295
1296
1297
1298
1299
1300
1301

    # 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,
1302
        help="Number of output tokens per request, used only for custom dataset.",
1303
1304
    )

1305
1306
1307
1308
1309
    spec_bench_group = parser.add_argument_group("spec bench dataset options")
    spec_bench_group.add_argument(
        "--spec-bench-output-len",
        type=int,
        default=256,
1310
        help="Num of output tokens per request, used only for spec bench dataset.",
1311
1312
1313
1314
1315
    )
    spec_bench_group.add_argument(
        "--spec-bench-category",
        type=str,
        default=None,
1316
        help="Category for spec bench dataset. If None, use all categories.",
1317
1318
    )

1319
1320
1321
1322
1323
    sonnet_group = parser.add_argument_group("sonnet dataset options")
    sonnet_group.add_argument(
        "--sonnet-input-len",
        type=int,
        default=550,
1324
        help="Number of input tokens per request, used only for sonnet dataset.",
1325
1326
1327
1328
1329
    )
    sonnet_group.add_argument(
        "--sonnet-output-len",
        type=int,
        default=150,
1330
        help="Number of output tokens per request, used only for sonnet dataset.",
1331
1332
1333
1334
1335
    )
    sonnet_group.add_argument(
        "--sonnet-prefix-len",
        type=int,
        default=200,
1336
        help="Number of prefix tokens per request, used only for sonnet dataset.",
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
    )

    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.",
    )

1348
1349
1350
1351
1352
    blazedit_group = parser.add_argument_group("blazedit dataset options")
    blazedit_group.add_argument(
        "--blazedit-min-distance",
        type=float,
        default=0.0,
1353
        help="Minimum distance for blazedit dataset. Min: 0, Max: 1.0",
1354
1355
1356
1357
1358
    )
    blazedit_group.add_argument(
        "--blazedit-max-distance",
        type=float,
        default=1.0,
1359
        help="Maximum distance for blazedit dataset. Min: 0, Max: 1.0",
1360
1361
    )

1362
1363
1364
1365
1366
    random_group = parser.add_argument_group("random dataset options")
    random_group.add_argument(
        "--random-input-len",
        type=int,
        default=1024,
1367
        help="Number of input tokens per request, used only for random sampling.",
1368
1369
1370
1371
1372
    )
    random_group.add_argument(
        "--random-output-len",
        type=int,
        default=128,
1373
        help="Number of output tokens per request, used only for random sampling.",
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
    )
    random_group.add_argument(
        "--random-range-ratio",
        type=float,
        default=0.0,
        help="Range ratio for sampling input/output length, "
        "used only for random sampling. Must be in the range [0, 1) to define "
        "a symmetric sampling range"
        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
    )
    random_group.add_argument(
        "--random-prefix-len",
        type=int,
        default=0,
1388
1389
1390
1391
1392
1393
1394
1395
        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)]."
        ),
1396
    )
1397
1398
1399
1400
    random_group.add_argument(
        "--random-batch-size",
        type=int,
        default=1,
1401
        help=("Batch size for random sampling. Only used for embeddings benchmark."),
1402
    )
1403
1404
1405
1406
1407
1408
1409
1410
    random_group.add_argument(
        "--no-reranker",
        action="store_true",
        help=(
            "Whether the model supports reranking natively."
            " Only used for reranker benchmark."
        ),
    )
1411

1412
1413
    # random multimodal dataset options
    random_mm_group = parser.add_argument_group(
1414
1415
        "random multimodal dataset options extended from random dataset"
    )
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
    random_mm_group.add_argument(
        "--random-mm-base-items-per-request",
        type=int,
        default=RandomMultiModalDataset.DEFAULT_BASE_ITEMS_PER_REQUEST,
        help=(
            "Base number of multimodal items per request for random-mm. "
            "Actual per-request count is sampled around this base using "
            "--random-mm-num-mm-items-range-ratio."
        ),
    )
    random_mm_group.add_argument(
        "--random-mm-num-mm-items-range-ratio",
        type=float,
        default=RandomMultiModalDataset.DEFAULT_NUM_MM_ITEMS_RANGE_RATIO,
        help=(
            "Range ratio r in [0, 1] for sampling items per request. "
            "We sample uniformly from the closed integer range "
            "[floor(n*(1-r)), ceil(n*(1+r))] "
            "where n is the base items per request. "
            "r=0 keeps it fixed; r=1 allows 0 items. The maximum is clamped "
            "to the sum of per-modality limits from "
            "--random-mm-limit-mm-per-prompt. "
            "An error is raised if the computed min exceeds the max."
        ),
    )
    random_mm_group.add_argument(
        "--random-mm-limit-mm-per-prompt",
        type=json.loads,
        default=RandomMultiModalDataset.DEFAULT_LIMIT_MM_PER_PROMPT,
        help=(
            "Per-modality hard caps for items attached per request, e.g. "
1447
            '\'{"image": 3, "video": 0}\'. The sampled per-request item '
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
            "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)
1464
1465
1466
1467
1468
                if not (
                    isinstance(key, tuple)
                    and len(key) == 3
                    and all(isinstance(x, int) for x in key)
                ):
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
                    raise ValueError(
                        f"Invalid bucket key {k!r}. Expected tuple (H, W, T)."
                    )
                out[(int(key[0]), int(key[1]), int(key[2]))] = float(val)
            return out

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

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

1506
    hf_group = parser.add_argument_group("hf dataset options")
1507
1508
1509
1510
1511
1512
    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."
    )
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
    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."
        ),
    )
1523
1524
1525
1526
1527
1528
1529
1530
    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.",
    )

1531
    prefix_repetition_group = parser.add_argument_group(
1532
1533
        "prefix repetition dataset options"
    )
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
    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.",
    )

1563
1564

def get_samples(args, tokenizer) -> list[SampleRequest]:
1565
1566
1567
    if not hasattr(args, "request_id_prefix"):
        args.request_id_prefix = ""

1568
    if args.dataset_name == "custom":
1569
1570
1571
        dataset = CustomDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1572
1573
1574
1575
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
1576
            skip_chat_template=args.skip_chat_template,
1577
            request_id_prefix=args.request_id_prefix,
1578
            no_oversample=args.no_oversample,
1579
1580
1581
        )

    elif args.dataset_name == "sonnet":
1582
1583
1584
        dataset = SonnetDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1585
        # For the "sonnet" dataset, formatting depends on the backend.
1586
        if args.backend == "openai-chat":
1587
1588
1589
1590
1591
1592
1593
            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,
1594
                request_id_prefix=args.request_id_prefix,
1595
                no_oversample=args.no_oversample,
1596
1597
1598
            )
        else:
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
1599
1600
                "Tokenizer/model must have chat template for sonnet dataset."
            )
1601
1602
1603
1604
1605
1606
1607
            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,
1608
                request_id_prefix=args.request_id_prefix,
1609
                no_oversample=args.no_oversample,
1610
1611
1612
1613
1614
            )

    elif args.dataset_name == "hf":
        # all following datasets are implemented from the
        # HuggingFaceDataset base class
1615
        hf_kwargs = {}
1616
1617
1618
1619
        if (
            args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS
        ):
1620
1621
1622
            dataset_class = VisionArenaDataset
            args.hf_split = "train"
            args.hf_subset = None
1623
1624
1625
1626
1627
1628
1629
        elif (
            args.dataset_path in MMVUDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMVUDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMVUDataset
            args.hf_split = "validation"
            args.hf_subset = None
1630
1631
1632
1633
        elif (
            args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS
        ):
1634
1635
            dataset_class = InstructCoderDataset
            args.hf_split = "train"
1636
1637
1638
1639
        elif (
            args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MTBenchDataset.SUPPORTED_DATASET_PATHS
        ):
1640
1641
            dataset_class = MTBenchDataset
            args.hf_split = "train"
1642
1643
1644
1645
        elif (
            args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS
        ):
1646
            dataset_class = ConversationDataset
1647
1648
1649
1650
        elif (
            args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS
        ):
1651
1652
            dataset_class = AIMODataset
            args.hf_split = "train"
1653
        elif (
1654
            args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS  # noqa: E501
1655
1656
            or args.hf_name in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS
        ):
1657
1658
            dataset_class = NextEditPredictionDataset
            args.hf_split = "train"
1659
1660
1661
1662
        elif (
            args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS
        ):
1663
1664
            dataset_class = ASRDataset
            args.hf_split = "train"
1665
1666
1667
1668
1669
1670
1671
        elif args.dataset_path in BlazeditDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = BlazeditDataset
            args.hf_split = "train"
            hf_kwargs = {
                "min_distance": args.blazedit_min_distance,
                "max_distance": args.blazedit_max_distance,
            }
1672
1673
1674
1675
        elif (
            args.dataset_path in MLPerfDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MLPerfDataset.SUPPORTED_DATASET_PATHS
        ):
1676
1677
            dataset_class = MLPerfDataset
            args.hf_split = "train"
1678
1679
1680
1681
1682
1683
1684
        elif (
            args.dataset_path in MMStarDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMStarDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMStarDataset
            args.hf_split = "val"
            args.hf_subset = None
1685
        else:
1686
1687
1688
1689
1690
1691
1692
            supported_datasets = set(
                [
                    dataset_name
                    for cls in HuggingFaceDataset.__subclasses__()
                    for dataset_name in cls.SUPPORTED_DATASET_PATHS
                ]
            )
1693
1694
1695
1696
1697
            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 "
1698
1699
                "like to add support for additional dataset formats."
            )
1700

1701
1702
        if dataset_class.IS_MULTIMODAL and not (
            args.backend in ("openai-chat", "openai-audio")
1703
            or "embeddings-" in args.backend
1704
        ):
1705
1706
            # multi-modal benchmark is only available on OpenAI Chat
            # endpoint-type.
1707
1708
            raise ValueError(
                "Multi-modal content is only supported on 'openai-chat' and "
1709
1710
                "'openai-audio' backends."
            )
1711
1712
1713
1714
1715
        input_requests = dataset_class(
            dataset_path=args.dataset_path,
            dataset_subset=args.hf_subset,
            dataset_split=args.hf_split,
            random_seed=args.seed,
1716
            no_stream=args.no_stream,
1717
            hf_name=args.hf_name,
1718
            disable_shuffle=args.disable_shuffle,
1719
1720
1721
1722
        ).sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.hf_output_len,
1723
            request_id_prefix=args.request_id_prefix,
1724
            no_oversample=args.no_oversample,
1725
            skip_chat_template=args.skip_chat_template,
1726
            **hf_kwargs,
1727
1728
1729
1730
1731
        )

    else:
        # For datasets that follow a similar structure, use a mapping.
        dataset_mapping = {
1732
            "spec_bench": lambda: SpecBench(
1733
1734
1735
                dataset_path=args.dataset_path,
                category=args.spec_bench_category,
                disable_shuffle=args.disable_shuffle,
1736
            ).sample(
1737
1738
1739
1740
                num_requests=args.num_prompts,
                tokenizer=tokenizer,
                output_len=args.spec_bench_output_len,
                request_id_prefix=args.request_id_prefix,
1741
                no_oversample=args.no_oversample,
1742
            ),
1743
            "sharegpt": lambda: ShareGPTDataset(
1744
1745
1746
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1747
1748
1749
1750
1751
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
                request_id_prefix=args.request_id_prefix,
1752
                no_oversample=args.no_oversample,
1753
1754
            ),
            "burstgpt": lambda: BurstGPTDataset(
1755
1756
1757
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1758
1759
1760
1761
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                request_id_prefix=args.request_id_prefix,
1762
                no_oversample=args.no_oversample,
1763
1764
            ),
            "random": lambda: RandomDataset(
1765
1766
1767
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1768
            ).sample(
1769
1770
1771
1772
1773
1774
                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,
1775
                request_id_prefix=args.request_id_prefix,
1776
                batchsize=args.random_batch_size,
1777
                no_oversample=args.no_oversample,
1778
            ),
1779
            "random-mm": lambda: RandomMultiModalDataset(
1780
1781
1782
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                prefix_len=args.random_prefix_len,
                range_ratio=args.random_range_ratio,
                input_len=args.random_input_len,
                output_len=args.random_output_len,
                base_items_per_request=args.random_mm_base_items_per_request,
                limit_mm_per_prompt=args.random_mm_limit_mm_per_prompt,
                num_mm_items_range_ratio=args.random_mm_num_mm_items_range_ratio,
                bucket_config=args.random_mm_bucket_config,
                request_id_prefix=args.request_id_prefix,
1795
                no_oversample=args.no_oversample,
1796
            ),
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
            "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,
            ),
1810
            "prefix_repetition": lambda: PrefixRepetitionRandomDataset(
1811
1812
1813
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1814
1815
1816
1817
1818
1819
1820
            ).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,
1821
                request_id_prefix=args.request_id_prefix,
1822
                no_oversample=args.no_oversample,
1823
            ),
1824
1825
1826
        }

        try:
1827
            # Enforce endpoint compatibility for multimodal datasets.
1828
            if args.dataset_name == "random-mm" and args.backend not in ["openai-chat"]:
1829
1830
1831
1832
                raise ValueError(
                    "Multi-modal content (images) is only supported on "
                    "'openai-chat' backend."
                )
1833
1834
1835
1836
1837
1838
1839
            input_requests = dataset_mapping[args.dataset_name]()
        except KeyError as err:
            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err

    return input_requests


1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
# -----------------------------------------------------------------------------
# Custom Dataset Implementation
# -----------------------------------------------------------------------------


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

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

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

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

        # Load the JSONL file
        if self.dataset_path.endswith(".jsonl"):
1873
            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886

            # 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(
1887
1888
                "Only JSONL format is supported for CustomDataset."
            )
1889
1890

        random.seed(self.random_seed)
1891
1892
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1893
1894
1895
1896
1897

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
1898
1899
1900
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1901
1902
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
1903
        request_id_prefix: str = "",
1904
        no_oversample: bool = False,
1905
1906
        **kwargs,
    ) -> list:
1907
1908
1909
1910
        # load all data if needed
        self.num_available_samples = len(self.data)
        if num_requests <= 0:
            num_requests = self.num_available_samples
1911
1912
1913
1914
1915
            logger.info(
                "num_requests is set to 0 or negative, "
                "so using all available samples: %d",
                num_requests,
            )
1916

1917
        sampled_requests = []
1918
        for i, item in enumerate(self.data):
1919
1920
1921
1922
1923
1924
1925
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["prompt"]

            # apply template
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
1926
                    [{"role": "user", "content": prompt}],
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
                    add_generation_prompt=True,
                    tokenize=False,
                )

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
1937
                    request_id=request_id_prefix + str(i),
1938
1939
1940
1941
1942
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
1943
1944
1945
1946

        return sampled_requests


1947
1948
1949
1950
1951
1952
1953
1954
# -----------------------------------------------------------------------------
# Spec Bench Dataset Implementation
# -----------------------------------------------------------------------------


class SpecBench(CustomDataset):
    """
    Implements the SpecBench dataset: https://github.com/hemingkx/Spec-Bench
1955
    Download the dataset using:
1956
    wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
1957
    """  # noqa: E501
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970

    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
1971
        jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
1972
1973
1974
1975
1976
1977
1978

        # 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
1979
            if (not self.category) or (self.category == row["category"]):
1980
1981
1982
1983
                prompt = row["turns"][0]
                self.data.append({"prompt": prompt})

        random.seed(self.random_seed)
1984
1985
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1986
1987
1988
1989

    def sample(self, **kwargs) -> list:
        # leverage CustomDataset sample
        return super().sample(**kwargs)
1990
1991


1992
1993
1994
1995
# -----------------------------------------------------------------------------
# Sonnet Dataset Implementation
# -----------------------------------------------------------------------------

1996

1997
1998
1999
@deprecated(
    "SonnetDataset is deprecated and will be removed in a future version.",
)
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
class SonnetDataset(BenchmarkDataset):
    """
    Simplified implementation of the Sonnet dataset.  Loads poem lines from a
    text file and generates sample requests.  Default values here copied from
    `benchmark_serving.py` for the sonnet dataset.
    """

    DEFAULT_PREFIX_LEN = 200
    DEFAULT_INPUT_LEN = 550
    DEFAULT_OUTPUT_LEN = 150

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

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

    def sample(
        self,
        tokenizer,
        num_requests: int,
        prefix_len: int = DEFAULT_PREFIX_LEN,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
        return_prompt_formatted: bool = False,
2032
        request_id_prefix: str = "",
2033
        no_oversample: bool = False,
2034
2035
2036
2037
        **kwargs,
    ) -> list:
        # Calculate average token length for a poem line.
        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
2038
        avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines)
2039
2040
2041
2042

        # 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}]
2043
2044
2045
        base_fmt = tokenizer.apply_chat_template(
            base_msg, add_generation_prompt=True, tokenize=False
        )
2046
2047
2048
2049
        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 "
2050
2051
                f"({base_offset})."
            )
2052
2053
2054
2055
2056
2057
2058

        # 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 = []
2059
        ind = 0
2060
        while len(samples) < num_requests:
2061
2062
2063
            extra_lines = random.choices(
                self.data, k=num_input_lines - num_prefix_lines
            )
2064
2065
2066
            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
            msg = [{"role": "user", "content": prompt}]
            prompt_formatted = tokenizer.apply_chat_template(
2067
2068
                msg, add_generation_prompt=True, tokenize=False
            )
2069
2070
2071
2072
            prompt_len = len(tokenizer(prompt_formatted).input_ids)
            if prompt_len <= input_len:
                samples.append(
                    SampleRequest(
2073
                        prompt=prompt_formatted if return_prompt_formatted else prompt,
2074
2075
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
2076
2077
2078
                        request_id=request_id_prefix + str(ind),
                    )
                )
2079
                ind += 1
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
        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()

2099
2100
2101
    def load_data(
        self,
    ):
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
        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):
2115
            data = self.data.sample(n=num_requests, random_state=self.random_seed)
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
        else:
            data = self.data.sample(
                n=num_requests,
                random_state=self.random_seed,
                replace=True,
            )
        # Convert the dataframe to a list of lists.
        return data.values.tolist()

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2129
2130
        max_loras: int | None = None,
        lora_path: str | None = None,
2131
        request_id_prefix: str = "",
2132
        no_oversample: bool = False,
2133
2134
2135
2136
2137
2138
2139
        **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])
2140
            lora_req = self.get_random_lora_request(
2141
2142
                max_loras=max_loras, lora_path=lora_path
            )
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
            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,
2154
                    request_id=request_id_prefix + str(i),
2155
2156
                )
            )
2157
2158
2159
2160
2161
2162
2163
2164
2165
        return samples


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

2166
    SUPPORTED_DATASET_PATHS: set[str] | dict[str, Callable] = set()
2167
2168
2169
2170
2171

    def __init__(
        self,
        dataset_path: str,
        dataset_split: str,
2172
        no_stream: bool = False,
2173
2174
        dataset_subset: str | None = None,
        hf_name: str | None = None,
2175
2176
2177
2178
2179
2180
        **kwargs,
    ) -> None:
        super().__init__(dataset_path=dataset_path, **kwargs)

        self.dataset_split = dataset_split
        self.dataset_subset = dataset_subset
2181
        self.load_stream = not no_stream
2182
        self.hf_name = hf_name or dataset_path
2183
2184
2185
2186
2187
2188
2189
2190
        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,
2191
            streaming=self.load_stream,
2192
        )
2193
2194
        if not getattr(self, "disable_shuffle", False):
            self.data = self.data.shuffle(seed=self.random_seed)
2195
2196
2197
2198
2199
2200
2201
2202
2203


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


class ConversationDataset(HuggingFaceDataset):
    """Dataset for conversation data with multimodal support."""
2204

2205
    SUPPORTED_DATASET_PATHS = {
2206
2207
        "lmms-lab/LLaVA-OneVision-Data",
        "Aeala/ShareGPT_Vicuna_unfiltered",
2208
    }
2209
    IS_MULTIMODAL = True
2210

2211
2212
2213
2214
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2215
        output_len: int | None = None,
2216
2217
2218
2219
2220
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2221
        # Filter examples with at least 2 conversations
2222
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
2223
        sampled_requests = []
2224
        ind = 0
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
        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
2239
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
2240
                continue
2241
            mm_content = process_image(item["image"]) if "image" in item else None
2242
2243
2244
2245
            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
2246
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2247
2248
2249
2250
2251
2252
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2253
                    request_id=request_id_prefix + str(ind),
2254
2255
                )
            )
2256
            ind += 1
2257
2258
2259
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
        return sampled_requests


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


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

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2275
2276
        "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"],
2277
    }
2278
    IS_MULTIMODAL = True
2279
2280
2281
2282
2283

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2284
        output_len: int | None = None,
2285
        enable_multimodal_chat: bool = False,
2286
        request_id_prefix: str = "",
2287
        no_oversample: bool = False,
2288
2289
        **kwargs,
    ) -> list:
2290
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2291
        sampled_requests = []
2292
        for i, item in enumerate(self.data):
2293
2294
            if len(sampled_requests) >= num_requests:
                break
2295
            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
2296
            if parser_fn is None:
2297
                raise ValueError(f"Unsupported dataset path: {self.hf_name}")
2298
2299
2300
2301
2302
2303
2304
            prompt = parser_fn(item)
            mm_content = process_image(item["images"][0])
            prompt_len = len(tokenizer(prompt).input_ids)
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len
2305
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2306
2307
2308
2309
2310
2311
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2312
                    request_id=request_id_prefix + str(i),
2313
2314
2315
2316
2317
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2318
2319
2320
        return sampled_requests


2321
2322
2323
2324
2325
2326
2327
2328
class MMVUDataset(HuggingFaceDataset):
    """
    MMVU Dataset.
    https://huggingface.co/datasets/yale-nlp/MMVU
    """

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2329
2330
2331
        "yale-nlp/MMVU": lambda x: x["question"]
        + " "
        + (" ".join(f"{k}.{v}" for k, v in x["choices"].items())),
2332
2333
2334
2335
2336
2337
    }

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2338
        output_len: int | None = None,
2339
2340
2341
2342
2343
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2344
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
        sampled_requests = []
        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
            if parser_fn is None:
                raise ValueError(f"Unsupported dataset path: {self.hf_name}")
            prompt = parser_fn(item)
            mm_content = process_video(item["video"])
            prompt_len = len(tokenizer(prompt).input_ids)
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len
2359
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2360
2361
2362
2363
2364
2365
2366
            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),
2367
2368
2369
2370
2371
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2372
2373
2374
        return sampled_requests


2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
# -----------------------------------------------------------------------------
# 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",
    }

2395
2396
2397
2398
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2399
        output_len: int | None = None,
2400
2401
2402
2403
2404
2405
2406
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2407
        sampled_requests = []
2408
        for i, item in enumerate(self.data):
2409
2410
            if len(sampled_requests) >= num_requests:
                break
2411
2412
2413
2414
            prompt = (
                f"{item['input']}\n\n{item['instruction']} Just output "
                "the code, do not include any explanation."
            )
2415
2416

            # apply template
2417
2418
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2419
                    [{"role": "user", "content": prompt}],
2420
2421
2422
                    add_generation_prompt=True,
                    tokenize=False,
                )
2423

2424
2425
2426
2427
2428
2429
            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2430
                    request_id=request_id_prefix + str(i),
2431
2432
2433
2434
2435
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2436
2437
2438
        return sampled_requests


2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
# -----------------------------------------------------------------------------
# MT-Bench Dataset Implementation
# -----------------------------------------------------------------------------


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

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

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

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2463
        output_len: int | None = None,
2464
        enable_multimodal_chat: bool = False,
2465
        skip_chat_template: bool = False,
2466
        request_id_prefix: str = "",
2467
        no_oversample: bool = False,
2468
2469
        **kwargs,
    ) -> list:
2470
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2471
2472
        sampled_requests = []

2473
        for i, item in enumerate(self.data):
2474
2475
2476
2477
2478
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["turns"][0]

            # apply template
2479
2480
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2481
                    [{"role": "user", "content": prompt}],
2482
2483
2484
                    add_generation_prompt=True,
                    tokenize=False,
                )
2485
2486
2487
2488
2489
2490
2491

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2492
                    request_id=request_id_prefix + str(i),
2493
2494
2495
2496
2497
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2498
2499
2500
        return sampled_requests


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
# -----------------------------------------------------------------------------
# Blazedit Dataset Implementation
# -----------------------------------------------------------------------------


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

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

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

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2529
        output_len: int | None = None,
2530
        skip_chat_template: bool = False,
2531
        request_id_prefix: str = "",
2532
        no_oversample: bool = False,
2533
2534
2535
2536
        min_distance: float = 0.0,
        max_distance: float = 1.0,
        **kwargs,
    ) -> list:
2537
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
        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
2550
2551

            # template copied from
2552
            # https://github.com/ise-uiuc/blazedit/blob/7765137e656fd62de877422d2e4cf8de51228054/dataset/create_refined_dataset.py#L94-L105 # noqa: E501
2553
            prompt = f"""Given a code file, please apply the change requests and generate the new file.
2554
2555
2556
2557
2558
2559
2560
2561
2562

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

Change request:
{change_request}

2563
Please generate the new code file in the "New file" section below."""  # noqa: E501
2564
2565

            # apply template
2566
2567
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2568
                    [{"role": "user", "content": prompt}],
2569
2570
2571
                    add_generation_prompt=True,
                    tokenize=False,
                )
2572
2573
2574
2575
2576
2577
2578
2579
2580

            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),
2581
2582
2583
2584
2585
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2586

2587
2588
2589
        return sampled_requests


2590
2591
2592
2593
2594
2595
2596
2597
2598
# -----------------------------------------------------------------------------
# AIMO Dataset Implementation
# -----------------------------------------------------------------------------


class AIMODataset(HuggingFaceDataset):
    """
    Dataset class for processing a AIMO dataset with reasoning questions.
    """
2599

2600
    SUPPORTED_DATASET_PATHS = {
2601
2602
2603
        "AI-MO/aimo-validation-aime",
        "AI-MO/NuminaMath-1.5",
        "AI-MO/NuminaMath-CoT",
2604
2605
    }

2606
2607
2608
2609
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2610
        output_len: int | None = None,
2611
2612
2613
2614
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2615
        sampled_requests = []
2616
        ind = 0
2617
2618
2619
2620
2621
        dynamic_output = output_len is None

        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
2622
            prompt, completion = item["problem"], item["solution"]
2623
2624
2625
2626
2627
2628
2629

            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
2630
2631
2632
            if dynamic_output and not is_valid_sequence(
                prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000
            ):
2633
2634
2635
2636
2637
2638
2639
                continue
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=None,
2640
                    request_id=request_id_prefix + str(ind),
2641
2642
                )
            )
2643
            ind += 1
2644
2645
2646
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2647
        return sampled_requests
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667


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

2668
"""  # noqa: E501
2669
2670
2671


def _format_zeta_prompt(
2672
2673
    sample: dict, original_start_marker: str = "<|editable_region_start|>"
) -> dict:
2674
    """Format the zeta prompt for the Next Edit Prediction (NEP) dataset.
2675
2676
2677

    This function formats examples from the NEP dataset
    into prompts and expected outputs. It could be
2678
    further extended to support more NEP datasets.
2679

2680
    Args:
2681
        sample: The dataset sample containing events,
2682
            inputs, and outputs.
2683
2684
        original_start_marker: The marker indicating the
            start of the editable region. Defaults to
2685
            "<|editable_region_start|>".
2686

2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
    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,
    }

2716
2717
2718
2719
2720
2721
2722
2723
    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ):
2724
        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.hf_name)
2725
        if formatting_prompt_func is None:
2726
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")
2727
        samples = []
2728
        for i, sample in enumerate(self.data):
2729
2730
2731
2732
2733
2734
            sample = formatting_prompt_func(sample)
            samples.append(
                SampleRequest(
                    prompt=sample["prompt"],
                    prompt_len=len(tokenizer(sample["prompt"]).input_ids),
                    expected_output_len=len(
2735
2736
                        tokenizer(sample["expected_output"]).input_ids
                    ),
2737
                    request_id=request_id_prefix + str(i),
2738
2739
                )
            )
2740
2741
            if len(samples) >= num_requests:
                break
2742
2743
2744
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
2745
        return samples
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784


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


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

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

    """  # noqa: E501

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

    DEFAULT_OUTPUT_LEN = 128
    IS_MULTIMODAL = True

    # TODO Whisper-specific. Abstract interface when more models are supported.
2785
    TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>"
2786
2787
2788
2789
2790
2791
    skip_long_audios: bool = True

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2792
        output_len: int | None = None,
2793
        request_id_prefix: str = "",
2794
        no_oversample: bool = False,
2795
2796
        **kwargs,
    ) -> list:
2797
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2798
2799
2800
        prompt = ASRDataset.TRANSCRIPTION_PREAMBLE
        prompt_len = len(tokenizer(prompt).input_ids)
        sampled_requests = []
2801
        ind = 0
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
        skipped = 0
        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
            audio = item["audio"]
            y, sr = audio["array"], audio["sampling_rate"]
            duration_s = librosa.get_duration(y=y, sr=sr)
            # Whisper max supported duration
            if self.skip_long_audios and duration_s > 30:
                skipped += 1
                continue

            mm_content = {"audio": (y, sr)}
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2821
                    request_id=request_id_prefix + str(ind),
2822
2823
                )
            )
2824
            ind += 1
2825
2826
2827
2828
2829
2830
2831
        if skipped:
            logger.warning(
                "%d samples discarded from dataset due to"
                " their length being greater than"
                " what Whisper supports.",
                skipped,
            )
2832
2833
2834
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2835
        return sampled_requests
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869


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


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

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

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

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

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

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
2870
        output_len: int | None = None,
2871
        request_id_prefix: str = "",
2872
        no_oversample: bool = False,
2873
2874
2875
2876
2877
        **kwargs,
    ) -> list[SampleRequest]:
        # Force dynamic output length based on reference completion.
        dynamic_output = output_len is None
        sampled_requests: list[SampleRequest] = []
2878
        ind = 0
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912

        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,
2913
                    request_id=request_id_prefix + str(ind),
2914
2915
                )
            )
2916
            ind += 1
2917

2918
2919
2920
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2921
        return sampled_requests
2922
2923
2924
2925
2926
2927
2928
2929


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


class PrefixRepetitionRandomDataset(BenchmarkDataset):
2930
    # Default values copied from benchmark_serving.py for the repeated prefix
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
    # dataset.
    DEFAULT_PREFIX_LEN = 256
    DEFAULT_SUFFIX_LEN = 256
    DEFAULT_NUM_PREFIXES = 10
    DEFAULT_OUTPUT_LEN = 128

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

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
        prefix_len: int = DEFAULT_PREFIX_LEN,
        suffix_len: int = DEFAULT_SUFFIX_LEN,
        num_prefixes: int = DEFAULT_NUM_PREFIXES,
        output_len: int = DEFAULT_OUTPUT_LEN,
2953
        request_id_prefix: str = "",
2954
        no_oversample: bool = False,
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
        **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
2969
            tokens = np.random.randint(0, vocab_size, size=target_length).tolist()
2970

2971
            _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(  # noqa: E501
2972
2973
2974
2975
2976
2977
                tokenizer=tokenizer,
                token_sequence=tokens,
                target_token_len=target_length,
                add_special_tokens=False,
            )
            return adjusted_tokens, token_mismatch
2978
2979

        requests = []
2980
        token_mismatch_total = 0
2981
        for _ in range(num_prefixes):
2982
2983
            prefix_tokens, prefix_mismatch = _generate_exact_length_tokens(prefix_len)
            token_mismatch_total += prefix_mismatch
2984
2985

            for _ in range(prompts_per_prefix):
2986
                suffix_tokens, suffix_mismatch = _generate_exact_length_tokens(
2987
                    suffix_len
2988
                )
2989
                token_mismatch_total += suffix_mismatch
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
                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,
                    )
                )

3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
        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,
            )
3011
3012
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(requests)
3013
        return requests
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025


# -----------------------------------------------------------------------------
# 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
    """
3026

3027
3028
3029
3030
3031
3032
3033
3034
    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {"Lin-Chen/MMStar"}
    IS_MULTIMODAL = True

    def sample(
        self,
        tokenizer: PreTrainedTokenizerBase,
        num_requests: int,
3035
        output_len: int | None = None,
3036
3037
3038
3039
3040
3041
        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.
3042
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
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
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
        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