datasets.py 126 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
7
8
9
10
11
12
13
"""
This module defines a framework for sampling benchmark requests from various
datasets. Each dataset subclass of BenchmarkDataset must implement sample
generation. Supported dataset types include:
  - ShareGPT
  - Random (synthetic)
  - Sonnet
  - BurstGPT
  - HuggingFace
  - VisionArena
"""
14

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

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

from vllm.lora.request import LoRARequest
from vllm.lora.utils import get_adapter_absolute_path
from vllm.multimodal import MultiModalDataDict
40
from vllm.multimodal.image import convert_image_mode
41
from vllm.tokenizers import TokenizerLike
42
from vllm.utils.argparse_utils import FlexibleArgumentParser
43
from vllm.utils.import_utils import PlaceholderModule
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

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

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

try:
    import librosa
except ImportError:
    librosa = PlaceholderModule("librosa")
60
61
62
63
64
65
66
67
68
69
70
71
72
73

logger = logging.getLogger(__name__)

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


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

74
    prompt: str | list[str]
75
76
    prompt_len: int
    expected_output_len: int
77
78
79
    multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None
    lora_request: LoRARequest | None = None
    request_id: str | None = None
80
81
82
83
84
85
86
87
88


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


class BenchmarkDataset(ABC):
    DEFAULT_SEED = 0
89
    IS_MULTIMODAL = False
90
91
92

    def __init__(
        self,
93
        dataset_path: str | None = None,
94
        random_seed: int = DEFAULT_SEED,
95
96
        disable_shuffle: bool = False,
        **kwargs,
97
98
99
    ) -> None:
        """
        Initialize the BenchmarkDataset with an optional dataset path and random
100
101
        seed.

102
103
        Args:
            dataset_path (Optional[str]): Path to the dataset. If None, it
104
                indicates that a default or random dataset might be used.
105
            random_seed (int): Seed value for reproducible shuffling or
106
                sampling. Defaults to DEFAULT_SEED.
107
108
109
110
        """
        self.dataset_path = dataset_path
        # Set the random seed, ensuring that a None value is replaced with the
        # default seed.
111
        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
112
        self.disable_shuffle = disable_shuffle
113
114
115
        self.data = None

    def apply_multimodal_chat_transformation(
116
117
        self,
        prompt: str,
118
        mm_content: MultiModalDataDict | dict | list[dict] | None = None,
119
    ) -> list[dict]:
120
121
122
123
124
125
126
        """
        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:
127
128
129
130
131
            if isinstance(mm_content, list):
                content.extend(cast(list[dict[str, Any]], mm_content))
            elif isinstance(mm_content, dict):
                content.append(mm_content)
            else:
132
                raise TypeError(
133
                    f"Could not process multimodal content of type: {type(mm_content)}"
134
                )
135
136
137
138
139
140
141
142
143
144
145
146
147
        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
148
        raise NotImplementedError("load_data must be implemented in subclasses.")
149
150
151

    def get_random_lora_request(
        self,
152
153
154
        max_loras: int | None = None,
        lora_path: str | None = None,
    ) -> LoRARequest | None:
155
        """
156
        Optionally select a random LoRA request.
157
158

        This method is used when LoRA parameters are provided.  It randomly
159
        selects a LoRA based on max_loras.
160
161

        Args:
162
163
164
165
            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.
166
167

        Returns:
168
169
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
170
171
        """
        if max_loras is None or lora_path is None:
172
            return None
173
174
175
176
177
178
179
180

        # 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),
        )
181
        return lora_request
182
183

    @abstractmethod
184
185
    def sample(
        self,
186
        tokenizer: TokenizerLike,
187
188
189
190
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
    ) -> list[SampleRequest]:
191
192
193
194
195
196
197
        """
        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:
198
            tokenizer (TokenizerLike): The tokenizer to be used
199
                for processing the dataset's text.
200
            num_requests (int): The number of sample requests to generate.
201
            request_id_prefix (str): The prefix of request_id.
202
203
204
205
206
207
208

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

209
210
211
212
213
    def maybe_oversample_requests(
        self,
        requests: list[SampleRequest],
        num_requests: int,
        request_id_prefix: str = "",
214
        no_oversample: bool = False,
215
    ) -> None:
216
217
218
219
220
221
        """
        Oversamples the list of requests if its size is less than the desired
        number.

        Args:
            requests (List[SampleRequest]): The current list of sampled
222
223
                requests.
            num_requests (int): The target number of requests.
224
225
            request_id_prefix (str): The prefix applied to generated request
                identifiers.
226

227
        """
228
        if no_oversample:
229
            logger.info("Skipping oversampling. Total samples: %d.", len(requests))
230
231
            return

232
233
        if len(requests) < num_requests:
            random.seed(self.random_seed)
234
235
236
237
            needed = num_requests - len(requests)
            additional = []
            for i in range(needed):
                req = deepcopy(random.choice(requests))
238
                req.request_id = request_id_prefix + str(len(requests) + i)
239
                additional.append(req)
240
            requests.extend(additional)
241
            logger.info("Oversampled requests to reach %d total samples.", num_requests)
242

243
244
        ids = [req.request_id for req in requests]
        if len(ids) != len(set(ids)):
245
246
247
248
249
            raise ValueError(
                "Duplicate request_id found in the sampled "
                "requests. Please ensure that each request_id "
                "is unique."
            )
250

251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273

# -----------------------------------------------------------------------------
# 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
274
    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
275
276
277
278
    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
279
280
281
    return not (
        prompt_too_short or output_too_short or prompt_too_long or combined_too_long
    )
282
283
284
285
286
287
288
289


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


# Global cache for LoRA tokenizers.
290
lora_tokenizer_cache: dict[int, TokenizerLike] = {}
291
292
293
294
295
296


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

297
    Supports the following input types:
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312

    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.
    """
313
314
    if isinstance(image, dict) and "bytes" in image:
        image = Image.open(BytesIO(image["bytes"]))
315
    if isinstance(image, Image.Image):
316
        image = convert_image_mode(image, "RGB")
317
318
        with io.BytesIO() as image_data:
            image.save(image_data, format="JPEG")
319
            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
320
321
        return {
            "type": "image_url",
322
            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
323
324
325
        }

    if isinstance(image, str):
326
327
328
329
330
        image_url = (
            image
            if image.startswith(("http://", "https://", "file://"))
            else f"file://{image}"
        )
331
332
        return {"type": "image_url", "image_url": {"url": image_url}}

333
334
335
336
    raise ValueError(
        f"Invalid image input {image}. Must be a PIL.Image.Image"
        " or str or dictionary with raw image bytes."
    )
337
338


339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
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.
    """
355
356
    if isinstance(video, dict) and "bytes" in video:
        video_bytes = video["bytes"]
357
358
359
        video_base64 = base64.b64encode(video_bytes).decode("utf-8")
        return {
            "type": "video_url",
360
            "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
361
362
363
        }

    if isinstance(video, str):
364
365
366
367
368
        video_url = (
            video
            if video.startswith(("http://", "https://", "file://"))
            else f"file://{video}"
        )
369
370
371
372
373
374
        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
    )

375
376

def gen_prompt_decode_to_target_len(
377
    tokenizer: TokenizerLike,
378
379
380
381
    token_sequence: list[int],
    target_token_len: int,
    max_retry: int = 10,
    add_special_tokens: bool = False,
382
    rng: np.random.Generator | None = None,
383
) -> tuple[str, list[int], int]:
384
385
386
387
    """
    Ensure decoded-then-encoded prompt length matches the target token length.

    This function decodes an initial token sequence to text and re-encodes it
388
389
    , iteratively adjusting the token sequence length to match a target.
    This is necessary because some tokenizers do not guarantee a 1:1 mapping
390
391
392
393
394
    between consecutive tokens and the decoded-then-encoded sequence length.
    For example, for GPT2Tokenizer:
    [6880, 6881] -> ['Ġcalls', 'here'] ->
    [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']

395
396
397
    Returns a tuple of the final prompt string, the adjusted token sequence,
    and the token mismatch (final_len - target_token_len) if the retry budget
    is exhausted.
398
399
400
401
402
    """
    remain_num_try = max_retry
    token_mismatch = 0
    while True:
        prompt = tokenizer.decode(token_sequence)
403
        token_sequence = tokenizer.encode(prompt, add_special_tokens=add_special_tokens)
404
405
406
407
        if remain_num_try <= 0:
            if len(token_sequence) != target_token_len:
                token_mismatch = len(token_sequence) - target_token_len
            break
408

409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
        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

432

433
434
435
436
# -----------------------------------------------------------------------------
# Random Dataset Implementation (Synthetic Data)
# -----------------------------------------------------------------------------

437

438
class RandomDataset(BenchmarkDataset):
439
440
441
442
443
444
445
446
447
448
449
450
    """
    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.
    """
451

452
453
454
455
456
457
    # 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

458
    def __init__(self, **kwargs) -> None:
459
        super().__init__(**kwargs)
460
461
462
463
        # 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)
464
465
466

    def sample(
        self,
467
        tokenizer: TokenizerLike,
468
        num_requests: int,
469
        request_id_prefix: str = "",
470
        no_oversample: bool = False,
471
472
473
474
        prefix_len: int = DEFAULT_PREFIX_LEN,
        range_ratio: float = DEFAULT_RANGE_RATIO,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
475
        batchsize: int = 1,
476
477
        **kwargs,
    ) -> list[SampleRequest]:
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
        # validate total input tokens (prefix + sampled) is at least 1.
        num_special = int(tokenizer.num_special_tokens_to_add())
        real_input_len = max(0, int(input_len) - num_special)
        min_sampled_input = math.floor(real_input_len * (1.0 - float(range_ratio)))
        min_total_input = int(prefix_len) + min_sampled_input
        if min_total_input < 1:
            raise ValueError(
                "--random-input-len is too small: with tokenizer special "
                f"tokens {num_special} and --random-range-ratio {range_ratio}, "
                "the minimum possible total input tokens (prefix + sampled) is "
                f"{min_total_input}. Increase --random-input-len and/or "
                "--random-prefix-len, or decrease --random-range-ratio so that "
                "prefix_len + floor(max(0, random_input_len - num_special)) "
                "* (1 - range_ratio) >= 1."
            )

494
495
        input_lens, output_lens, offsets = self.get_sampling_params(
            num_requests, range_ratio, input_len, output_len, tokenizer
496
497
498
        )

        vocab_size = tokenizer.vocab_size
499
500
501
502
503
        prohibited_tokens = tokenizer.all_special_ids
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))

        # Generate prefix once
504
        prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len)
505

506
        requests = []
507
        token_mismatch_total = 0
508
        for i in range(num_requests):
509
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
510
511
512
513
514
515
516
                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,
517
                allowed_tokens=allowed_tokens,
518
            )
519
            token_mismatch_total += token_mismatch
520
521
522
523
524
525
526
527
            requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
                    request_id=request_id_prefix + str(i),
                )
            )
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
        # 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
543

544
545
546
547
548
549
550
551
552
553
554
        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,
            )

555
556
557
        return requests

    def get_prefix(
558
        self,
559
        tokenizer: TokenizerLike,
560
561
        allowed_tokens: np.ndarray,
        prefix_len: int,
562
563
564
565
    ) -> list[int]:
        """
        Get the prefix for the dataset.
        """
566
567
568
569
570
571
572
573
574
575
576
577
        if prefix_len <= 0:
            return []

        prefix_tokens = allowed_tokens[
            self._rng.integers(0, len(allowed_tokens), size=prefix_len)
        ].tolist()
        _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(
            tokenizer=tokenizer,
            token_sequence=prefix_tokens,
            target_token_len=prefix_len,
            add_special_tokens=False,
            rng=self._rng,
578
        )
579
580
581
582
583
584
585
586
587
588
        if token_mismatch != 0:
            sign = "more" if token_mismatch > 0 else "fewer"
            logger.warning(
                "Prefix tokenization produced %d %s tokens than expected "
                "after decoding and re-encoding. This is expected due to "
                "the imperfect nature of the sampling procedure",
                abs(token_mismatch),
                sign,
            )
        return adjusted_tokens
589

590
591
592
593
594
595
    def get_sampling_params(
        self,
        num_requests: int,
        range_ratio: float,
        input_len: int,
        output_len: int,
596
        tokenizer: TokenizerLike,
597
598
599
600
601
602
603
604
605
606
607
608
609
610
    ) -> 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))
611
612
        # Ensure the lower bound for output length is at least 1 to
        # prevent sampling 0 tokens.
613
        output_low = max(output_low, 1)
614
        output_high = max(output_high, 1)
615
616
617

        if input_low > input_high:
            raise ValueError(
618
                f"Invalid input sampling interval: low={input_low} > high={input_high}"
619
620
621
622
623
624
            )
        if output_low > output_high:
            raise ValueError(
                "Invalid output sampling interval: "
                f"low={output_low} > high={output_high}"
            )
625

626
627
        logger.info(
            "Sampling input_len from [%s, %s] and output_len from [%s, %s]",
628
629
630
631
632
            input_low,
            input_high,
            output_low,
            output_high,
        )
633

634
635
636
        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)
637
        return input_lens, output_lens, offsets
638

639
640
641
    def generate_token_sequence(
        self,
        *,
642
        tokenizer: TokenizerLike,
643
644
645
646
647
648
        prefix_token_ids: list[int],
        prefix_len: int,
        vocab_size: int,
        input_len: int,
        offset: int,
        index: int,
649
        allowed_tokens: np.ndarray,
650
    ) -> tuple[str, int, int]:
651
652
653
654
655
656
657
658
659
660
        """
        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,
661
        the encoded sequence is truncated before being decoded again.
662
        """
663
664
665
666
667
        # Build the inner sequence by sampling
        # sequentially from the allowed tokens
        inner_seq = allowed_tokens[
            (offset + index + np.arange(input_len)) % len(allowed_tokens)
        ].tolist()
668
669
670
671
        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)
672
        prompt, adjusted_token_sequence, token_mismatch = (
673
            gen_prompt_decode_to_target_len(
674
675
676
677
678
679
                tokenizer=tokenizer,
                token_sequence=token_sequence,
                target_token_len=total_input_len,
                add_special_tokens=False,
                rng=self._rng,
            )
680
681
682
        )
        total_input_len = len(adjusted_token_sequence)
        return prompt, total_input_len, token_mismatch
683
684


685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
# -----------------------------------------------------------------------------
# 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,
702
        tokenizer: TokenizerLike,
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
        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
        )
732

733
        vocab_size = tokenizer.vocab_size
734
735
736
        prohibited_tokens = tokenizer.all_special_ids
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))
737
738
739
740
741
742
743
744
745
746

        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,
747
                allowed_tokens=allowed_tokens,
748
749
750
751
752
753
754
755
756
757
758
759
760
            )
        )

        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,
761
                allowed_tokens=allowed_tokens,
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
            )
            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


797
798
799
800
# -----------------------------------------------------------------------------
# MultiModalDataset Implementation
# -----------------------------------------------------------------------------

801

802
803
804
805
806
807
class RandomMultiModalDataset(RandomDataset):
    """
    Synthetic multimodal dataset (text + images) that extends RandomDataset.

    Status:
    - Images: supported via synthetic RGB data.
808
    - Video: supported via synthetic RGB data.
809
810
811
812
813
814
815
816
    - 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
817
       mapping (height, width, num_frames) → probability. We treat
818
       `num_frames`=1 as image and `num_frames` > 1 as video.
819
       Entries with zero probability are removed and the rest are renormalized
820
821
822
823
824
825
826
       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}
827
828
      - Two image buckets (`num_frames`=1) and one video bucket
      (`num_frames`=16).
829
830
831
832
    OBS.: Only image sampling is supported for now.
    """

    IS_MULTIMODAL = True
833
    DEFAULT_LIMIT_MM_PER_PROMPT = {"image": 255, "video": 1}
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848

    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.
849
850
851

        NOTE: iid pixel sampling results in worst-case compression
        (good for stressing I/O), but very unlike real photos.
852
853
854
855
856
857
858
859
860
861
862
        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)

863
864
865
    def generate_synthetic_video(
        self, width: int, height: int, num_frames: int
    ) -> dict:
866
        """Generate synthetic video with random values.
867

868
869
        Creates a video with random pixel values, encodes it to MP4 format,
        and returns the content as bytes.
870
        """
871
872
        import cv2

873
874
875
876
877
878
879
880
881
882
883
        random_pixels = self._rng.integers(
            0,
            256,
            (num_frames, height, width, 3),
            dtype=np.uint8,
        )

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

884
        with NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
            temp_path = temp_file.name

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

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

            for frame in random_pixels:
                video_writer.write(frame)

            video_writer.release()
            temp_file.close()

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

            return {"bytes": video_content}
906
907
908
909
910
911
912
913
914
915

    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}")

916
917
918
    def normalize_bucket_config(
        self, bucket_config: dict[tuple[int, int, int], float]
    ) -> dict[tuple[int, int, int], float]:
919
920
921
922
923
924
925
926
927
928
929
        """
        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:
930
931
932
            raise ValueError(
                "Got invalid bucket config. Bucket config values must be non-zero."
            )
933
934
935
936
        # 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()}

937
938
939
940
    def generate_mm_item(
        self,
        mm_item_config: tuple[int, int, int],
    ) -> Mapping[str, Any]:
941
        """
942
        Create synthetic images and videos and
943
944
945
946
        apply process_image/process_video respectively.
        This follows the OpenAI API chat completions
        https://github.com/openai/openai-python
        """
947

948
        if self.map_config_to_modality(mm_item_config) == "image":
949
950
951
            return process_image(
                self.generate_synthetic_image(mm_item_config[1], mm_item_config[0])
            )
952
        elif self.map_config_to_modality(mm_item_config) == "video":
953
954
955
956
957
            return process_video(
                self.generate_synthetic_video(
                    mm_item_config[1], mm_item_config[0], mm_item_config[2]
                )
            )
958
        else:
959
            raise ValueError(f"Invalid multimodal item configuration: {mm_item_config}")
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979

    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:
980
981
982
983
984
                raise ValueError(
                    f"Modality {modality} is not in "
                    f"limit_mm_per_prompt: "
                    f"{limit_mm_per_prompt.keys()}"
                )
985

986
        # Remove zero probability entries
987
988
989
        # and normalize bucket config to sum to 1
        bucket_config = self.normalize_bucket_config(bucket_config)
        logger.info(
990
991
            "Normalized bucket config: %s",
            bucket_config,
992
993
        )
        # Only consider limit per prompt for modalities in bucket config
994
        allowed_modalities = {self.map_config_to_modality(cfg) for cfg in bucket_config}
995
        limit_mm_per_prompt = {
996
997
            k: v for k, v in limit_mm_per_prompt.items() if k in allowed_modalities
        }
998
        if not limit_mm_per_prompt:
999
            raise ValueError("No valid limits for modalities present in bucket_config.")
1000
1001

        logger.info(
1002
1003
            "Updated mm-limit-per-prompt: %s",
            limit_mm_per_prompt,
1004
1005
1006
1007
1008
        )

        # 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(
1009
            sum(limit_mm_per_prompt.values()),
1010
            math.ceil(base_items_per_request * (1 + num_mm_items_range_ratio)),
1011
1012
1013
        )
        # Ensure min num mm items is at least 0
        min_num_mm_items = max(
1014
            0, math.floor(base_items_per_request * (1 - num_mm_items_range_ratio))
1015
1016
1017
        )
        # Raise error if min num mm items is greater than max num mm items
        if min_num_mm_items > max_num_mm_items:
1018
1019
1020
1021
            raise ValueError(
                f"Min num mm items is greater than max mm items: "
                f"{min_num_mm_items} > {max_num_mm_items}"
            )
1022

1023
1024
        logger.info(
            "Sampling number of multimodal items from [%s, %s]",
1025
1026
            min_num_mm_items,
            max_num_mm_items,
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
        )

        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],
1042
    ) -> Iterator[tuple[int, int, int]]:
1043
1044
1045
1046
1047
        """
        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.
1048
1049
        Loop until the number of multimodal items sampled is equal to
        request_num_mm_items or limit of multimodal items per prompt
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
        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)
1061
        )
1062
1063
1064
1065
        # If request_num_mm_items is 0, yield an empty iterator
        if request_num_mm_items == 0:
            return
        # Initialize modality counters
1066
        modality_counter = {self.map_config_to_modality(k): 0 for k in bucket_config}
1067
1068
1069
1070
1071
        # 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
1072
1073
1074
            mm_item_config = self._rng.choice(
                list(bucket_config_copy.keys()), p=list(bucket_config_copy.values())
            )
1075
1076
1077
1078
            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
1079
                yield (mm_item_config)
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
            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()):
1090
1091
1092
                    logger.warning(
                        "Exhausted all multimodal items of modality %s", modality
                    )
1093
1094
                    break
                # Renormalize the bucket config
1095
                bucket_config_copy = self.normalize_bucket_config(bucket_config_copy)
1096
1097
1098

    def sample(
        self,
1099
        tokenizer: TokenizerLike,
1100
1101
        num_requests: int,
        request_id_prefix: str = "",
1102
        no_oversample: bool = False,
1103
1104
1105
1106
1107
1108
1109
        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,
1110
1111
1112
        bucket_config: dict[
            tuple[int, int, int], float
        ] = DEFAULT_MM_ITEM_BUCKET_CONFIG,
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
        enable_multimodal_chat: bool = DEFAULT_ENABLE_MULTIMODAL_CHAT,
        **kwargs,
    ) -> list[SampleRequest]:
        # Get the sampling parameters for the dataset
        input_lens, output_lens, offsets = self.get_sampling_params(
            num_requests, range_ratio, input_len, output_len, tokenizer
        )

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

        vocab_size = tokenizer.vocab_size
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
        # Can't use tokenizer.all_special_ids since
        # it returns ONLY ids from special_tokens_map.json
        # We want to exclude placeholder tokens and all
        # tokens that indicate start/end of image as it
        # may break prompt replacement logic.
        prohibited_tokens = list(
            tok_id
            for tok_id, token in tokenizer.added_tokens_decoder.items()
            if token.special
        )
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))
        logger.debug(
            "Sampling from %d out of %d (vocab size)", len(allowed_tokens), vocab_size
        )
        # Generate prefix once
1150
        prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len)
1151
1152
        # Add synthetic multimodal items to each request
        mm_requests = []
1153
        token_mismatch_total = 0
1154
        for i in range(num_requests):
1155
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
1156
1157
1158
1159
1160
1161
1162
                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,
1163
                allowed_tokens=allowed_tokens,
1164
            )
1165
            token_mismatch_total += token_mismatch
1166
1167
1168
1169
1170
1171
1172
1173
            # 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,
            )

1174
1175
1176
1177
1178
1179
1180
            mm_content = cast(
                list[dict[str, Any]],
                [
                    self.generate_mm_item(mm_item_config)
                    for mm_item_config in mm_item_iterator
                ],
            )
1181
1182

            if enable_multimodal_chat:
1183
                # NOTE: For now this option is only provided for completeness
1184
1185
1186
                # given that the serve.py benchmark currently does not use it.
                mm_chat_prompt: Any = prompt
                mm_chat_prompt = self.apply_multimodal_chat_transformation(
1187
1188
                    prompt, mm_content
                )
1189
1190
1191
1192
1193
1194
1195
1196
1197
                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(
1198
1199
1200
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
1201
                    multi_modal_data=mm_content,
1202
                    request_id=request_id_prefix + str(i),
1203
1204
                )
            mm_requests.append(sample_request)
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216

        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,
            )

1217
        return mm_requests
1218

1219

1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
# -----------------------------------------------------------------------------
# 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 = [
1243
1244
            entry
            for entry in self.data
1245
1246
1247
            if "conversations" in entry and len(entry["conversations"]) >= 2
        ]
        random.seed(self.random_seed)
1248
1249
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1250
1251
1252

    def sample(
        self,
1253
        tokenizer: TokenizerLike,
1254
        num_requests: int,
1255
1256
1257
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1258
        enable_multimodal_chat: bool = False,
1259
        request_id_prefix: str = "",
1260
        no_oversample: bool = False,
1261
1262
1263
        **kwargs,
    ) -> list:
        samples: list = []
1264
        ind = 0
1265
1266
1267
1268
1269
1270
1271
1272
        for entry in self.data:
            if len(samples) >= num_requests:
                break
            prompt, completion = (
                entry["conversations"][0]["value"],
                entry["conversations"][1]["value"],
            )

1273
            lora_request = self.get_random_lora_request(
1274
1275
                max_loras=max_loras, lora_path=lora_path
            )
1276
1277
1278
            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
1279
1280
1281
1282
1283
1284
            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,
            ):
1285
                continue
1286
1287
1288
            if image_path := entry.get("image"):
                mm_content = process_image(image_path)
            elif video_path := entry.get("video"):
1289
                mm_content = process_video(video_path)
1290
            else:
1291
                mm_content = None
1292
            if enable_multimodal_chat:
1293
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
1294
1295
1296
1297
1298
1299
            samples.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=new_output_len,
                    lora_request=lora_request,
1300
                    multi_modal_data=mm_content,
1301
                    request_id=request_id_prefix + str(ind),
1302
1303
                )
            )
1304
            ind += 1
1305
1306
1307
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
1308
1309
1310
        return samples


1311
1312
class _ValidateDatasetArgs(argparse.Action):
    """Argparse action to validate dataset name and path compatibility."""
1313

1314
1315
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, values)
1316

1317
        # Get current values of both dataset_name and dataset_path
1318
1319
        dataset_name = getattr(namespace, "dataset_name", "random")
        dataset_path = getattr(namespace, "dataset_path", None)
1320

1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
        # 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}"
            )


1331
def add_dataset_parser(parser: FlexibleArgumentParser):
1332
1333
1334
1335
1336
    parser.add_argument(
        "--trust-remote-code",
        action="store_true",
        help="Trust remote code from huggingface",
    )
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
    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",
1348
        action=_ValidateDatasetArgs,
1349
        choices=[
1350
1351
1352
1353
1354
            "sharegpt",
            "burstgpt",
            "sonnet",
            "random",
            "random-mm",
1355
            "random-rerank",
1356
1357
            "hf",
            "custom",
1358
            "custom_mm",
1359
1360
            "prefix_repetition",
            "spec_bench",
1361
        ],
1362
1363
        help="Name of the dataset to benchmark on.",
    )
1364
1365
1366
1367
1368
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
1369
1370
1371
1372
    parser.add_argument(
        "--dataset-path",
        type=str,
        default=None,
1373
        action=_ValidateDatasetArgs,
1374
1375
1376
        help="Path to the sharegpt/sonnet dataset. "
        "Or the huggingface dataset ID if using HF dataset.",
    )
1377
1378
1379
    parser.add_argument(
        "--no-oversample",
        action="store_true",
1380
        help="Do not oversample if the dataset has fewer samples than num-prompts.",
1381
    )
1382
1383
1384
    parser.add_argument(
        "--skip-chat-template",
        action="store_true",
1385
        help="Skip applying chat template to prompt for datasets that support it.",
1386
    )
1387
1388
1389
1390
1391
    parser.add_argument(
        "--enable-multimodal-chat",
        action="store_true",
        help="Enable multimodal chat transformation for datasets that support it.",
    )
1392
1393
1394
1395
1396
    parser.add_argument(
        "--disable-shuffle",
        action="store_true",
        help="Disable shuffling of dataset samples for deterministic ordering.",
    )
1397
1398
1399
1400
1401
1402
1403

    # 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,
1404
1405
1406
        help="Number of output tokens per request. Unless it is set to -1, the "
        "value overrides potential output length loaded from the dataset. It is "
        "used only for custom dataset.",
1407
1408
    )

1409
1410
1411
1412
1413
    spec_bench_group = parser.add_argument_group("spec bench dataset options")
    spec_bench_group.add_argument(
        "--spec-bench-output-len",
        type=int,
        default=256,
1414
        help="Num of output tokens per request, used only for spec bench dataset.",
1415
1416
1417
1418
1419
    )
    spec_bench_group.add_argument(
        "--spec-bench-category",
        type=str,
        default=None,
1420
        help="Category for spec bench dataset. If None, use all categories.",
1421
1422
    )

1423
1424
1425
1426
1427
    sonnet_group = parser.add_argument_group("sonnet dataset options")
    sonnet_group.add_argument(
        "--sonnet-input-len",
        type=int,
        default=550,
1428
        help="Number of input tokens per request, used only for sonnet dataset.",
1429
1430
1431
1432
1433
    )
    sonnet_group.add_argument(
        "--sonnet-output-len",
        type=int,
        default=150,
1434
        help="Number of output tokens per request, used only for sonnet dataset.",
1435
1436
1437
1438
1439
    )
    sonnet_group.add_argument(
        "--sonnet-prefix-len",
        type=int,
        default=200,
1440
        help="Number of prefix tokens per request, used only for sonnet dataset.",
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
    )

    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.",
    )

1452
1453
1454
1455
1456
    blazedit_group = parser.add_argument_group("blazedit dataset options")
    blazedit_group.add_argument(
        "--blazedit-min-distance",
        type=float,
        default=0.0,
1457
        help="Minimum distance for blazedit dataset. Min: 0, Max: 1.0",
1458
1459
1460
1461
1462
    )
    blazedit_group.add_argument(
        "--blazedit-max-distance",
        type=float,
        default=1.0,
1463
        help="Maximum distance for blazedit dataset. Min: 0, Max: 1.0",
1464
1465
    )

1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
    asr_group = parser.add_argument_group("asr dataset options")
    asr_group.add_argument(
        "--asr-max-audio-len-sec",
        type=float,
        default=float("inf"),
        help="Maximum audio length in seconds for ASR dataset.",
    )
    asr_group.add_argument(
        "--asr-min-audio-len-sec",
        type=float,
        default=0.0,
        help="Minimum audio length in seconds for ASR dataset.",
    )

1480
    random_group = parser.add_argument_group("random dataset options")
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
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
    add_random_dataset_base_args(random_group)

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

    hf_group = parser.add_argument_group("hf dataset options")
    hf_group.add_argument(
        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
    )
    hf_group.add_argument(
        "--hf-split", type=str, default=None, help="Split of the HF dataset."
    )
    hf_group.add_argument(
        "--hf-name",
        type=str,
        default=None,
        help=(
            "Name of the dataset on HuggingFace "
            "(e.g., 'lmarena-ai/VisionArena-Chat'). "
            "Specify this if your dataset-path is a local path."
        ),
    )
    hf_group.add_argument(
        "--hf-output-len",
        type=int,
        default=None,
        help="Output length for each request. Overrides the output lengths "
        "from the sampled HF dataset.",
    )

    prefix_repetition_group = parser.add_argument_group(
        "prefix repetition dataset options"
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-prefix-len",
        type=int,
        default=256,
        help="Number of prefix tokens per request, used only for prefix "
        "repetition dataset.",
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-suffix-len",
        type=int,
        default=256,
        help="Number of suffix tokens per request, used only for prefix "
        "repetition dataset. Total input length is prefix_len + suffix_len.",
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-num-prefixes",
        type=int,
        default=10,
        help="Number of prefixes to generate, used only for prefix repetition "
        "dataset. Prompts per prefix is num_requests // num_prefixes.",
    )
    prefix_repetition_group.add_argument(
        "--prefix-repetition-output-len",
        type=int,
        default=128,
        help="Number of output tokens per request, used only for prefix "
        "repetition dataset.",
    )


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

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

    Args:
        parser_or_group: Either a parser or an argument group to add arguments to.
    """
    parser_or_group.add_argument(
1560
1561
1562
        "--random-input-len",
        type=int,
        default=1024,
1563
        help="Number of input tokens per request, used only for random sampling.",
1564
    )
1565
    parser_or_group.add_argument(
1566
1567
1568
        "--random-output-len",
        type=int,
        default=128,
1569
        help="Number of output tokens per request, used only for random sampling.",
1570
    )
1571
    parser_or_group.add_argument(
1572
1573
1574
1575
1576
1577
1578
1579
        "--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)].",
    )
1580
    parser_or_group.add_argument(
1581
1582
1583
        "--random-prefix-len",
        type=int,
        default=0,
1584
1585
1586
1587
1588
1589
1590
1591
        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)]."
        ),
1592
    )
1593
    parser_or_group.add_argument(
1594
1595
1596
        "--random-batch-size",
        type=int,
        default=1,
1597
        help=("Batch size for random sampling. Only used for embeddings benchmark."),
1598
    )
1599
    parser_or_group.add_argument(
1600
1601
1602
1603
1604
1605
1606
        "--no-reranker",
        action="store_true",
        help=(
            "Whether the model supports reranking natively."
            " Only used for reranker benchmark."
        ),
    )
1607

1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620

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

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

    Args:
        parser_or_group: Either a parser or an argument group to add arguments to.
    """
    parser_or_group.add_argument(
1621
1622
1623
1624
1625
1626
1627
1628
1629
        "--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."
        ),
    )
1630
    parser_or_group.add_argument(
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
        "--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."
        ),
    )
1645
    parser_or_group.add_argument(
1646
1647
1648
1649
1650
        "--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. "
1651
            '\'{"image": 3, "video": 0}\'. The sampled per-request item '
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
            "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)
1668
1669
1670
1671
1672
                if not (
                    isinstance(key, tuple)
                    and len(key) == 3
                    and all(isinstance(x, int) for x in key)
                ):
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
                    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.")

1689
    parser_or_group.add_argument(
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
        "--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."
        ),
1708
1709
    )

1710

1711
def get_samples(args, tokenizer: TokenizerLike) -> list[SampleRequest]:
1712
1713
1714
    if not hasattr(args, "request_id_prefix"):
        args.request_id_prefix = ""

1715
    if args.dataset_name == "custom":
1716
1717
1718
        dataset = CustomDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1719
1720
1721
1722
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
1723
            skip_chat_template=args.skip_chat_template,
1724
            request_id_prefix=args.request_id_prefix,
1725
            no_oversample=args.no_oversample,
1726
1727
        )

1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
    elif args.dataset_name == "custom_mm":
        dataset = CustomMMDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
            enable_multimodal_chat=args.enable_multimodal_chat,
            request_id_prefix=args.request_id_prefix,
            no_oversample=args.no_oversample,
        )

1741
    elif args.dataset_name == "sonnet":
1742
1743
1744
        dataset = SonnetDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1745
        # For the "sonnet" dataset, formatting depends on the backend.
1746
        if args.backend == "openai-chat":
1747
1748
1749
1750
1751
1752
1753
            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,
1754
                request_id_prefix=args.request_id_prefix,
1755
                no_oversample=args.no_oversample,
1756
1757
1758
            )
        else:
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
1759
1760
                "Tokenizer/model must have chat template for sonnet dataset."
            )
1761
1762
1763
1764
1765
1766
1767
            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,
1768
                request_id_prefix=args.request_id_prefix,
1769
                no_oversample=args.no_oversample,
1770
1771
1772
1773
1774
            )

    elif args.dataset_name == "hf":
        # all following datasets are implemented from the
        # HuggingFaceDataset base class
1775
        hf_kwargs = {}
1776
1777
1778
1779
        if (
            args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS
        ):
1780
            dataset_class = VisionArenaDataset
1781
            args.hf_split = args.hf_split if args.hf_split else "train"
1782
            args.hf_subset = None
1783
1784
1785
1786
1787
        elif (
            args.dataset_path in MMVUDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMVUDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMVUDataset
1788
            args.hf_split = args.hf_split if args.hf_split else "validation"
1789
            args.hf_subset = None
1790
1791
1792
1793
        elif (
            args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS
        ):
1794
            dataset_class = InstructCoderDataset
1795
            args.hf_split = args.hf_split if args.hf_split else "train"
1796
1797
1798
1799
        elif (
            args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MTBenchDataset.SUPPORTED_DATASET_PATHS
        ):
1800
            dataset_class = MTBenchDataset
1801
            args.hf_split = args.hf_split if args.hf_split else "train"
1802
1803
1804
1805
1806
        elif (
            args.dataset_path in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MultiModalConversationDataset
1807
1808
1809
1810
        elif (
            args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS
        ):
1811
            dataset_class = ConversationDataset
1812
1813
1814
1815
        elif (
            args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS
        ):
1816
            dataset_class = AIMODataset
1817
            args.hf_split = args.hf_split if args.hf_split else "train"
1818
        elif (
1819
            args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS  # noqa: E501
1820
1821
            or args.hf_name in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS
        ):
1822
            dataset_class = NextEditPredictionDataset
1823
            args.hf_split = args.hf_split if args.hf_split else "train"
1824
1825
1826
1827
        elif (
            args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS
        ):
1828
            dataset_class = ASRDataset
1829
1830
1831
1832
1833
            args.hf_split = args.hf_split if args.hf_split else "train"
            hf_kwargs = {
                "asr_min_audio_len_sec": args.asr_min_audio_len_sec,
                "asr_max_audio_len_sec": args.asr_max_audio_len_sec,
            }
1834
1835
        elif args.dataset_path in BlazeditDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = BlazeditDataset
1836
            args.hf_split = args.hf_split if args.hf_split else "train"
1837
1838
1839
1840
            hf_kwargs = {
                "min_distance": args.blazedit_min_distance,
                "max_distance": args.blazedit_max_distance,
            }
1841
1842
1843
1844
        elif (
            args.dataset_path in MLPerfDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MLPerfDataset.SUPPORTED_DATASET_PATHS
        ):
1845
            dataset_class = MLPerfDataset
1846
            args.hf_split = args.hf_split if args.hf_split else "train"
1847
1848
1849
1850
1851
        elif (
            args.dataset_path in MMStarDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMStarDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMStarDataset
1852
            args.hf_split = args.hf_split if args.hf_split else "val"
1853
            args.hf_subset = None
1854
        else:
1855
1856
1857
1858
1859
1860
1861
            supported_datasets = set(
                [
                    dataset_name
                    for cls in HuggingFaceDataset.__subclasses__()
                    for dataset_name in cls.SUPPORTED_DATASET_PATHS
                ]
            )
1862
1863
1864
1865
1866
            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 "
1867
1868
                "like to add support for additional dataset formats."
            )
1869

1870
1871
        if dataset_class.IS_MULTIMODAL and not (
            args.backend in ("openai-chat", "openai-audio")
1872
            or "embeddings-" in args.backend
1873
        ):
1874
1875
            # multi-modal benchmark is only available on OpenAI Chat
            # endpoint-type.
1876
1877
            raise ValueError(
                "Multi-modal content is only supported on 'openai-chat' and "
1878
1879
                "'openai-audio' backends."
            )
1880
1881
1882
1883
1884
        input_requests = dataset_class(
            dataset_path=args.dataset_path,
            dataset_subset=args.hf_subset,
            dataset_split=args.hf_split,
            random_seed=args.seed,
1885
            no_stream=args.no_stream,
1886
            hf_name=args.hf_name,
1887
            disable_shuffle=args.disable_shuffle,
1888
            trust_remote_code=args.trust_remote_code,
1889
1890
1891
1892
        ).sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.hf_output_len,
1893
            enable_multimodal_chat=args.enable_multimodal_chat,
1894
            request_id_prefix=args.request_id_prefix,
1895
            no_oversample=args.no_oversample,
1896
            skip_chat_template=args.skip_chat_template,
1897
            **hf_kwargs,
1898
1899
1900
1901
1902
        )

    else:
        # For datasets that follow a similar structure, use a mapping.
        dataset_mapping = {
1903
            "spec_bench": lambda: SpecBench(
1904
1905
1906
                dataset_path=args.dataset_path,
                category=args.spec_bench_category,
                disable_shuffle=args.disable_shuffle,
1907
            ).sample(
1908
1909
1910
                num_requests=args.num_prompts,
                tokenizer=tokenizer,
                output_len=args.spec_bench_output_len,
1911
                enable_multimodal_chat=args.enable_multimodal_chat,
1912
                request_id_prefix=args.request_id_prefix,
1913
                no_oversample=args.no_oversample,
1914
            ),
1915
            "sharegpt": lambda: ShareGPTDataset(
1916
1917
1918
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1919
1920
1921
1922
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
1923
                enable_multimodal_chat=args.enable_multimodal_chat,
1924
                request_id_prefix=args.request_id_prefix,
1925
                no_oversample=args.no_oversample,
1926
1927
            ),
            "burstgpt": lambda: BurstGPTDataset(
1928
1929
1930
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1931
1932
1933
1934
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                request_id_prefix=args.request_id_prefix,
1935
                no_oversample=args.no_oversample,
1936
1937
            ),
            "random": lambda: RandomDataset(
1938
1939
1940
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1941
            ).sample(
1942
1943
1944
1945
1946
1947
                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,
1948
                request_id_prefix=args.request_id_prefix,
1949
                batchsize=args.random_batch_size,
1950
                no_oversample=args.no_oversample,
1951
            ),
1952
            "random-mm": lambda: RandomMultiModalDataset(
1953
1954
1955
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
            ).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,
1967
                enable_multimodal_chat=args.enable_multimodal_chat,
1968
                request_id_prefix=args.request_id_prefix,
1969
                no_oversample=args.no_oversample,
1970
            ),
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
            "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,
            ),
1984
            "prefix_repetition": lambda: PrefixRepetitionRandomDataset(
1985
1986
1987
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
1988
1989
1990
1991
1992
1993
1994
            ).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,
1995
                request_id_prefix=args.request_id_prefix,
1996
                no_oversample=args.no_oversample,
1997
            ),
1998
1999
2000
        }

        try:
2001
            # Enforce endpoint compatibility for multimodal datasets.
2002
            if args.dataset_name == "random-mm" and args.backend not in ["openai-chat"]:
2003
2004
2005
2006
                raise ValueError(
                    "Multi-modal content (images) is only supported on "
                    "'openai-chat' backend."
                )
2007
2008
2009
2010
2011
2012
2013
            input_requests = dataset_mapping[args.dataset_name]()
        except KeyError as err:
            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err

    return input_requests


2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
# -----------------------------------------------------------------------------
# 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.,
    ```
2024
2025
2026
    {"prompt": "What is the capital of India?", "output_tokens": 10}
    {"prompt": "What is the capital of Iran?", "output_tokens": 1520}
    {"prompt": "What is the capital of China?", "output_tokens": 819}
2027
    ```
2028
2029
    Note that 'output_tokens' column is optional and has to be provided only if
    'custom-output-len' argument is None or -1.
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
    """

    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"):
2049
            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062

            # 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(
2063
2064
                "Only JSONL format is supported for CustomDataset."
            )
2065
2066

        random.seed(self.random_seed)
2067
2068
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2069
2070
2071

    def sample(
        self,
2072
        tokenizer: TokenizerLike,
2073
        num_requests: int,
2074
2075
2076
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
2077
2078
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
2079
        request_id_prefix: str = "",
2080
        no_oversample: bool = False,
2081
2082
        **kwargs,
    ) -> list:
2083
2084
2085
2086
        # load all data if needed
        self.num_available_samples = len(self.data)
        if num_requests <= 0:
            num_requests = self.num_available_samples
2087
2088
2089
2090
2091
            logger.info(
                "num_requests is set to 0 or negative, "
                "so using all available samples: %d",
                num_requests,
            )
2092

2093
        sampled_requests = []
2094
        for i, item in enumerate(self.data):
2095
2096
2097
2098
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["prompt"]

2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
            if tokenizer is None:
                new_output_len = 1
            else:
                new_output_len = output_len
                if output_len is None or output_len == -1:
                    # check that the request has an 'output_tokens' field
                    if "output_tokens" not in item:
                        raise ValueError(
                            "If no output length is provided the "
                            "custom dataset must contain an 'output_tokens' field."
                        )
                    # Use number of output tokens from the request data
                    try:
                        new_output_len = int(item["output_tokens"])
                    except (ValueError, TypeError) as e:
                        raise ValueError(
                            f"Invalid value for 'output_tokens' in custom dataset: "
                            f"'{item['output_tokens']}'. Must be an integer."
                        ) from e

            if tokenizer is None:
                prompt_len = 1
            else:
                # apply template
                if not skip_chat_template:
                    prompt = tokenizer.apply_chat_template(
                        [{"role": "user", "content": prompt}],
                        add_generation_prompt=True,
                        tokenize=False,
2128
                    )
2129

2130
                prompt_len = len(tokenizer(prompt).input_ids)
2131
2132
2133
2134
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
2135
                    expected_output_len=new_output_len,
2136
                    request_id=request_id_prefix + str(i),
2137
2138
2139
2140
2141
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2142
2143
2144
2145

        return sampled_requests


2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
class CustomMMDataset(CustomDataset):
    """
    Implements the Custom MultiModal dataset. Loads data from a JSONL file and generates
    sample requests based on conversation turns. E.g.,
    ```
    {
        "prompt": "How many red blocks in the given images?",
        "image_files": ["path/to/image1.png", "path/to/image2.png"],
    }
    {
        "prompt": "Which country has the most pokemons based on the given graphs?",
        "image_files": ["path/to/image.png"],
    }
    ```

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

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

    IS_MULTIMODAL = True

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

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

            prompt_len = len(tokenizer(prompt).input_ids)
            images = item["image_files"]
            if len(images) > 1:
                logger.warning(
                    "Multiple image files found for sample %d. "
                    "Only the first image will be used.",
                    i,
                )
            mm_content = process_image(images[0])
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)

            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
                    request_id=request_id_prefix + str(i),
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )

        return sampled_requests


2225
2226
2227
2228
2229
2230
2231
2232
# -----------------------------------------------------------------------------
# Spec Bench Dataset Implementation
# -----------------------------------------------------------------------------


class SpecBench(CustomDataset):
    """
    Implements the SpecBench dataset: https://github.com/hemingkx/Spec-Bench
2233
    Download the dataset using:
2234
    wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
2235
    """  # noqa: E501
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248

    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
2249
        jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2250
2251
2252
2253
2254
2255
2256

        # 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
2257
            if (not self.category) or (self.category == row["category"]):
2258
2259
2260
2261
                prompt = row["turns"][0]
                self.data.append({"prompt": prompt})

        random.seed(self.random_seed)
2262
2263
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2264
2265
2266
2267

    def sample(self, **kwargs) -> list:
        # leverage CustomDataset sample
        return super().sample(**kwargs)
2268
2269


2270
2271
2272
2273
# -----------------------------------------------------------------------------
# Sonnet Dataset Implementation
# -----------------------------------------------------------------------------

2274

2275
2276
2277
@deprecated(
    "SonnetDataset is deprecated and will be removed in a future version.",
)
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
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,
2304
        tokenizer: TokenizerLike,
2305
2306
2307
2308
2309
        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,
2310
        request_id_prefix: str = "",
2311
        no_oversample: bool = False,
2312
2313
2314
2315
        **kwargs,
    ) -> list:
        # Calculate average token length for a poem line.
        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
2316
        avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines)
2317
2318
2319
2320

        # 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}]
2321
2322
2323
        base_fmt = tokenizer.apply_chat_template(
            base_msg, add_generation_prompt=True, tokenize=False
        )
2324
2325
2326
2327
        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 "
2328
2329
                f"({base_offset})."
            )
2330
2331
2332
2333
2334
2335
2336

        # 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 = []
2337
        ind = 0
2338
        while len(samples) < num_requests:
2339
2340
2341
            extra_lines = random.choices(
                self.data, k=num_input_lines - num_prefix_lines
            )
2342
2343
2344
            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
            msg = [{"role": "user", "content": prompt}]
            prompt_formatted = tokenizer.apply_chat_template(
2345
2346
                msg, add_generation_prompt=True, tokenize=False
            )
2347
2348
2349
2350
            prompt_len = len(tokenizer(prompt_formatted).input_ids)
            if prompt_len <= input_len:
                samples.append(
                    SampleRequest(
2351
                        prompt=prompt_formatted if return_prompt_formatted else prompt,
2352
2353
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
2354
2355
2356
                        request_id=request_id_prefix + str(ind),
                    )
                )
2357
                ind += 1
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
        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()

2377
2378
2379
    def load_data(
        self,
    ):
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
        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):
2393
            data = self.data.sample(n=num_requests, random_state=self.random_seed)
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
        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,
2405
        tokenizer: TokenizerLike,
2406
        num_requests: int,
2407
2408
        max_loras: int | None = None,
        lora_path: str | None = None,
2409
        request_id_prefix: str = "",
2410
        no_oversample: bool = False,
2411
2412
2413
2414
2415
2416
2417
        **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])
2418
            lora_req = self.get_random_lora_request(
2419
2420
                max_loras=max_loras, lora_path=lora_path
            )
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
            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,
2432
                    request_id=request_id_prefix + str(i),
2433
2434
                )
            )
2435
2436
2437
2438
2439
2440
2441
2442
2443
        return samples


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

2444
    SUPPORTED_DATASET_PATHS: set[str] | dict[str, Callable] = set()
2445
2446
2447
2448
2449

    def __init__(
        self,
        dataset_path: str,
        dataset_split: str,
2450
        no_stream: bool = False,
2451
2452
        dataset_subset: str | None = None,
        hf_name: str | None = None,
2453
        trust_remote_code: bool = False,
2454
2455
2456
2457
2458
2459
        **kwargs,
    ) -> None:
        super().__init__(dataset_path=dataset_path, **kwargs)

        self.dataset_split = dataset_split
        self.dataset_subset = dataset_subset
2460
        self.load_stream = not no_stream
2461
        self.hf_name = hf_name or dataset_path
2462
        self.trust_remote_code = trust_remote_code
2463
2464
2465
2466
2467
2468
2469
2470
        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,
2471
            streaming=self.load_stream,
2472
            trust_remote_code=self.trust_remote_code,
2473
        )
2474
2475
        if not getattr(self, "disable_shuffle", False):
            self.data = self.data.shuffle(seed=self.random_seed)
2476
2477
2478
2479
2480
2481
2482
2483


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


class ConversationDataset(HuggingFaceDataset):
2484
    """Dataset for text-only conversation data."""
2485

2486
    SUPPORTED_DATASET_PATHS = {
2487
        "Aeala/ShareGPT_Vicuna_unfiltered",
2488
    }
2489
2490
2491
2492
    IS_MULTIMODAL = False

    def sample(
        self,
2493
        tokenizer: TokenizerLike,
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
        num_requests: int,
        output_len: int | None = None,
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
        # Filter examples with at least 2 conversations
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
        sampled_requests = []
        ind = 0
        dynamic_output = output_len is None

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

            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
            completion_len = len(completion_ids)
            output_len = completion_len if dynamic_output else output_len
            assert isinstance(output_len, int) and output_len > 0
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
                continue
            mm_content = process_image(item["image"]) if "image" in item else None
            if enable_multimodal_chat:
                # Note: when chat is enabled the request prompt_len is no longer
                # accurate and we will be using request output to count the
                # actual prompt len and output len
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
                    request_id=request_id_prefix + str(ind),
                )
            )
            ind += 1
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
        return sampled_requests


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

    SUPPORTED_DATASET_PATHS = {
        "lmms-lab/LLaVA-OneVision-Data",
    }
2549
    IS_MULTIMODAL = True
2550

2551
2552
    def sample(
        self,
2553
        tokenizer: TokenizerLike,
2554
        num_requests: int,
2555
        output_len: int | None = None,
2556
2557
2558
2559
2560
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2561
        # Filter examples with at least 2 conversations
2562
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
2563
        sampled_requests = []
2564
        ind = 0
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
        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
2579
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
2580
                continue
2581
            mm_content = process_image(item["image"]) if "image" in item else None
2582
2583
2584
2585
            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
2586
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2587
2588
2589
2590
2591
2592
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2593
                    request_id=request_id_prefix + str(ind),
2594
2595
                )
            )
2596
            ind += 1
2597
2598
2599
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
        return sampled_requests


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


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

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2615
2616
        "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"],
2617
    }
2618
    IS_MULTIMODAL = True
2619
2620
2621

    def sample(
        self,
2622
        tokenizer: TokenizerLike,
2623
        num_requests: int,
2624
        output_len: int | None = None,
2625
        enable_multimodal_chat: bool = False,
2626
        request_id_prefix: str = "",
2627
        no_oversample: bool = False,
2628
2629
        **kwargs,
    ) -> list:
2630
2631
2632
2633
        parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
        if parser_fn is None:
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")

2634
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2635

2636
        sampled_requests = []
2637
        for i, item in enumerate(self.data):
2638
2639
            if len(sampled_requests) >= num_requests:
                break
2640

2641
2642
            prompt = parser_fn(item)
            mm_content = process_image(item["images"][0])
2643
            prompt_len = len(tokenizer.encode(prompt))
2644
2645
2646
2647
            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
2648
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2649

2650
2651
2652
2653
2654
2655
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2656
                    request_id=request_id_prefix + str(i),
2657
2658
                )
            )
2659

2660
2661
2662
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2663
2664
2665
        return sampled_requests


2666
2667
2668
2669
2670
2671
2672
2673
class MMVUDataset(HuggingFaceDataset):
    """
    MMVU Dataset.
    https://huggingface.co/datasets/yale-nlp/MMVU
    """

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2674
2675
2676
        "yale-nlp/MMVU": lambda x: x["question"]
        + " "
        + (" ".join(f"{k}.{v}" for k, v in x["choices"].items())),
2677
2678
2679
2680
    }

    def sample(
        self,
2681
        tokenizer: TokenizerLike,
2682
        num_requests: int,
2683
        output_len: int | None = None,
2684
2685
2686
2687
2688
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2689
2690
2691
2692
        parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
        if parser_fn is None:
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")

2693
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2694

2695
2696
2697
2698
        sampled_requests = []
        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
2699

2700
2701
            prompt = parser_fn(item)
            mm_content = process_video(item["video"])
2702
            prompt_len = len(tokenizer.encode(prompt))
2703
2704
2705
2706
            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
2707
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2708

2709
2710
2711
2712
2713
2714
2715
            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),
2716
2717
                )
            )
2718

2719
2720
2721
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2722
2723
2724
        return sampled_requests


2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
# -----------------------------------------------------------------------------
# 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",
    }

2745
2746
    def sample(
        self,
2747
        tokenizer: TokenizerLike,
2748
        num_requests: int,
2749
        output_len: int | None = None,
2750
2751
2752
2753
2754
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
2755
    ) -> list[SampleRequest]:
2756
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2757
        sampled_requests = []
2758
        for i, prompt in enumerate(self.sample_prompts(n=num_requests)):
2759
            # apply template
2760
2761
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2762
                    [{"role": "user", "content": prompt}],
2763
2764
2765
                    add_generation_prompt=True,
                    tokenize=False,
                )
2766

2767
2768
2769
2770
2771
2772
            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2773
                    request_id=request_id_prefix + str(i),
2774
2775
2776
2777
2778
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2779
2780
        return sampled_requests

2781
2782
2783
2784
2785
2786
2787
2788
    def sample_prompts(self, n: int) -> Iterator[str]:
        for item in self.data.take(n):
            prompt = (
                f"{item['input']}\n\n{item['instruction']} Just output "
                "the code, do not include any explanation."
            )
            yield prompt

2789

2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
# -----------------------------------------------------------------------------
# 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,
2812
        tokenizer: TokenizerLike,
2813
        num_requests: int,
2814
        output_len: int | None = None,
2815
        enable_multimodal_chat: bool = False,
2816
        skip_chat_template: bool = False,
2817
        request_id_prefix: str = "",
2818
        no_oversample: bool = False,
2819
2820
        **kwargs,
    ) -> list:
2821
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2822
2823
        sampled_requests = []

2824
        for i, item in enumerate(self.data):
2825
2826
2827
2828
2829
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["turns"][0]

            # apply template
2830
2831
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2832
                    [{"role": "user", "content": prompt}],
2833
2834
2835
                    add_generation_prompt=True,
                    tokenize=False,
                )
2836
2837
2838
2839
2840
2841
2842

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2843
                    request_id=request_id_prefix + str(i),
2844
2845
2846
2847
2848
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2849
2850
2851
        return sampled_requests


2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
# -----------------------------------------------------------------------------
# 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,
2878
        tokenizer: TokenizerLike,
2879
        num_requests: int,
2880
        output_len: int | None = None,
2881
        skip_chat_template: bool = False,
2882
        request_id_prefix: str = "",
2883
        no_oversample: bool = False,
2884
2885
2886
2887
        min_distance: float = 0.0,
        max_distance: float = 1.0,
        **kwargs,
    ) -> list:
2888
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
        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
2901
2902

            # template copied from
2903
            # https://github.com/ise-uiuc/blazedit/blob/7765137e656fd62de877422d2e4cf8de51228054/dataset/create_refined_dataset.py#L94-L105 # noqa: E501
2904
            prompt = f"""Given a code file, please apply the change requests and generate the new file.
2905
2906
2907
2908
2909
2910
2911
2912
2913

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

Change request:
{change_request}

2914
Please generate the new code file in the "New file" section below."""  # noqa: E501
2915
2916

            # apply template
2917
2918
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2919
                    [{"role": "user", "content": prompt}],
2920
2921
2922
                    add_generation_prompt=True,
                    tokenize=False,
                )
2923
2924
2925
2926
2927
2928
2929
2930
2931

            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),
2932
2933
2934
2935
2936
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2937

2938
2939
2940
        return sampled_requests


2941
2942
2943
2944
2945
2946
2947
2948
2949
# -----------------------------------------------------------------------------
# AIMO Dataset Implementation
# -----------------------------------------------------------------------------


class AIMODataset(HuggingFaceDataset):
    """
    Dataset class for processing a AIMO dataset with reasoning questions.
    """
2950

2951
    SUPPORTED_DATASET_PATHS = {
2952
2953
2954
        "AI-MO/aimo-validation-aime",
        "AI-MO/NuminaMath-1.5",
        "AI-MO/NuminaMath-CoT",
2955
2956
    }

2957
2958
    def sample(
        self,
2959
        tokenizer: TokenizerLike,
2960
        num_requests: int,
2961
        output_len: int | None = None,
2962
2963
2964
2965
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2966
        sampled_requests = []
2967
        ind = 0
2968
2969
2970
2971
2972
        dynamic_output = output_len is None

        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
2973
            prompt, completion = item["problem"], item["solution"]
2974
2975
2976
2977
2978
2979
2980

            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
2981
2982
2983
            if dynamic_output and not is_valid_sequence(
                prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000
            ):
2984
2985
2986
2987
2988
2989
2990
                continue
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=None,
2991
                    request_id=request_id_prefix + str(ind),
2992
2993
                )
            )
2994
            ind += 1
2995
2996
2997
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2998
        return sampled_requests
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018


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

3019
"""  # noqa: E501
3020
3021
3022


def _format_zeta_prompt(
3023
3024
    sample: dict, original_start_marker: str = "<|editable_region_start|>"
) -> dict:
3025
    """Format the zeta prompt for the Next Edit Prediction (NEP) dataset.
3026
3027
3028

    This function formats examples from the NEP dataset
    into prompts and expected outputs. It could be
3029
    further extended to support more NEP datasets.
3030

3031
    Args:
3032
        sample: The dataset sample containing events,
3033
            inputs, and outputs.
3034
3035
        original_start_marker: The marker indicating the
            start of the editable region. Defaults to
3036
            "<|editable_region_start|>".
3037

3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
    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,
    }

3067
3068
    def sample(
        self,
3069
        tokenizer: TokenizerLike,
3070
3071
3072
3073
3074
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ):
3075
        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.hf_name)
3076
        if formatting_prompt_func is None:
3077
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")
3078
        samples = []
3079
        for i, sample in enumerate(self.data):
3080
3081
3082
3083
3084
3085
            sample = formatting_prompt_func(sample)
            samples.append(
                SampleRequest(
                    prompt=sample["prompt"],
                    prompt_len=len(tokenizer(sample["prompt"]).input_ids),
                    expected_output_len=len(
3086
3087
                        tokenizer(sample["expected_output"]).input_ids
                    ),
3088
                    request_id=request_id_prefix + str(i),
3089
3090
                )
            )
3091
3092
            if len(samples) >= num_requests:
                break
3093
3094
3095
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
3096
        return samples
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131


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

3132
    DEFAULT_OUTPUT_LEN = 1024
3133
3134
3135
3136
    IS_MULTIMODAL = True

    def sample(
        self,
3137
        tokenizer: TokenizerLike,
3138
        num_requests: int,
3139
        output_len: int | None = None,
3140
        request_id_prefix: str = "",
3141
        no_oversample: bool = False,
3142
3143
        **kwargs,
    ) -> list:
3144
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
3145
3146
3147
3148
        if "openai" in tokenizer.name_or_path:
            prompt = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>"
        else:
            prompt = ""
3149
3150
        prompt_len = len(tokenizer(prompt).input_ids)
        sampled_requests = []
3151
        ind = 0
3152
        skipped = 0
3153
3154
3155
        asr_min_audio_len_sec = kwargs.get("asr_min_audio_len_sec")
        asr_max_audio_len_sec = kwargs.get("asr_max_audio_len_sec")
        durations = []
3156
3157
3158
3159
3160
3161
        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)
3162
            if duration_s < asr_min_audio_len_sec or duration_s > asr_max_audio_len_sec:
3163
3164
3165
                skipped += 1
                continue

3166
            durations.append(duration_s)
3167
3168
3169
3170
3171
3172
3173
            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,
3174
                    request_id=request_id_prefix + str(ind),
3175
3176
                )
            )
3177
            ind += 1
3178
3179
3180
3181
3182
3183
3184
        if skipped:
            logger.warning(
                "%d samples discarded from dataset due to"
                " their length being greater than"
                " what Whisper supports.",
                skipped,
            )
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198

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

3199
3200
3201
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3202
        return sampled_requests
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234


# -----------------------------------------------------------------------------
# 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,
3235
        tokenizer: TokenizerLike,
3236
        num_requests: int,
3237
        output_len: int | None = None,
3238
        request_id_prefix: str = "",
3239
        no_oversample: bool = False,
3240
3241
3242
3243
3244
        **kwargs,
    ) -> list[SampleRequest]:
        # Force dynamic output length based on reference completion.
        dynamic_output = output_len is None
        sampled_requests: list[SampleRequest] = []
3245
        ind = 0
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279

        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,
3280
                    request_id=request_id_prefix + str(ind),
3281
3282
                )
            )
3283
            ind += 1
3284

3285
3286
3287
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3288
        return sampled_requests
3289
3290
3291
3292
3293
3294
3295
3296


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


class PrefixRepetitionRandomDataset(BenchmarkDataset):
3297
    # Default values copied from benchmark_serving.py for the repeated prefix
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
    # 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,
3314
        tokenizer: TokenizerLike,
3315
3316
3317
3318
3319
        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,
3320
        request_id_prefix: str = "",
3321
        no_oversample: bool = False,
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
        **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
3336
            tokens = np.random.randint(0, vocab_size, size=target_length).tolist()
3337

3338
            _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(  # noqa: E501
3339
3340
3341
3342
3343
3344
                tokenizer=tokenizer,
                token_sequence=tokens,
                target_token_len=target_length,
                add_special_tokens=False,
            )
            return adjusted_tokens, token_mismatch
3345
3346

        requests = []
3347
        token_mismatch_total = 0
3348
        for _ in range(num_prefixes):
3349
3350
            prefix_tokens, prefix_mismatch = _generate_exact_length_tokens(prefix_len)
            token_mismatch_total += prefix_mismatch
3351
3352

            for _ in range(prompts_per_prefix):
3353
                suffix_tokens, suffix_mismatch = _generate_exact_length_tokens(
3354
                    suffix_len
3355
                )
3356
                token_mismatch_total += suffix_mismatch
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
                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,
                    )
                )

3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
        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,
            )
3378
3379
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(requests)
3380
        return requests
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392


# -----------------------------------------------------------------------------
# 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
    """
3393

3394
3395
3396
3397
3398
3399
    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {"Lin-Chen/MMStar"}
    IS_MULTIMODAL = True

    def sample(
        self,
3400
        tokenizer: TokenizerLike,
3401
        num_requests: int,
3402
        output_len: int | None = None,
3403
3404
3405
3406
3407
3408
        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.
3409
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
        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