datasets.py 129 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
import io
import json
import logging
20
import math
21
22
import random
from abc import ABC, abstractmethod
23
from collections.abc import Callable, Iterator, Mapping
24
from contextlib import suppress
25
from copy import deepcopy
26
27
28
from dataclasses import dataclass
from functools import cache
from io import BytesIO
29
from tempfile import NamedTemporaryFile
30
from typing import Any, cast
31
32

import numpy as np
33
import pybase64 as base64
34
from huggingface_hub import snapshot_download
35
from PIL import Image
36
from typing_extensions import deprecated
37
38
39
40

from vllm.lora.request import LoRARequest
from vllm.lora.utils import get_adapter_absolute_path
from vllm.multimodal import MultiModalDataDict
41
from vllm.multimodal.image import convert_image_mode
42
from vllm.tokenizers import TokenizerLike
43
from vllm.utils.argparse_utils import FlexibleArgumentParser
44
from vllm.utils.import_utils import PlaceholderModule
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

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")
61
62
63

logger = logging.getLogger(__name__)

64
65
DEFAULT_NUM_PROMPTS = 1000

66
67
68
69
70
71
72
73
74
75
76
# -----------------------------------------------------------------------------
# Data Classes
# -----------------------------------------------------------------------------


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

77
    prompt: str | list[str]
78
79
    prompt_len: int
    expected_output_len: int
80
81
82
    multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None
    lora_request: LoRARequest | None = None
    request_id: str | None = None
83
84
85
86
87
88
89
90
91


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


class BenchmarkDataset(ABC):
    DEFAULT_SEED = 0
92
    IS_MULTIMODAL = False
93
94
95

    def __init__(
        self,
96
        dataset_path: str | None = None,
97
        random_seed: int = DEFAULT_SEED,
98
99
        disable_shuffle: bool = False,
        **kwargs,
100
101
102
    ) -> None:
        """
        Initialize the BenchmarkDataset with an optional dataset path and random
103
104
        seed.

105
106
        Args:
            dataset_path (Optional[str]): Path to the dataset. If None, it
107
                indicates that a default or random dataset might be used.
108
            random_seed (int): Seed value for reproducible shuffling or
109
                sampling. Defaults to DEFAULT_SEED.
110
111
112
113
        """
        self.dataset_path = dataset_path
        # Set the random seed, ensuring that a None value is replaced with the
        # default seed.
114
        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
115
        self.disable_shuffle = disable_shuffle
116
117
118
        self.data = None

    def apply_multimodal_chat_transformation(
119
120
        self,
        prompt: str,
121
        mm_content: MultiModalDataDict | dict | list[dict] | None = None,
122
    ) -> list[dict]:
123
124
125
126
127
128
129
        """
        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:
130
131
132
133
134
            if isinstance(mm_content, list):
                content.extend(cast(list[dict[str, Any]], mm_content))
            elif isinstance(mm_content, dict):
                content.append(mm_content)
            else:
135
                raise TypeError(
136
                    f"Could not process multimodal content of type: {type(mm_content)}"
137
                )
138
139
140
141
142
143
144
145
146
147
148
149
150
        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
151
        raise NotImplementedError("load_data must be implemented in subclasses.")
152
153
154

    def get_random_lora_request(
        self,
155
156
157
        max_loras: int | None = None,
        lora_path: str | None = None,
    ) -> LoRARequest | None:
158
        """
159
        Optionally select a random LoRA request.
160
161

        This method is used when LoRA parameters are provided.  It randomly
162
        selects a LoRA based on max_loras.
163
164

        Args:
165
166
167
168
            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.
169
170

        Returns:
171
172
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
173
174
        """
        if max_loras is None or lora_path is None:
175
            return None
176
177
178
179
180
181
182
183

        # 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),
        )
184
        return lora_request
185

186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
    def get_round_robin_lora_request(
        self,
        index: int,
        max_loras: int | None = None,
        lora_path: str | None = None,
    ) -> LoRARequest | None:
        """
        Optionally select a LoRA request using deterministic round-robin.

        This method cycles through LoRA IDs in order based on the request
        index, providing reproducible LoRA assignment.

        Args:
            index (int): The request index used for round-robin selection.
            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.

        Returns:
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
        """
        if max_loras is None or lora_path is None:
            return None

        # Deterministic round-robin: cycle through [1, max_loras]
        lora_id = index % max_loras + 1
        lora_request = LoRARequest(
            lora_name=str(lora_id),
            lora_int_id=lora_id,
            lora_path=lora_path_on_disk(lora_path),
        )
        return lora_request

    def get_lora_request(
        self,
        index: int,
        max_loras: int | None = None,
        lora_path: str | None = None,
        lora_assignment: str = "random",
    ) -> LoRARequest | None:
        """
        Select a LoRA request using the specified assignment strategy.

        Args:
            index (int): The request index (used for round-robin).
            max_loras (Optional[int]): The maximum number of LoRAs available.
            lora_path (Optional[str]): Path to the LoRA parameters on disk.
            lora_assignment (str): Strategy for LoRA selection.
                'random' (default) or 'round-robin'.

        Returns:
            A new [`LoRARequest`][vllm.lora.request.LoRARequest]
            (or `None` if not applicable).
        """
        if lora_assignment == "round-robin":
            return self.get_round_robin_lora_request(
                index=index, max_loras=max_loras, lora_path=lora_path
            )
        return self.get_random_lora_request(max_loras=max_loras, lora_path=lora_path)

248
    @abstractmethod
249
250
    def sample(
        self,
251
        tokenizer: TokenizerLike,
252
253
254
255
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
    ) -> list[SampleRequest]:
256
257
258
259
260
261
262
        """
        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:
263
            tokenizer (TokenizerLike): The tokenizer to be used
264
                for processing the dataset's text.
265
            num_requests (int): The number of sample requests to generate.
266
            request_id_prefix (str): The prefix of request_id.
267
268
269
270
271
272
273

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

274
275
276
277
278
    def maybe_oversample_requests(
        self,
        requests: list[SampleRequest],
        num_requests: int,
        request_id_prefix: str = "",
279
        no_oversample: bool = False,
280
    ) -> None:
281
282
283
284
285
286
        """
        Oversamples the list of requests if its size is less than the desired
        number.

        Args:
            requests (List[SampleRequest]): The current list of sampled
287
288
                requests.
            num_requests (int): The target number of requests.
289
290
            request_id_prefix (str): The prefix applied to generated request
                identifiers.
291

292
        """
293
        if no_oversample:
294
            logger.info("Skipping oversampling. Total samples: %d.", len(requests))
295
296
            return

297
298
        if len(requests) < num_requests:
            random.seed(self.random_seed)
299
300
301
302
            needed = num_requests - len(requests)
            additional = []
            for i in range(needed):
                req = deepcopy(random.choice(requests))
303
                req.request_id = request_id_prefix + str(len(requests) + i)
304
                additional.append(req)
305
            requests.extend(additional)
306
            logger.info("Oversampled requests to reach %d total samples.", num_requests)
307

308
309
        ids = [req.request_id for req in requests]
        if len(ids) != len(set(ids)):
310
311
312
313
314
            raise ValueError(
                "Duplicate request_id found in the sampled "
                "requests. Please ensure that each request_id "
                "is unique."
            )
315

316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338

# -----------------------------------------------------------------------------
# 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
339
    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
340
341
342
343
    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
344
345
346
    return not (
        prompt_too_short or output_too_short or prompt_too_long or combined_too_long
    )
347
348
349
350
351
352
353
354


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


# Global cache for LoRA tokenizers.
355
lora_tokenizer_cache: dict[int, TokenizerLike] = {}
356
357
358
359
360
361


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

362
    Supports the following input types:
363
364
365
366
367
368
369
370

    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.

371
372
373
374
375
    3. String input: - Treats the string as a URL, local file path, or base64
       encoded data.  - If string starts with "data:image/", treats as base64.
       - If string starts with "http://", "https://", or "file://", treats as URL.
       - Otherwise treats as local file path and prepends "file://".
       - Returns a dictionary with the image URL or base64 data.
376
377
378
379

    Raises:
        ValueError: If the input is not a supported type.
    """
380
381
    if isinstance(image, dict) and "bytes" in image:
        image = Image.open(BytesIO(image["bytes"]))
382
    if isinstance(image, Image.Image):
383
        image = convert_image_mode(image, "RGB")
384
385
        with io.BytesIO() as image_data:
            image.save(image_data, format="JPEG")
386
            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
387
388
        return {
            "type": "image_url",
389
            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
390
391
392
        }

    if isinstance(image, str):
393
394
        image_url = (
            image
395
            if image.startswith(("http://", "https://", "file://", "data:image/"))
396
397
            else f"file://{image}"
        )
398
399
        return {"type": "image_url", "image_url": {"url": image_url}}

400
    raise ValueError(
401
402
        f"Invalid image input {image}. Must be a PIL.Image.Image, "
        "str (URL, file path, or base64 data URL), or dictionary with raw image bytes."
403
    )
404
405


406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
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.
    """
422
423
    if isinstance(video, dict) and "bytes" in video:
        video_bytes = video["bytes"]
424
425
426
        video_base64 = base64.b64encode(video_bytes).decode("utf-8")
        return {
            "type": "video_url",
427
            "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
428
429
430
        }

    if isinstance(video, str):
431
432
433
434
435
        video_url = (
            video
            if video.startswith(("http://", "https://", "file://"))
            else f"file://{video}"
        )
436
437
438
439
440
441
        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
    )

442
443

def gen_prompt_decode_to_target_len(
444
    tokenizer: TokenizerLike,
445
446
447
448
    token_sequence: list[int],
    target_token_len: int,
    max_retry: int = 10,
    add_special_tokens: bool = False,
449
    rng: np.random.Generator | None = None,
450
) -> tuple[str, list[int], int]:
451
452
453
454
    """
    Ensure decoded-then-encoded prompt length matches the target token length.

    This function decodes an initial token sequence to text and re-encodes it
455
456
    , iteratively adjusting the token sequence length to match a target.
    This is necessary because some tokenizers do not guarantee a 1:1 mapping
457
458
459
460
461
    between consecutive tokens and the decoded-then-encoded sequence length.
    For example, for GPT2Tokenizer:
    [6880, 6881] -> ['Ġcalls', 'here'] ->
    [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']

462
463
464
    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.
465
466
467
468
469
    """
    remain_num_try = max_retry
    token_mismatch = 0
    while True:
        prompt = tokenizer.decode(token_sequence)
470
        token_sequence = tokenizer.encode(prompt, add_special_tokens=add_special_tokens)
471
472
473
474
        if remain_num_try <= 0:
            if len(token_sequence) != target_token_len:
                token_mismatch = len(token_sequence) - target_token_len
            break
475

476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
        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

499

500
501
502
503
# -----------------------------------------------------------------------------
# Random Dataset Implementation (Synthetic Data)
# -----------------------------------------------------------------------------

504

505
class RandomDataset(BenchmarkDataset):
506
507
508
509
510
511
512
513
514
515
516
517
    """
    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.
    """
518

519
520
521
522
523
524
    # 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

525
    def __init__(self, **kwargs) -> None:
526
        super().__init__(**kwargs)
527
528
529
530
        # 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)
531
532
533

    def sample(
        self,
534
        tokenizer: TokenizerLike,
535
        num_requests: int,
536
        request_id_prefix: str = "",
537
        no_oversample: bool = False,
538
539
540
541
        prefix_len: int = DEFAULT_PREFIX_LEN,
        range_ratio: float = DEFAULT_RANGE_RATIO,
        input_len: int = DEFAULT_INPUT_LEN,
        output_len: int = DEFAULT_OUTPUT_LEN,
542
        batchsize: int = 1,
543
544
545
        max_loras: int | None = None,
        lora_path: str | None = None,
        lora_assignment: str = "random",
546
547
        **kwargs,
    ) -> list[SampleRequest]:
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
        # 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."
            )

564
565
        input_lens, output_lens, offsets = self.get_sampling_params(
            num_requests, range_ratio, input_len, output_len, tokenizer
566
567
568
        )

        vocab_size = tokenizer.vocab_size
569
570
571
572
573
        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
574
        prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len)
575

576
        requests = []
577
        token_mismatch_total = 0
578
        for i in range(num_requests):
579
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
580
581
582
583
584
585
586
                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,
587
                allowed_tokens=allowed_tokens,
588
            )
589
            token_mismatch_total += token_mismatch
590
591
592
593
594
595
            lora_req = self.get_lora_request(
                index=i,
                max_loras=max_loras,
                lora_path=lora_path,
                lora_assignment=lora_assignment,
            )
596
597
598
599
600
            requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
601
                    lora_request=lora_req,
602
603
604
                    request_id=request_id_prefix + str(i),
                )
            )
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
        # 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
620

621
622
623
624
625
626
627
628
629
630
631
        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,
            )

632
633
634
        return requests

    def get_prefix(
635
        self,
636
        tokenizer: TokenizerLike,
637
638
        allowed_tokens: np.ndarray,
        prefix_len: int,
639
640
641
642
    ) -> list[int]:
        """
        Get the prefix for the dataset.
        """
643
644
645
646
647
648
649
650
651
652
653
654
        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,
655
        )
656
657
658
659
660
661
662
663
664
665
        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
666

667
668
669
670
671
672
    def get_sampling_params(
        self,
        num_requests: int,
        range_ratio: float,
        input_len: int,
        output_len: int,
673
        tokenizer: TokenizerLike,
674
675
676
677
678
679
680
681
682
683
684
685
686
687
    ) -> 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))
688
689
        # Ensure the lower bound for output length is at least 1 to
        # prevent sampling 0 tokens.
690
        output_low = max(output_low, 1)
691
        output_high = max(output_high, 1)
692
693
694

        if input_low > input_high:
            raise ValueError(
695
                f"Invalid input sampling interval: low={input_low} > high={input_high}"
696
697
698
699
700
701
            )
        if output_low > output_high:
            raise ValueError(
                "Invalid output sampling interval: "
                f"low={output_low} > high={output_high}"
            )
702

703
704
        logger.info(
            "Sampling input_len from [%s, %s] and output_len from [%s, %s]",
705
706
707
708
709
            input_low,
            input_high,
            output_low,
            output_high,
        )
710

711
712
713
        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)
714
        return input_lens, output_lens, offsets
715

716
717
718
    def generate_token_sequence(
        self,
        *,
719
        tokenizer: TokenizerLike,
720
721
722
723
724
725
        prefix_token_ids: list[int],
        prefix_len: int,
        vocab_size: int,
        input_len: int,
        offset: int,
        index: int,
726
        allowed_tokens: np.ndarray,
727
    ) -> tuple[str, int, int]:
728
729
730
731
732
733
734
735
736
737
        """
        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,
738
        the encoded sequence is truncated before being decoded again.
739
        """
740
741
742
743
744
        # 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()
745
746
747
748
        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)
749
        prompt, adjusted_token_sequence, token_mismatch = (
750
            gen_prompt_decode_to_target_len(
751
752
753
754
755
756
                tokenizer=tokenizer,
                token_sequence=token_sequence,
                target_token_len=total_input_len,
                add_special_tokens=False,
                rng=self._rng,
            )
757
758
759
        )
        total_input_len = len(adjusted_token_sequence)
        return prompt, total_input_len, token_mismatch
760
761


762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
# -----------------------------------------------------------------------------
# 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,
779
        tokenizer: TokenizerLike,
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
        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
        )
809

810
        vocab_size = tokenizer.vocab_size
811
812
813
        prohibited_tokens = tokenizer.all_special_ids
        all_tokens = np.arange(vocab_size)
        allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens)))
814
815
816
817
818
819
820
821
822
823

        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,
824
                allowed_tokens=allowed_tokens,
825
826
827
828
829
830
831
832
833
834
835
836
837
            )
        )

        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,
838
                allowed_tokens=allowed_tokens,
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
            )
            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


874
875
876
877
# -----------------------------------------------------------------------------
# MultiModalDataset Implementation
# -----------------------------------------------------------------------------

878

879
880
881
882
883
884
class RandomMultiModalDataset(RandomDataset):
    """
    Synthetic multimodal dataset (text + images) that extends RandomDataset.

    Status:
    - Images: supported via synthetic RGB data.
885
    - Video: supported via synthetic RGB data.
886
887
888
889
890
891
892
893
    - 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
894
       mapping (height, width, num_frames) → probability. We treat
895
       `num_frames`=1 as image and `num_frames` > 1 as video.
896
       Entries with zero probability are removed and the rest are renormalized
897
898
899
900
901
902
903
       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}
904
905
      - Two image buckets (`num_frames`=1) and one video bucket
      (`num_frames`=16).
906
907
908
909
    OBS.: Only image sampling is supported for now.
    """

    IS_MULTIMODAL = True
910
    DEFAULT_LIMIT_MM_PER_PROMPT = {"image": 255, "video": 1}
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925

    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.
926
927
928

        NOTE: iid pixel sampling results in worst-case compression
        (good for stressing I/O), but very unlike real photos.
929
930
931
932
933
934
935
936
937
938
939
        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)

940
941
942
    def generate_synthetic_video(
        self, width: int, height: int, num_frames: int
    ) -> dict:
943
        """Generate synthetic video with random values.
944

945
946
        Creates a video with random pixel values, encodes it to MP4 format,
        and returns the content as bytes.
947
        """
948
949
        import cv2

950
951
952
953
954
955
956
957
958
959
960
        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

961
        with NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
            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}
983
984
985
986
987
988
989
990
991
992

    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}")

993
994
995
    def normalize_bucket_config(
        self, bucket_config: dict[tuple[int, int, int], float]
    ) -> dict[tuple[int, int, int], float]:
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
        """
        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:
1007
1008
1009
            raise ValueError(
                "Got invalid bucket config. Bucket config values must be non-zero."
            )
1010
1011
1012
1013
        # 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()}

1014
1015
1016
1017
    def generate_mm_item(
        self,
        mm_item_config: tuple[int, int, int],
    ) -> Mapping[str, Any]:
1018
        """
1019
        Create synthetic images and videos and
1020
1021
1022
1023
        apply process_image/process_video respectively.
        This follows the OpenAI API chat completions
        https://github.com/openai/openai-python
        """
1024

1025
        if self.map_config_to_modality(mm_item_config) == "image":
1026
1027
1028
            return process_image(
                self.generate_synthetic_image(mm_item_config[1], mm_item_config[0])
            )
1029
        elif self.map_config_to_modality(mm_item_config) == "video":
1030
1031
1032
1033
1034
            return process_video(
                self.generate_synthetic_video(
                    mm_item_config[1], mm_item_config[0], mm_item_config[2]
                )
            )
1035
        else:
1036
            raise ValueError(f"Invalid multimodal item configuration: {mm_item_config}")
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056

    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:
1057
1058
1059
1060
1061
                raise ValueError(
                    f"Modality {modality} is not in "
                    f"limit_mm_per_prompt: "
                    f"{limit_mm_per_prompt.keys()}"
                )
1062

1063
        # Remove zero probability entries
1064
1065
1066
        # and normalize bucket config to sum to 1
        bucket_config = self.normalize_bucket_config(bucket_config)
        logger.info(
1067
1068
            "Normalized bucket config: %s",
            bucket_config,
1069
1070
        )
        # Only consider limit per prompt for modalities in bucket config
1071
        allowed_modalities = {self.map_config_to_modality(cfg) for cfg in bucket_config}
1072
        limit_mm_per_prompt = {
1073
1074
            k: v for k, v in limit_mm_per_prompt.items() if k in allowed_modalities
        }
1075
        if not limit_mm_per_prompt:
1076
            raise ValueError("No valid limits for modalities present in bucket_config.")
1077
1078

        logger.info(
1079
1080
            "Updated mm-limit-per-prompt: %s",
            limit_mm_per_prompt,
1081
1082
1083
1084
1085
        )

        # 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(
1086
            sum(limit_mm_per_prompt.values()),
1087
            math.ceil(base_items_per_request * (1 + num_mm_items_range_ratio)),
1088
1089
1090
        )
        # Ensure min num mm items is at least 0
        min_num_mm_items = max(
1091
            0, math.floor(base_items_per_request * (1 - num_mm_items_range_ratio))
1092
1093
1094
        )
        # Raise error if min num mm items is greater than max num mm items
        if min_num_mm_items > max_num_mm_items:
1095
1096
1097
1098
            raise ValueError(
                f"Min num mm items is greater than max mm items: "
                f"{min_num_mm_items} > {max_num_mm_items}"
            )
1099

1100
1101
        logger.info(
            "Sampling number of multimodal items from [%s, %s]",
1102
1103
            min_num_mm_items,
            max_num_mm_items,
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
        )

        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],
1119
    ) -> Iterator[tuple[int, int, int]]:
1120
1121
1122
1123
1124
        """
        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.
1125
1126
        Loop until the number of multimodal items sampled is equal to
        request_num_mm_items or limit of multimodal items per prompt
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
        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)
1138
        )
1139
1140
1141
1142
        # If request_num_mm_items is 0, yield an empty iterator
        if request_num_mm_items == 0:
            return
        # Initialize modality counters
1143
        modality_counter = {self.map_config_to_modality(k): 0 for k in bucket_config}
1144
1145
1146
1147
1148
        # 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
1149
1150
1151
            mm_item_config = self._rng.choice(
                list(bucket_config_copy.keys()), p=list(bucket_config_copy.values())
            )
1152
1153
1154
1155
            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
1156
                yield (mm_item_config)
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
            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()):
1167
1168
1169
                    logger.warning(
                        "Exhausted all multimodal items of modality %s", modality
                    )
1170
1171
                    break
                # Renormalize the bucket config
1172
                bucket_config_copy = self.normalize_bucket_config(bucket_config_copy)
1173
1174
1175

    def sample(
        self,
1176
        tokenizer: TokenizerLike,
1177
1178
        num_requests: int,
        request_id_prefix: str = "",
1179
        no_oversample: bool = False,
1180
1181
1182
1183
1184
1185
1186
        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,
1187
1188
1189
        bucket_config: dict[
            tuple[int, int, int], float
        ] = DEFAULT_MM_ITEM_BUCKET_CONFIG,
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
        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
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
        # 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
1227
        prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len)
1228
1229
        # Add synthetic multimodal items to each request
        mm_requests = []
1230
        token_mismatch_total = 0
1231
        for i in range(num_requests):
1232
            prompt, total_input_len, token_mismatch = self.generate_token_sequence(  # noqa: E501
1233
1234
1235
1236
1237
1238
1239
                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,
1240
                allowed_tokens=allowed_tokens,
1241
            )
1242
            token_mismatch_total += token_mismatch
1243
1244
1245
1246
1247
1248
1249
1250
            # 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,
            )

1251
1252
1253
1254
1255
1256
1257
            mm_content = cast(
                list[dict[str, Any]],
                [
                    self.generate_mm_item(mm_item_config)
                    for mm_item_config in mm_item_iterator
                ],
            )
1258
1259

            if enable_multimodal_chat:
1260
                # NOTE: For now this option is only provided for completeness
1261
1262
1263
                # given that the serve.py benchmark currently does not use it.
                mm_chat_prompt: Any = prompt
                mm_chat_prompt = self.apply_multimodal_chat_transformation(
1264
1265
                    prompt, mm_content
                )
1266
1267
1268
1269
1270
1271
1272
1273
1274
                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(
1275
1276
1277
                    prompt=prompt,
                    prompt_len=total_input_len,
                    expected_output_len=int(output_lens[i]),
1278
                    multi_modal_data=mm_content,
1279
                    request_id=request_id_prefix + str(i),
1280
1281
                )
            mm_requests.append(sample_request)
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293

        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,
            )

1294
        return mm_requests
1295

1296

1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
# -----------------------------------------------------------------------------
# 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 = [
1320
1321
            entry
            for entry in self.data
1322
1323
1324
            if "conversations" in entry and len(entry["conversations"]) >= 2
        ]
        random.seed(self.random_seed)
1325
1326
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
1327
1328
1329

    def sample(
        self,
1330
        tokenizer: TokenizerLike,
1331
        num_requests: int,
1332
1333
1334
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
1335
        enable_multimodal_chat: bool = False,
1336
        request_id_prefix: str = "",
1337
        no_oversample: bool = False,
1338
        lora_assignment: str = "random",
1339
1340
1341
        **kwargs,
    ) -> list:
        samples: list = []
1342
        ind = 0
1343
1344
1345
1346
1347
1348
1349
1350
        for entry in self.data:
            if len(samples) >= num_requests:
                break
            prompt, completion = (
                entry["conversations"][0]["value"],
                entry["conversations"][1]["value"],
            )

1351
1352
1353
1354
1355
            lora_request = self.get_lora_request(
                index=ind,
                max_loras=max_loras,
                lora_path=lora_path,
                lora_assignment=lora_assignment,
1356
            )
1357
1358
1359
            prompt_ids = tokenizer(prompt).input_ids
            completion_ids = tokenizer(completion).input_ids
            prompt_len = len(prompt_ids)
1360
1361
1362
1363
1364
1365
            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,
            ):
1366
                continue
1367
1368
1369
            if image_path := entry.get("image"):
                mm_content = process_image(image_path)
            elif video_path := entry.get("video"):
1370
                mm_content = process_video(video_path)
1371
            else:
1372
                mm_content = None
1373
            if enable_multimodal_chat:
1374
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
1375
1376
1377
1378
1379
1380
            samples.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=new_output_len,
                    lora_request=lora_request,
1381
                    multi_modal_data=mm_content,
1382
                    request_id=request_id_prefix + str(ind),
1383
1384
                )
            )
1385
            ind += 1
1386
1387
1388
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
1389
1390
1391
        return samples


1392
1393
class _ValidateDatasetArgs(argparse.Action):
    """Argparse action to validate dataset name and path compatibility."""
1394

1395
1396
    def __call__(self, parser, namespace, values, option_string=None):
        setattr(namespace, self.dest, values)
1397

1398
        # Get current values of both dataset_name and dataset_path
1399
1400
        dataset_name = getattr(namespace, "dataset_name", "random")
        dataset_path = getattr(namespace, "dataset_path", None)
1401

1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
        # 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}"
            )


1412
def add_dataset_parser(parser: FlexibleArgumentParser):
1413
1414
1415
1416
1417
    parser.add_argument(
        "--trust-remote-code",
        action="store_true",
        help="Trust remote code from huggingface",
    )
1418
1419
1420
1421
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument(
        "--num-prompts",
        type=int,
1422
        default=DEFAULT_NUM_PROMPTS,
1423
1424
1425
1426
1427
1428
        help="Number of prompts to process.",
    )
    parser.add_argument(
        "--dataset-name",
        type=str,
        default="random",
1429
        action=_ValidateDatasetArgs,
1430
        choices=[
1431
1432
1433
1434
1435
            "sharegpt",
            "burstgpt",
            "sonnet",
            "random",
            "random-mm",
1436
            "random-rerank",
1437
1438
            "hf",
            "custom",
1439
            "custom_mm",
1440
1441
            "prefix_repetition",
            "spec_bench",
1442
        ],
1443
1444
        help="Name of the dataset to benchmark on.",
    )
1445
1446
1447
1448
1449
    parser.add_argument(
        "--no-stream",
        action="store_true",
        help="Do not load the dataset in streaming mode.",
    )
1450
1451
1452
1453
    parser.add_argument(
        "--dataset-path",
        type=str,
        default=None,
1454
        action=_ValidateDatasetArgs,
1455
1456
1457
        help="Path to the sharegpt/sonnet dataset. "
        "Or the huggingface dataset ID if using HF dataset.",
    )
1458
1459
1460
    parser.add_argument(
        "--no-oversample",
        action="store_true",
1461
        help="Do not oversample if the dataset has fewer samples than num-prompts.",
1462
    )
1463
1464
1465
    parser.add_argument(
        "--skip-chat-template",
        action="store_true",
1466
        help="Skip applying chat template to prompt for datasets that support it.",
1467
    )
1468
1469
1470
1471
1472
    parser.add_argument(
        "--enable-multimodal-chat",
        action="store_true",
        help="Enable multimodal chat transformation for datasets that support it.",
    )
1473
1474
1475
1476
1477
    parser.add_argument(
        "--disable-shuffle",
        action="store_true",
        help="Disable shuffling of dataset samples for deterministic ordering.",
    )
1478
1479
1480
1481
1482
1483
1484

    # 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,
1485
1486
1487
        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.",
1488
1489
    )

1490
1491
1492
1493
1494
    spec_bench_group = parser.add_argument_group("spec bench dataset options")
    spec_bench_group.add_argument(
        "--spec-bench-output-len",
        type=int,
        default=256,
1495
        help="Num of output tokens per request, used only for spec bench dataset.",
1496
1497
1498
1499
1500
    )
    spec_bench_group.add_argument(
        "--spec-bench-category",
        type=str,
        default=None,
1501
        help="Category for spec bench dataset. If None, use all categories.",
1502
1503
    )

1504
1505
1506
1507
1508
    sonnet_group = parser.add_argument_group("sonnet dataset options")
    sonnet_group.add_argument(
        "--sonnet-input-len",
        type=int,
        default=550,
1509
        help="Number of input tokens per request, used only for sonnet dataset.",
1510
1511
1512
1513
1514
    )
    sonnet_group.add_argument(
        "--sonnet-output-len",
        type=int,
        default=150,
1515
        help="Number of output tokens per request, used only for sonnet dataset.",
1516
1517
1518
1519
1520
    )
    sonnet_group.add_argument(
        "--sonnet-prefix-len",
        type=int,
        default=200,
1521
        help="Number of prefix tokens per request, used only for sonnet dataset.",
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
    )

    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.",
    )

1533
1534
1535
1536
1537
    blazedit_group = parser.add_argument_group("blazedit dataset options")
    blazedit_group.add_argument(
        "--blazedit-min-distance",
        type=float,
        default=0.0,
1538
        help="Minimum distance for blazedit dataset. Min: 0, Max: 1.0",
1539
1540
1541
1542
1543
    )
    blazedit_group.add_argument(
        "--blazedit-max-distance",
        type=float,
        default=1.0,
1544
        help="Maximum distance for blazedit dataset. Min: 0, Max: 1.0",
1545
1546
    )

1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
    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.",
    )

1561
    random_group = parser.add_argument_group("random dataset options")
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
    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(
1641
1642
1643
        "--random-input-len",
        type=int,
        default=1024,
1644
        help="Number of input tokens per request, used only for random sampling.",
1645
    )
1646
    parser_or_group.add_argument(
1647
1648
1649
        "--random-output-len",
        type=int,
        default=128,
1650
        help="Number of output tokens per request, used only for random sampling.",
1651
    )
1652
    parser_or_group.add_argument(
1653
1654
1655
1656
1657
1658
1659
1660
        "--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)].",
    )
1661
    parser_or_group.add_argument(
1662
1663
1664
        "--random-prefix-len",
        type=int,
        default=0,
1665
1666
1667
1668
1669
1670
1671
1672
        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)]."
        ),
1673
    )
1674
    parser_or_group.add_argument(
1675
1676
1677
        "--random-batch-size",
        type=int,
        default=1,
1678
        help=("Batch size for random sampling. Only used for embeddings benchmark."),
1679
    )
1680
    parser_or_group.add_argument(
1681
1682
1683
1684
1685
1686
1687
        "--no-reranker",
        action="store_true",
        help=(
            "Whether the model supports reranking natively."
            " Only used for reranker benchmark."
        ),
    )
1688

1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701

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(
1702
1703
1704
1705
1706
1707
1708
1709
1710
        "--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."
        ),
    )
1711
    parser_or_group.add_argument(
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
        "--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."
        ),
    )
1726
    parser_or_group.add_argument(
1727
1728
1729
1730
1731
        "--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. "
1732
            '\'{"image": 3, "video": 0}\'. The sampled per-request item '
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
            "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)
1749
1750
1751
1752
1753
                if not (
                    isinstance(key, tuple)
                    and len(key) == 3
                    and all(isinstance(x, int) for x in key)
                ):
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
                    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.")

1770
    parser_or_group.add_argument(
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
        "--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."
        ),
1789
1790
    )

1791

1792
def get_samples(args, tokenizer: TokenizerLike) -> list[SampleRequest]:
1793
1794
1795
    if not hasattr(args, "request_id_prefix"):
        args.request_id_prefix = ""

1796
    if args.dataset_name == "custom":
1797
1798
1799
        dataset = CustomDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1800
1801
1802
1803
        input_requests = dataset.sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.custom_output_len,
1804
            skip_chat_template=args.skip_chat_template,
1805
            request_id_prefix=args.request_id_prefix,
1806
            no_oversample=args.no_oversample,
1807
1808
        )

1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
    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,
        )

1822
    elif args.dataset_name == "sonnet":
1823
1824
1825
        dataset = SonnetDataset(
            dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle
        )
1826
        # For the "sonnet" dataset, formatting depends on the backend.
1827
        if args.backend == "openai-chat":
1828
1829
1830
1831
1832
1833
1834
            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,
1835
                request_id_prefix=args.request_id_prefix,
1836
                no_oversample=args.no_oversample,
1837
1838
1839
            )
        else:
            assert tokenizer.chat_template or tokenizer.default_chat_template, (
1840
1841
                "Tokenizer/model must have chat template for sonnet dataset."
            )
1842
1843
1844
1845
1846
1847
1848
            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,
1849
                request_id_prefix=args.request_id_prefix,
1850
                no_oversample=args.no_oversample,
1851
1852
1853
1854
1855
            )

    elif args.dataset_name == "hf":
        # all following datasets are implemented from the
        # HuggingFaceDataset base class
1856
        hf_kwargs = {}
1857
1858
1859
1860
        if (
            args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS
        ):
1861
            dataset_class = VisionArenaDataset
1862
            args.hf_split = args.hf_split if args.hf_split else "train"
1863
            args.hf_subset = None
1864
1865
1866
1867
1868
        elif (
            args.dataset_path in MMVUDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMVUDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMVUDataset
1869
            args.hf_split = args.hf_split if args.hf_split else "validation"
1870
            args.hf_subset = None
1871
1872
1873
1874
        elif (
            args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS
        ):
1875
            dataset_class = InstructCoderDataset
1876
            args.hf_split = args.hf_split if args.hf_split else "train"
1877
1878
1879
1880
        elif (
            args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MTBenchDataset.SUPPORTED_DATASET_PATHS
        ):
1881
            dataset_class = MTBenchDataset
1882
            args.hf_split = args.hf_split if args.hf_split else "train"
1883
1884
1885
1886
1887
        elif (
            args.dataset_path in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MultiModalConversationDataset
1888
1889
1890
1891
        elif (
            args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS
        ):
1892
            dataset_class = ConversationDataset
1893
1894
1895
1896
        elif (
            args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS
        ):
1897
            dataset_class = AIMODataset
1898
            args.hf_split = args.hf_split if args.hf_split else "train"
1899
        elif (
1900
            args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS  # noqa: E501
1901
1902
            or args.hf_name in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS
        ):
1903
            dataset_class = NextEditPredictionDataset
1904
            args.hf_split = args.hf_split if args.hf_split else "train"
1905
1906
1907
1908
        elif (
            args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS
        ):
1909
            dataset_class = ASRDataset
1910
1911
1912
1913
1914
            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,
            }
1915
1916
        elif args.dataset_path in BlazeditDataset.SUPPORTED_DATASET_PATHS:
            dataset_class = BlazeditDataset
1917
            args.hf_split = args.hf_split if args.hf_split else "train"
1918
1919
1920
1921
            hf_kwargs = {
                "min_distance": args.blazedit_min_distance,
                "max_distance": args.blazedit_max_distance,
            }
1922
1923
1924
1925
        elif (
            args.dataset_path in MLPerfDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MLPerfDataset.SUPPORTED_DATASET_PATHS
        ):
1926
            dataset_class = MLPerfDataset
1927
            args.hf_split = args.hf_split if args.hf_split else "train"
1928
1929
1930
1931
1932
        elif (
            args.dataset_path in MMStarDataset.SUPPORTED_DATASET_PATHS
            or args.hf_name in MMStarDataset.SUPPORTED_DATASET_PATHS
        ):
            dataset_class = MMStarDataset
1933
            args.hf_split = args.hf_split if args.hf_split else "val"
1934
            args.hf_subset = None
1935
        else:
1936
1937
1938
1939
1940
1941
1942
            supported_datasets = set(
                [
                    dataset_name
                    for cls in HuggingFaceDataset.__subclasses__()
                    for dataset_name in cls.SUPPORTED_DATASET_PATHS
                ]
            )
1943
1944
1945
1946
1947
            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 "
1948
1949
                "like to add support for additional dataset formats."
            )
1950

1951
1952
        if dataset_class.IS_MULTIMODAL and not (
            args.backend in ("openai-chat", "openai-audio")
1953
            or "embeddings-" in args.backend
1954
        ):
1955
1956
            # multi-modal benchmark is only available on OpenAI Chat
            # endpoint-type.
1957
1958
            raise ValueError(
                "Multi-modal content is only supported on 'openai-chat' and "
1959
1960
                "'openai-audio' backends."
            )
1961
1962
1963
1964
1965
        input_requests = dataset_class(
            dataset_path=args.dataset_path,
            dataset_subset=args.hf_subset,
            dataset_split=args.hf_split,
            random_seed=args.seed,
1966
            no_stream=args.no_stream,
1967
            hf_name=args.hf_name,
1968
            disable_shuffle=args.disable_shuffle,
1969
            trust_remote_code=args.trust_remote_code,
1970
1971
1972
1973
        ).sample(
            num_requests=args.num_prompts,
            tokenizer=tokenizer,
            output_len=args.hf_output_len,
1974
            enable_multimodal_chat=args.enable_multimodal_chat,
1975
            request_id_prefix=args.request_id_prefix,
1976
            no_oversample=args.no_oversample,
1977
            skip_chat_template=args.skip_chat_template,
1978
            **hf_kwargs,
1979
1980
1981
1982
1983
        )

    else:
        # For datasets that follow a similar structure, use a mapping.
        dataset_mapping = {
1984
            "spec_bench": lambda: SpecBench(
1985
1986
1987
                dataset_path=args.dataset_path,
                category=args.spec_bench_category,
                disable_shuffle=args.disable_shuffle,
1988
            ).sample(
1989
1990
1991
                num_requests=args.num_prompts,
                tokenizer=tokenizer,
                output_len=args.spec_bench_output_len,
1992
                enable_multimodal_chat=args.enable_multimodal_chat,
1993
                request_id_prefix=args.request_id_prefix,
1994
                no_oversample=args.no_oversample,
1995
            ),
1996
            "sharegpt": lambda: ShareGPTDataset(
1997
1998
1999
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
2000
2001
2002
2003
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                output_len=args.sharegpt_output_len,
2004
                enable_multimodal_chat=args.enable_multimodal_chat,
2005
                request_id_prefix=args.request_id_prefix,
2006
                no_oversample=args.no_oversample,
2007
2008
            ),
            "burstgpt": lambda: BurstGPTDataset(
2009
2010
2011
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
2012
2013
2014
2015
            ).sample(
                tokenizer=tokenizer,
                num_requests=args.num_prompts,
                request_id_prefix=args.request_id_prefix,
2016
                no_oversample=args.no_oversample,
2017
2018
            ),
            "random": lambda: RandomDataset(
2019
2020
2021
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
2022
            ).sample(
2023
2024
2025
2026
2027
2028
                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,
2029
                request_id_prefix=args.request_id_prefix,
2030
                batchsize=args.random_batch_size,
2031
                no_oversample=args.no_oversample,
2032
            ),
2033
            "random-mm": lambda: RandomMultiModalDataset(
2034
2035
2036
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
            ).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,
2048
                enable_multimodal_chat=args.enable_multimodal_chat,
2049
                request_id_prefix=args.request_id_prefix,
2050
                no_oversample=args.no_oversample,
2051
            ),
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
            "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,
            ),
2065
            "prefix_repetition": lambda: PrefixRepetitionRandomDataset(
2066
2067
2068
                random_seed=args.seed,
                dataset_path=args.dataset_path,
                disable_shuffle=args.disable_shuffle,
2069
2070
2071
2072
2073
2074
2075
            ).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,
2076
                request_id_prefix=args.request_id_prefix,
2077
                no_oversample=args.no_oversample,
2078
            ),
2079
2080
2081
        }

        try:
2082
            # Enforce endpoint compatibility for multimodal datasets.
2083
            if args.dataset_name == "random-mm" and args.backend not in ["openai-chat"]:
2084
2085
2086
2087
                raise ValueError(
                    "Multi-modal content (images) is only supported on "
                    "'openai-chat' backend."
                )
2088
2089
2090
2091
2092
2093
2094
            input_requests = dataset_mapping[args.dataset_name]()
        except KeyError as err:
            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err

    return input_requests


2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
# -----------------------------------------------------------------------------
# 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.,
    ```
2105
2106
2107
    {"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}
2108
    ```
2109
2110
    Note that 'output_tokens' column is optional and has to be provided only if
    'custom-output-len' argument is None or -1.
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
    """

    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"):
2130
            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143

            # 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(
2144
2145
                "Only JSONL format is supported for CustomDataset."
            )
2146
2147

        random.seed(self.random_seed)
2148
2149
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2150
2151
2152

    def sample(
        self,
2153
        tokenizer: TokenizerLike,
2154
        num_requests: int,
2155
2156
2157
        lora_path: str | None = None,
        max_loras: int | None = None,
        output_len: int | None = None,
2158
2159
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
2160
        request_id_prefix: str = "",
2161
        no_oversample: bool = False,
2162
2163
        **kwargs,
    ) -> list:
2164
2165
2166
2167
        # load all data if needed
        self.num_available_samples = len(self.data)
        if num_requests <= 0:
            num_requests = self.num_available_samples
2168
2169
2170
2171
2172
            logger.info(
                "num_requests is set to 0 or negative, "
                "so using all available samples: %d",
                num_requests,
            )
2173

2174
        sampled_requests = []
2175
        for i, item in enumerate(self.data):
2176
2177
2178
2179
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["prompt"]

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
            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,
2209
                    )
2210

2211
                prompt_len = len(tokenizer(prompt).input_ids)
2212
2213
2214
2215
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
2216
                    expected_output_len=new_output_len,
2217
                    request_id=request_id_prefix + str(i),
2218
2219
2220
2221
2222
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2223
2224
2225
2226

        return sampled_requests


2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
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
2304
2305
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


2306
2307
2308
2309
2310
2311
2312
2313
# -----------------------------------------------------------------------------
# Spec Bench Dataset Implementation
# -----------------------------------------------------------------------------


class SpecBench(CustomDataset):
    """
    Implements the SpecBench dataset: https://github.com/hemingkx/Spec-Bench
2314
    Download the dataset using:
2315
    wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl
2316
    """  # noqa: E501
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329

    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
2330
        jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
2331
2332
2333
2334
2335
2336
2337

        # 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
2338
            if (not self.category) or (self.category == row["category"]):
2339
2340
2341
2342
                prompt = row["turns"][0]
                self.data.append({"prompt": prompt})

        random.seed(self.random_seed)
2343
2344
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(self.data)
2345
2346
2347
2348

    def sample(self, **kwargs) -> list:
        # leverage CustomDataset sample
        return super().sample(**kwargs)
2349
2350


2351
2352
2353
2354
# -----------------------------------------------------------------------------
# Sonnet Dataset Implementation
# -----------------------------------------------------------------------------

2355

2356
2357
2358
@deprecated(
    "SonnetDataset is deprecated and will be removed in a future version.",
)
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
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,
2385
        tokenizer: TokenizerLike,
2386
2387
2388
2389
2390
        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,
2391
        request_id_prefix: str = "",
2392
        no_oversample: bool = False,
2393
2394
2395
2396
        **kwargs,
    ) -> list:
        # Calculate average token length for a poem line.
        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
2397
        avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines)
2398
2399
2400
2401

        # 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}]
2402
2403
2404
        base_fmt = tokenizer.apply_chat_template(
            base_msg, add_generation_prompt=True, tokenize=False
        )
2405
2406
2407
2408
        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 "
2409
2410
                f"({base_offset})."
            )
2411
2412
2413
2414
2415
2416
2417

        # 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 = []
2418
        ind = 0
2419
        while len(samples) < num_requests:
2420
2421
2422
            extra_lines = random.choices(
                self.data, k=num_input_lines - num_prefix_lines
            )
2423
2424
2425
            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
            msg = [{"role": "user", "content": prompt}]
            prompt_formatted = tokenizer.apply_chat_template(
2426
2427
                msg, add_generation_prompt=True, tokenize=False
            )
2428
2429
2430
2431
            prompt_len = len(tokenizer(prompt_formatted).input_ids)
            if prompt_len <= input_len:
                samples.append(
                    SampleRequest(
2432
                        prompt=prompt_formatted if return_prompt_formatted else prompt,
2433
2434
                        prompt_len=prompt_len,
                        expected_output_len=output_len,
2435
2436
2437
                        request_id=request_id_prefix + str(ind),
                    )
                )
2438
                ind += 1
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
        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()

2458
2459
2460
    def load_data(
        self,
    ):
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
        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):
2474
            data = self.data.sample(n=num_requests, random_state=self.random_seed)
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
        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,
2486
        tokenizer: TokenizerLike,
2487
        num_requests: int,
2488
2489
        max_loras: int | None = None,
        lora_path: str | None = None,
2490
        request_id_prefix: str = "",
2491
        no_oversample: bool = False,
2492
        lora_assignment: str = "random",
2493
2494
2495
2496
2497
2498
2499
        **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])
2500
2501
2502
2503
2504
            lora_req = self.get_lora_request(
                index=i,
                max_loras=max_loras,
                lora_path=lora_path,
                lora_assignment=lora_assignment,
2505
            )
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
            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,
2517
                    request_id=request_id_prefix + str(i),
2518
2519
                )
            )
2520
2521
2522
2523
2524
2525
2526
2527
2528
        return samples


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

2529
    SUPPORTED_DATASET_PATHS: set[str] | dict[str, Callable] = set()
2530
2531
2532
2533
2534

    def __init__(
        self,
        dataset_path: str,
        dataset_split: str,
2535
        no_stream: bool = False,
2536
2537
        dataset_subset: str | None = None,
        hf_name: str | None = None,
2538
        trust_remote_code: bool = False,
2539
2540
2541
2542
2543
2544
        **kwargs,
    ) -> None:
        super().__init__(dataset_path=dataset_path, **kwargs)

        self.dataset_split = dataset_split
        self.dataset_subset = dataset_subset
2545
        self.load_stream = not no_stream
2546
        self.hf_name = hf_name or dataset_path
2547
        self.trust_remote_code = trust_remote_code
2548
2549
2550
2551
2552
2553
2554
2555
        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,
2556
            streaming=self.load_stream,
2557
            trust_remote_code=self.trust_remote_code,
2558
        )
2559
2560
        if not getattr(self, "disable_shuffle", False):
            self.data = self.data.shuffle(seed=self.random_seed)
2561
2562
2563
2564
2565
2566
2567
2568


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


class ConversationDataset(HuggingFaceDataset):
2569
    """Dataset for text-only conversation data."""
2570

2571
    SUPPORTED_DATASET_PATHS = {
2572
        "Aeala/ShareGPT_Vicuna_unfiltered",
2573
    }
2574
2575
2576
2577
    IS_MULTIMODAL = False

    def sample(
        self,
2578
        tokenizer: TokenizerLike,
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
        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",
    }
2634
    IS_MULTIMODAL = True
2635

2636
2637
    def sample(
        self,
2638
        tokenizer: TokenizerLike,
2639
        num_requests: int,
2640
        output_len: int | None = None,
2641
2642
2643
2644
2645
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2646
        # Filter examples with at least 2 conversations
2647
        filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2)
2648
        sampled_requests = []
2649
        ind = 0
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
        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
2664
            if dynamic_output and not is_valid_sequence(prompt_len, completion_len):
2665
                continue
2666
            mm_content = process_image(item["image"]) if "image" in item else None
2667
2668
2669
2670
            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
2671
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2672
2673
2674
2675
2676
2677
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2678
                    request_id=request_id_prefix + str(ind),
2679
2680
                )
            )
2681
            ind += 1
2682
2683
2684
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
        return sampled_requests


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


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

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2700
2701
        "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"],
2702
    }
2703
    IS_MULTIMODAL = True
2704
2705
2706

    def sample(
        self,
2707
        tokenizer: TokenizerLike,
2708
        num_requests: int,
2709
        output_len: int | None = None,
2710
        enable_multimodal_chat: bool = False,
2711
        request_id_prefix: str = "",
2712
        no_oversample: bool = False,
2713
2714
        **kwargs,
    ) -> list:
2715
2716
2717
2718
        parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
        if parser_fn is None:
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")

2719
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2720

2721
        sampled_requests = []
2722
        for i, item in enumerate(self.data):
2723
2724
            if len(sampled_requests) >= num_requests:
                break
2725

2726
2727
            prompt = parser_fn(item)
            mm_content = process_image(item["images"][0])
2728
            prompt_len = len(tokenizer.encode(prompt))
2729
2730
2731
2732
            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
2733
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2734

2735
2736
2737
2738
2739
2740
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=mm_content,
2741
                    request_id=request_id_prefix + str(i),
2742
2743
                )
            )
2744

2745
2746
2747
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2748
2749
2750
        return sampled_requests


2751
2752
2753
2754
2755
2756
2757
2758
class MMVUDataset(HuggingFaceDataset):
    """
    MMVU Dataset.
    https://huggingface.co/datasets/yale-nlp/MMVU
    """

    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {
2759
2760
2761
        "yale-nlp/MMVU": lambda x: x["question"]
        + " "
        + (" ".join(f"{k}.{v}" for k, v in x["choices"].items())),
2762
2763
    }

2764
2765
2766
2767
2768
2769
2770
2771
    def __init__(self, **kwargs) -> None:
        super().__init__(**kwargs)

        self._remote_path_root = (
            f"https://huggingface.co/datasets/{self.hf_name}/resolve/main"
        )
        self._local_path_root = snapshot_download(self.hf_name, repo_type="dataset")

2772
2773
    def sample(
        self,
2774
        tokenizer: TokenizerLike,
2775
        num_requests: int,
2776
        output_len: int | None = None,
2777
2778
2779
2780
2781
        enable_multimodal_chat: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
2782
2783
2784
2785
        parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name)
        if parser_fn is None:
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")

2786
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2787

2788
2789
2790
2791
        sampled_requests = []
        for i, item in enumerate(self.data):
            if len(sampled_requests) >= num_requests:
                break
2792

2793
            prompt = parser_fn(item)
2794
2795
2796
            mm_content = process_video(
                item["video"].replace(self._remote_path_root, self._local_path_root)
            )
2797
            prompt_len = len(tokenizer.encode(prompt))
2798
2799
2800
2801
            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
2802
                prompt = self.apply_multimodal_chat_transformation(prompt, mm_content)
2803

2804
2805
2806
2807
2808
2809
2810
            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),
2811
2812
                )
            )
2813

2814
2815
2816
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2817
2818
2819
        return sampled_requests


2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
# -----------------------------------------------------------------------------
# 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",
    }

2840
2841
    def sample(
        self,
2842
        tokenizer: TokenizerLike,
2843
        num_requests: int,
2844
        output_len: int | None = None,
2845
2846
2847
2848
2849
        enable_multimodal_chat: bool = False,
        skip_chat_template: bool = False,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
2850
    ) -> list[SampleRequest]:
2851
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2852
        sampled_requests = []
2853
        for i, prompt in enumerate(self.sample_prompts(n=num_requests)):
2854
            # apply template
2855
2856
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2857
                    [{"role": "user", "content": prompt}],
2858
2859
2860
                    add_generation_prompt=True,
                    tokenize=False,
                )
2861

2862
2863
2864
2865
2866
2867
            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2868
                    request_id=request_id_prefix + str(i),
2869
2870
2871
2872
2873
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2874
2875
        return sampled_requests

2876
2877
2878
2879
2880
2881
2882
2883
    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

2884

2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
# -----------------------------------------------------------------------------
# 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,
2907
        tokenizer: TokenizerLike,
2908
        num_requests: int,
2909
        output_len: int | None = None,
2910
        enable_multimodal_chat: bool = False,
2911
        skip_chat_template: bool = False,
2912
        request_id_prefix: str = "",
2913
        no_oversample: bool = False,
2914
2915
        **kwargs,
    ) -> list:
2916
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2917
2918
        sampled_requests = []

2919
        for i, item in enumerate(self.data):
2920
2921
2922
2923
2924
            if len(sampled_requests) >= num_requests:
                break
            prompt = item["turns"][0]

            # apply template
2925
2926
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
2927
                    [{"role": "user", "content": prompt}],
2928
2929
2930
                    add_generation_prompt=True,
                    tokenize=False,
                )
2931
2932
2933
2934
2935
2936
2937

            prompt_len = len(tokenizer(prompt).input_ids)
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
2938
                    request_id=request_id_prefix + str(i),
2939
2940
2941
2942
2943
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
2944
2945
2946
        return sampled_requests


2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
# -----------------------------------------------------------------------------
# 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,
2973
        tokenizer: TokenizerLike,
2974
        num_requests: int,
2975
        output_len: int | None = None,
2976
        skip_chat_template: bool = False,
2977
        request_id_prefix: str = "",
2978
        no_oversample: bool = False,
2979
2980
2981
2982
        min_distance: float = 0.0,
        max_distance: float = 1.0,
        **kwargs,
    ) -> list:
2983
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
        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
2996
2997

            # template copied from
2998
            # https://github.com/ise-uiuc/blazedit/blob/7765137e656fd62de877422d2e4cf8de51228054/dataset/create_refined_dataset.py#L94-L105 # noqa: E501
2999
            prompt = f"""Given a code file, please apply the change requests and generate the new file.
3000
3001
3002
3003
3004
3005
3006
3007
3008

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

Change request:
{change_request}

3009
Please generate the new code file in the "New file" section below."""  # noqa: E501
3010
3011

            # apply template
3012
3013
            if not skip_chat_template:
                prompt = tokenizer.apply_chat_template(
3014
                    [{"role": "user", "content": prompt}],
3015
3016
3017
                    add_generation_prompt=True,
                    tokenize=False,
                )
3018
3019
3020
3021
3022
3023
3024
3025
3026

            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),
3027
3028
3029
3030
3031
                )
            )
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3032

3033
3034
3035
        return sampled_requests


3036
3037
3038
3039
3040
3041
3042
3043
3044
# -----------------------------------------------------------------------------
# AIMO Dataset Implementation
# -----------------------------------------------------------------------------


class AIMODataset(HuggingFaceDataset):
    """
    Dataset class for processing a AIMO dataset with reasoning questions.
    """
3045

3046
    SUPPORTED_DATASET_PATHS = {
3047
3048
3049
        "AI-MO/aimo-validation-aime",
        "AI-MO/NuminaMath-1.5",
        "AI-MO/NuminaMath-CoT",
3050
3051
    }

3052
3053
    def sample(
        self,
3054
        tokenizer: TokenizerLike,
3055
        num_requests: int,
3056
        output_len: int | None = None,
3057
3058
3059
3060
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ) -> list:
3061
        sampled_requests = []
3062
        ind = 0
3063
3064
3065
3066
3067
        dynamic_output = output_len is None

        for item in self.data:
            if len(sampled_requests) >= num_requests:
                break
3068
            prompt, completion = item["problem"], item["solution"]
3069
3070
3071
3072
3073
3074
3075

            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
3076
3077
3078
            if dynamic_output and not is_valid_sequence(
                prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000
            ):
3079
3080
3081
3082
3083
3084
3085
                continue
            sampled_requests.append(
                SampleRequest(
                    prompt=prompt,
                    prompt_len=prompt_len,
                    expected_output_len=output_len,
                    multi_modal_data=None,
3086
                    request_id=request_id_prefix + str(ind),
3087
3088
                )
            )
3089
            ind += 1
3090
3091
3092
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3093
        return sampled_requests
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113


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

3114
"""  # noqa: E501
3115
3116
3117


def _format_zeta_prompt(
3118
3119
    sample: dict, original_start_marker: str = "<|editable_region_start|>"
) -> dict:
3120
    """Format the zeta prompt for the Next Edit Prediction (NEP) dataset.
3121
3122
3123

    This function formats examples from the NEP dataset
    into prompts and expected outputs. It could be
3124
    further extended to support more NEP datasets.
3125

3126
    Args:
3127
        sample: The dataset sample containing events,
3128
            inputs, and outputs.
3129
3130
        original_start_marker: The marker indicating the
            start of the editable region. Defaults to
3131
            "<|editable_region_start|>".
3132

3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
    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,
    }

3162
3163
    def sample(
        self,
3164
        tokenizer: TokenizerLike,
3165
3166
3167
3168
3169
        num_requests: int,
        request_id_prefix: str = "",
        no_oversample: bool = False,
        **kwargs,
    ):
3170
        formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.hf_name)
3171
        if formatting_prompt_func is None:
3172
            raise ValueError(f"Unsupported dataset path: {self.hf_name}")
3173
        samples = []
3174
        for i, sample in enumerate(self.data):
3175
3176
3177
3178
3179
3180
            sample = formatting_prompt_func(sample)
            samples.append(
                SampleRequest(
                    prompt=sample["prompt"],
                    prompt_len=len(tokenizer(sample["prompt"]).input_ids),
                    expected_output_len=len(
3181
3182
                        tokenizer(sample["expected_output"]).input_ids
                    ),
3183
                    request_id=request_id_prefix + str(i),
3184
3185
                )
            )
3186
3187
            if len(samples) >= num_requests:
                break
3188
3189
3190
        self.maybe_oversample_requests(
            samples, num_requests, request_id_prefix, no_oversample
        )
3191
        return samples
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226


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

3227
    DEFAULT_OUTPUT_LEN = 1024
3228
3229
3230
3231
    IS_MULTIMODAL = True

    def sample(
        self,
3232
        tokenizer: TokenizerLike,
3233
        num_requests: int,
3234
        output_len: int | None = None,
3235
        request_id_prefix: str = "",
3236
        no_oversample: bool = False,
3237
3238
        **kwargs,
    ) -> list:
3239
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
Ekagra Ranjan's avatar
Ekagra Ranjan committed
3240
        if "openai" in getattr(tokenizer, "name_or_path", ""):
3241
3242
3243
            prompt = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>"
        else:
            prompt = ""
3244
3245
        prompt_len = len(tokenizer(prompt).input_ids)
        sampled_requests = []
3246
        ind = 0
3247
        skipped = 0
3248
3249
3250
        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 = []
3251
3252
3253
3254
3255
3256
        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)
3257
            if duration_s < asr_min_audio_len_sec or duration_s > asr_max_audio_len_sec:
3258
3259
3260
                skipped += 1
                continue

3261
            durations.append(duration_s)
3262
3263
3264
3265
3266
3267
3268
            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,
3269
                    request_id=request_id_prefix + str(ind),
3270
3271
                )
            )
3272
            ind += 1
3273
3274
3275
3276
3277
3278
3279
        if skipped:
            logger.warning(
                "%d samples discarded from dataset due to"
                " their length being greater than"
                " what Whisper supports.",
                skipped,
            )
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293

        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,
        )

3294
3295
3296
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3297
        return sampled_requests
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329


# -----------------------------------------------------------------------------
# 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,
3330
        tokenizer: TokenizerLike,
3331
        num_requests: int,
3332
        output_len: int | None = None,
3333
        request_id_prefix: str = "",
3334
        no_oversample: bool = False,
3335
3336
3337
3338
3339
        **kwargs,
    ) -> list[SampleRequest]:
        # Force dynamic output length based on reference completion.
        dynamic_output = output_len is None
        sampled_requests: list[SampleRequest] = []
3340
        ind = 0
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374

        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,
3375
                    request_id=request_id_prefix + str(ind),
3376
3377
                )
            )
3378
            ind += 1
3379

3380
3381
3382
        self.maybe_oversample_requests(
            sampled_requests, num_requests, request_id_prefix, no_oversample
        )
3383
        return sampled_requests
3384
3385
3386
3387
3388
3389
3390
3391


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


class PrefixRepetitionRandomDataset(BenchmarkDataset):
3392
    # Default values copied from benchmark_serving.py for the repeated prefix
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
    # 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,
3409
        tokenizer: TokenizerLike,
3410
3411
3412
3413
3414
        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,
3415
        request_id_prefix: str = "",
3416
        no_oversample: bool = False,
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
        **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
3431
            tokens = np.random.randint(0, vocab_size, size=target_length).tolist()
3432

3433
            _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len(  # noqa: E501
3434
3435
3436
3437
3438
3439
                tokenizer=tokenizer,
                token_sequence=tokens,
                target_token_len=target_length,
                add_special_tokens=False,
            )
            return adjusted_tokens, token_mismatch
3440
3441

        requests = []
3442
        token_mismatch_total = 0
3443
        for _ in range(num_prefixes):
3444
3445
            prefix_tokens, prefix_mismatch = _generate_exact_length_tokens(prefix_len)
            token_mismatch_total += prefix_mismatch
3446
3447

            for _ in range(prompts_per_prefix):
3448
                suffix_tokens, suffix_mismatch = _generate_exact_length_tokens(
3449
                    suffix_len
3450
                )
3451
                token_mismatch_total += suffix_mismatch
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
                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,
                    )
                )

3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
        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,
            )
3473
3474
        if not getattr(self, "disable_shuffle", False):
            random.shuffle(requests)
3475
        return requests
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487


# -----------------------------------------------------------------------------
# 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
    """
3488

3489
3490
3491
3492
3493
3494
    DEFAULT_OUTPUT_LEN = 128
    SUPPORTED_DATASET_PATHS = {"Lin-Chen/MMStar"}
    IS_MULTIMODAL = True

    def sample(
        self,
3495
        tokenizer: TokenizerLike,
3496
        num_requests: int,
3497
        output_len: int | None = None,
3498
3499
3500
3501
3502
3503
        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.
3504
        output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
        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