| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This module defines a framework for sampling benchmark requests from various |
| 4 | datasets. Each dataset subclass of BenchmarkDataset must implement sample |
| 5 | generation. Supported dataset types include: |
| 6 | - ShareGPT |
| 7 | - Random (synthetic) |
| 8 | - Sonnet |
| 9 | - BurstGPT |
| 10 | - HuggingFace |
| 11 | - VisionArena |
| 12 | """ |
| 13 | |
| 14 | import argparse |
| 15 | import ast |
| 16 | import io |
| 17 | import json |
| 18 | import logging |
| 19 | import math |
| 20 | import random |
| 21 | from abc import ABC, abstractmethod |
| 22 | from collections.abc import Callable, Iterator, Mapping |
| 23 | from contextlib import suppress |
| 24 | from dataclasses import dataclass, replace |
| 25 | from functools import cache |
| 26 | from io import BytesIO |
| 27 | from pathlib import Path |
| 28 | from tempfile import TemporaryDirectory |
| 29 | from typing import Any, cast |
| 30 | |
| 31 | import numpy as np |
| 32 | import pybase64 as base64 |
| 33 | from PIL import Image |
| 34 | from typing_extensions import deprecated |
| 35 | |
| 36 | from vllm.benchmarks.datasets.utils import ( |
| 37 | RangeRatio, |
| 38 | _resolve_range_ratios, |
| 39 | get_sampling_params, |
| 40 | ) |
| 41 | from vllm.inputs import MultiModalDataDict |
| 42 | from vllm.lora.request import LoRARequest |
| 43 | from vllm.lora.utils import get_adapter_absolute_path |
| 44 | from vllm.multimodal.audio import get_audio_duration |
| 45 | from vllm.multimodal.image import convert_image_mode |
| 46 | from vllm.multimodal.utils import encode_image_url, fetch_image |
| 47 | from vllm.tokenizers import TokenizerLike |
| 48 | from vllm.transformers_utils.repo_utils import hf_api |
| 49 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 50 | from vllm.utils.import_utils import PlaceholderModule |
| 51 | from vllm.utils.mistral import is_mistral_tokenizer |
| 52 | |
| 53 | try: |
| 54 | from datasets import load_dataset |
| 55 | except ImportError: |
| 56 | datasets = PlaceholderModule("datasets") |
| 57 | load_dataset = datasets.placeholder_attr("load_dataset") |
| 58 | |
| 59 | try: |
| 60 | import pandas as pd |
| 61 | except ImportError: |
| 62 | pd = PlaceholderModule("pandas") |
| 63 | |
| 64 | try: |
| 65 | import soundfile as sf |
| 66 | except ImportError: |
| 67 | sf = PlaceholderModule("soundfile") |
| 68 | |
| 69 | |
| 70 | logger = logging.getLogger(__name__) |
| 71 | |
| 72 | DEFAULT_NUM_PROMPTS = 1000 |
| 73 | |
| 74 | |
| 75 | @dataclass |
| 76 | class SampleRequest: |
| 77 | """Represents a single inference request for benchmarking.""" |
| 78 | |
| 79 | prompt: str | list[str] | list[int] | list[list[int]] | list[dict] |
| 80 | prompt_len: int |
| 81 | expected_output_len: int = 0 |
| 82 | multi_modal_data: MultiModalDataDict | dict | list[dict] | None = None |
| 83 | lora_request: LoRARequest | None = None |
| 84 | request_id: str | None = None |
| 85 | timestamp: float | None = None |
| 86 | # Pre-built chat messages. When set, the chat backend uses this list |
| 87 | # directly and skips constructing messages from `prompt` + multimodal |
| 88 | # content. Mutually exclusive with the `prompt`-based path. |
| 89 | chat_messages: list[dict[str, Any]] | None = None |
| 90 | # Per-request fields merged into the request body (e.g. tools, |
| 91 | # tool_choice, response_format). Shallow-merged with --extra-body at |
| 92 | # dispatch time; per-request keys win. |
| 93 | request_overrides: dict | None = None |
| 94 | |
| 95 | |
| 96 | # ----------------------------------------------------------------------------- |
| 97 | # Benchmark Dataset Base Class |
| 98 | # ----------------------------------------------------------------------------- |
| 99 | |
| 100 | |
| 101 | class BenchmarkDataset(ABC): |
| 102 | DEFAULT_SEED = 0 |
| 103 | IS_MULTIMODAL = False |
| 104 | |
| 105 | def __init__( |
| 106 | self, |
| 107 | dataset_path: str | None = None, |
| 108 | random_seed: int = DEFAULT_SEED, |
| 109 | disable_shuffle: bool = False, |
| 110 | **kwargs, |
| 111 | ) -> None: |
| 112 | """Initialize the BenchmarkDataset with an optional dataset path and random |
| 113 | seed. |
| 114 | |
| 115 | Args: |
| 116 | dataset_path (Optional[str]): Path to the dataset. If None, it |
| 117 | indicates that a default or random dataset might be used. |
| 118 | random_seed (int): Seed value for reproducible shuffling or |
| 119 | sampling. Defaults to DEFAULT_SEED. |
| 120 | disable_shuffle (bool): If True, keep the dataset in its original |
| 121 | order instead of shuffling it. |
| 122 | |
| 123 | """ |
| 124 | self.dataset_path = dataset_path |
| 125 | # Set the random seed, ensuring that a None value is replaced with the |
| 126 | # default seed. |
| 127 | self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED |
| 128 | self.disable_shuffle = disable_shuffle |
| 129 | self.data: Any | None = None |
| 130 | |
| 131 | def apply_multimodal_chat_transformation( |
| 132 | self, |
| 133 | prompt: str, |
| 134 | mm_content: MultiModalDataDict | dict | list[dict] | None = None, |
| 135 | ) -> list[dict]: |
| 136 | """Transform a prompt and optional multimodal content into a chat format. |
| 137 | This method is used for chat models that expect a specific conversation |
| 138 | format. |
| 139 | """ |
| 140 | content = [{"text": prompt, "type": "text"}] |
| 141 | if mm_content is not None: |
| 142 | if isinstance(mm_content, list): |
| 143 | content.extend(cast(list[dict[str, Any]], mm_content)) |
| 144 | elif isinstance(mm_content, dict): |
| 145 | content.append(mm_content) |
| 146 | else: |
| 147 | raise TypeError( |
| 148 | f"Could not process multimodal content of type: {type(mm_content)}" |
| 149 | ) |
| 150 | return [{"role": "user", "content": content}] |
| 151 | |
| 152 | def load_data(self) -> None: |
| 153 | """Load data from the dataset path into self.data. |
| 154 | |
| 155 | This method must be overridden by subclasses since the method to load |
| 156 | data will vary depending on the dataset format and source. |
| 157 | |
| 158 | Raises: |
| 159 | NotImplementedError: If a subclass does not implement this method. |
| 160 | |
| 161 | """ |
| 162 | # TODO (jenniferzhao): add support for downloading data |
| 163 | raise NotImplementedError("load_data must be implemented in subclasses.") |
| 164 | |
| 165 | @staticmethod |
| 166 | def get_random_lora_request( |
| 167 | max_loras: int | None = None, |
| 168 | lora_path: str | None = None, |
| 169 | ) -> LoRARequest | None: |
| 170 | """Optionally select a random LoRA request. |
| 171 | |
| 172 | This method is used when LoRA parameters are provided. It randomly |
| 173 | selects a LoRA based on max_loras. |
| 174 | |
| 175 | Args: |
| 176 | max_loras (Optional[int]): The maximum number of LoRAs available. |
| 177 | If `None`, LoRA is not used. |
| 178 | lora_path (Optional[str]): Path to the LoRA parameters on disk. |
| 179 | If `None`, LoRA is not used. |
| 180 | |
| 181 | Returns: |
| 182 | A new [`LoRARequest`][vllm.lora.request.LoRARequest] |
| 183 | (or `None` if not applicable). |
| 184 | |
| 185 | """ |
| 186 | if max_loras is None or lora_path is None: |
| 187 | return None |
| 188 | |
| 189 | # Generate a random LoRA ID in the range [1, max_loras]. |
| 190 | lora_id = random.randint(1, max_loras) |
| 191 | lora_request = LoRARequest( |
| 192 | lora_name=str(lora_id), |
| 193 | lora_int_id=lora_id, |
| 194 | lora_path=lora_path_on_disk(lora_path), |
| 195 | ) |
| 196 | return lora_request |
| 197 | |
| 198 | @staticmethod |
| 199 | def get_round_robin_lora_request( |
| 200 | index: int, |
| 201 | max_loras: int | None = None, |
| 202 | lora_path: str | None = None, |
| 203 | ) -> LoRARequest | None: |
| 204 | """Optionally select a LoRA request using deterministic round-robin. |
| 205 | |
| 206 | This method cycles through LoRA IDs in order based on the request |
| 207 | index, providing reproducible LoRA assignment. |
| 208 | |
| 209 | Args: |
| 210 | index (int): The request index used for round-robin selection. |
| 211 | max_loras (Optional[int]): The maximum number of LoRAs available. |
| 212 | If `None`, LoRA is not used. |
| 213 | lora_path (Optional[str]): Path to the LoRA parameters on disk. |
| 214 | If `None`, LoRA is not used. |
| 215 | |
| 216 | Returns: |
| 217 | A new [`LoRARequest`][vllm.lora.request.LoRARequest] |
| 218 | (or `None` if not applicable). |
| 219 | |
| 220 | """ |
| 221 | if max_loras is None or lora_path is None: |
| 222 | return None |
| 223 | |
| 224 | # Deterministic round-robin: cycle through [1, max_loras] |
| 225 | lora_id = index % max_loras + 1 |
| 226 | lora_request = LoRARequest( |
| 227 | lora_name=str(lora_id), |
| 228 | lora_int_id=lora_id, |
| 229 | lora_path=lora_path_on_disk(lora_path), |
| 230 | ) |
| 231 | return lora_request |
| 232 | |
| 233 | @staticmethod |
| 234 | def get_lora_request( |
| 235 | index: int, |
| 236 | max_loras: int | None = None, |
| 237 | lora_path: str | None = None, |
| 238 | lora_assignment: str = "random", |
| 239 | ) -> LoRARequest | None: |
| 240 | """Select a LoRA request using the specified assignment strategy. |
| 241 | |
| 242 | Args: |
| 243 | index (int): The request index (used for round-robin). |
| 244 | max_loras (Optional[int]): The maximum number of LoRAs available. |
| 245 | lora_path (Optional[str]): Path to the LoRA parameters on disk. |
| 246 | lora_assignment (str): Strategy for LoRA selection. |
| 247 | 'random' (default) or 'round-robin'. |
| 248 | |
| 249 | Returns: |
| 250 | A new [`LoRARequest`][vllm.lora.request.LoRARequest] |
| 251 | (or `None` if not applicable). |
| 252 | |
| 253 | """ |
| 254 | if lora_assignment == "round-robin": |
| 255 | return BenchmarkDataset.get_round_robin_lora_request( |
| 256 | index=index, max_loras=max_loras, lora_path=lora_path |
| 257 | ) |
| 258 | return BenchmarkDataset.get_random_lora_request( |
| 259 | max_loras=max_loras, lora_path=lora_path |
| 260 | ) |
| 261 | |
| 262 | @abstractmethod |
| 263 | def sample( |
| 264 | self, |
| 265 | tokenizer: TokenizerLike, |
| 266 | num_requests: int, |
| 267 | request_id_prefix: str = "", |
| 268 | no_oversample: bool = False, |
| 269 | **kwargs, |
| 270 | ) -> list[SampleRequest]: |
| 271 | """Abstract method to generate sample requests from the dataset. |
| 272 | |
| 273 | Subclasses must override this method to implement dataset-specific logic |
| 274 | for generating a list of SampleRequest objects. |
| 275 | |
| 276 | Args: |
| 277 | tokenizer (TokenizerLike): The tokenizer to be used |
| 278 | for processing the dataset's text. |
| 279 | num_requests (int): The number of sample requests to generate. |
| 280 | request_id_prefix (str): The prefix of request_id. |
| 281 | no_oversample (bool): If True, do not oversample the dataset to |
| 282 | reach num_requests. |
| 283 | |
| 284 | Returns: |
| 285 | list[SampleRequest]: A list of sample requests generated from the |
| 286 | dataset. |
| 287 | |
| 288 | """ |
| 289 | raise NotImplementedError("sample must be implemented in subclasses.") |
| 290 | |
| 291 | def maybe_oversample_requests( |
| 292 | self, |
| 293 | requests: list[SampleRequest], |
| 294 | num_requests: int, |
| 295 | request_id_prefix: str = "", |
| 296 | no_oversample: bool = False, |
| 297 | ) -> None: |
| 298 | """Oversamples the list of requests if its size is less than the desired |
| 299 | number. |
| 300 | |
| 301 | Args: |
| 302 | requests (List[SampleRequest]): The current list of sampled |
| 303 | requests. |
| 304 | num_requests (int): The target number of requests. |
| 305 | request_id_prefix (str): The prefix applied to generated request |
| 306 | identifiers. |
| 307 | no_oversample (bool): If True, return the requests unchanged |
| 308 | instead of oversampling up to num_requests. |
| 309 | |
| 310 | """ |
| 311 | if no_oversample: |
| 312 | logger.info("Skipping oversampling. Total samples: %d.", len(requests)) |
| 313 | return |
| 314 | |
| 315 | if len(requests) < num_requests: |
| 316 | random.seed(self.random_seed) |
| 317 | needed = num_requests - len(requests) |
| 318 | additional = [] |
| 319 | for i in range(needed): |
| 320 | req = replace( |
| 321 | random.choice(requests), |
| 322 | request_id=request_id_prefix + str(len(requests) + i), |
| 323 | ) |
| 324 | additional.append(req) |
| 325 | requests.extend(additional) |
| 326 | logger.info("Oversampled requests to reach %d total samples.", num_requests) |
| 327 | |
| 328 | ids = [req.request_id for req in requests] |
| 329 | if len(ids) != len(set(ids)): |
| 330 | raise ValueError( |
| 331 | "Duplicate request_id found in the sampled " |
| 332 | "requests. Please ensure that each request_id " |
| 333 | "is unique." |
| 334 | ) |
| 335 | |
| 336 | |
| 337 | # ----------------------------------------------------------------------------- |
| 338 | # Utility Functions and Global Caches |
| 339 | # ----------------------------------------------------------------------------- |
| 340 | |
| 341 | |
| 342 | def is_valid_sequence( |
| 343 | prompt_len: int, |
| 344 | output_len: int, |
| 345 | min_len: int = 4, |
| 346 | max_prompt_len: int = 1024, |
| 347 | max_total_len: int = 2048, |
| 348 | skip_min_output_len_check: bool = False, |
| 349 | ) -> bool: |
| 350 | """Validate a sequence based on prompt and output lengths. |
| 351 | |
| 352 | Default pruning criteria are copied from the original `sample_hf_requests` |
| 353 | and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as |
| 354 | from `sample_requests` in benchmark_throughput.py. |
| 355 | """ |
| 356 | # Check for invalid conditions |
| 357 | prompt_too_short = prompt_len < min_len |
| 358 | output_too_short = (not skip_min_output_len_check) and (output_len < min_len) |
| 359 | prompt_too_long = prompt_len > max_prompt_len |
| 360 | combined_too_long = (prompt_len + output_len) > max_total_len |
| 361 | |
| 362 | # Return True if none of the invalid conditions are met |
| 363 | return not ( |
| 364 | prompt_too_short or output_too_short or prompt_too_long or combined_too_long |
| 365 | ) |
| 366 | |
| 367 | |
| 368 | @cache |
| 369 | def lora_path_on_disk(lora_path: str) -> str: |
| 370 | return get_adapter_absolute_path(lora_path) |
| 371 | |
| 372 | |
| 373 | # Global cache for LoRA tokenizers. |
| 374 | lora_tokenizer_cache: dict[int, TokenizerLike] = {} |
| 375 | |
| 376 | |
| 377 | def process_image( |
| 378 | image: Any, |
| 379 | *, |
| 380 | ensure_client_side_data: bool = False, |
| 381 | ) -> Mapping[str, Any]: |
| 382 | """Process a single image input and return a multimedia content dictionary. |
| 383 | |
| 384 | Supports the following input types: |
| 385 | |
| 386 | 1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key |
| 387 | containing raw image data. - Loads the bytes as a PIL.Image.Image. |
| 388 | |
| 389 | 2. PIL.Image.Image input: - Converts the image to RGB. - Saves the image as |
| 390 | a JPEG in memory. - Encodes the JPEG data as a base64 string. - Returns |
| 391 | a dictionary with the image as a base64 data URL. |
| 392 | |
| 393 | 3. String input: - Treats the string as a URL, local file path, or base64 |
| 394 | encoded data. - If string starts with "data:image/", treats as base64. |
| 395 | - If string starts with "http://", "https://", or "file://", treats as URL. |
| 396 | - Otherwise treats as local file path and prepends "file://". |
| 397 | - If ensure_client_side_data is True, local and HTTP(S) image references |
| 398 | are loaded and encoded as base64 image data URLs. Existing data:image |
| 399 | URLs are kept unchanged. |
| 400 | - Returns a dictionary with the image URL or base64 data. |
| 401 | |
| 402 | Raises: |
| 403 | ValueError: If the input is not a supported type. |
| 404 | |
| 405 | """ |
| 406 | if isinstance(image, dict) and "bytes" in image: |
| 407 | image = Image.open(BytesIO(image["bytes"])) |
| 408 | if isinstance(image, Image.Image): |
| 409 | image = convert_image_mode(image, "RGB") |
| 410 | with io.BytesIO() as image_data: |
| 411 | image.save(image_data, format="JPEG") |
| 412 | image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8") |
| 413 | return { |
| 414 | "type": "image_url", |
| 415 | "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}, |
| 416 | } |
| 417 | |
| 418 | if isinstance(image, str): |
| 419 | image_url = ( |
| 420 | image |
| 421 | if image.startswith(("http://", "https://", "file://", "data:image/")) |
| 422 | else f"file://{image}" |
| 423 | ) |
| 424 | |
| 425 | if ensure_client_side_data and not image_url.startswith("data:image/"): |
| 426 | try: |
| 427 | fetched_image = fetch_image(image_url) |
| 428 | image_url = encode_image_url(fetched_image) |
| 429 | except Exception as e: |
| 430 | raise ValueError(f"Invalid image URL: {image_url}") from e |
| 431 | return {"type": "image_url", "image_url": {"url": image_url}} |
| 432 | |
| 433 | raise ValueError( |
| 434 | f"Invalid image input {image}. Must be a PIL.Image.Image, " |
| 435 | "str (URL, file path, or base64 data URL), or dictionary with raw image bytes." |
| 436 | ) |
| 437 | |
| 438 | |
| 439 | def process_video(video: Any) -> Mapping[str, Any]: |
| 440 | """Process a single video input and return a multimedia content dictionary. |
| 441 | |
| 442 | Supports the following input types: |
| 443 | |
| 444 | 1. Dictionary with raw video bytes: - Expects a dict with a 'bytes' key |
| 445 | containing raw video data. |
| 446 | |
| 447 | 2. String input: - Treats the string as a URL or local file path. - |
| 448 | Prepends "file://" if the string doesn't start with "http://" or |
| 449 | "file://". - Returns a dictionary with the image URL. |
| 450 | |
| 451 | Raises: |
| 452 | ValueError: If the input is not a supported type. |
| 453 | |
| 454 | """ |
| 455 | if isinstance(video, dict) and "bytes" in video: |
| 456 | video_bytes = video["bytes"] |
| 457 | video_base64 = base64.b64encode(video_bytes).decode("utf-8") |
| 458 | return { |
| 459 | "type": "video_url", |
| 460 | "video_url": {"url": f"data:video/mp4;base64,{video_base64}"}, |
| 461 | } |
| 462 | |
| 463 | if isinstance(video, str): |
| 464 | video_url = ( |
| 465 | video |
| 466 | if video.startswith(("http://", "https://", "file://")) |
| 467 | else f"file://{video}" |
| 468 | ) |
| 469 | return {"type": "video_url", "video_url": {"url": video_url}} |
| 470 | |
| 471 | raise ValueError( |
| 472 | 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 |
| 473 | ) |
| 474 | |
| 475 | |
| 476 | def process_audio(audio: Any) -> tuple: |
| 477 | """Process a single audio input and return a (array, sample_rate) tuple. |
| 478 | |
| 479 | Supports: |
| 480 | 1. String: treated as a file path, loaded with soundfile. |
| 481 | 2. Dict with 'array' and 'sampling_rate' keys: HuggingFace audio format. |
| 482 | 3. Tuple (array, sr): passed through directly. |
| 483 | """ |
| 484 | if isinstance(audio, str): |
| 485 | return sf.read(audio) |
| 486 | if isinstance(audio, dict) and "array" in audio and "sampling_rate" in audio: |
| 487 | return audio["array"], audio["sampling_rate"] |
| 488 | if isinstance(audio, tuple) and len(audio) == 2: |
| 489 | return audio |
| 490 | raise ValueError( |
| 491 | f"Invalid audio input {audio}. Must be a file path string, " |
| 492 | "a dict with 'array' and 'sampling_rate', or a (array, sr) tuple." |
| 493 | ) |
| 494 | |
| 495 | |
| 496 | def gen_prompt_decode_to_target_len( |
| 497 | tokenizer: TokenizerLike, |
| 498 | token_sequence: list[int], |
| 499 | target_token_len: int, |
| 500 | max_retry: int = 10, |
| 501 | add_special_tokens: bool = False, |
| 502 | rng: np.random.Generator | None = None, |
| 503 | ) -> tuple[str, list[int], int]: |
| 504 | """Ensure decoded-then-encoded prompt length matches the target token length. |
| 505 | |
| 506 | This function decodes an initial token sequence to text and re-encodes it |
| 507 | , iteratively adjusting the token sequence length to match a target. |
| 508 | This is necessary because some tokenizers do not guarantee a 1:1 mapping |
| 509 | between consecutive tokens and the decoded-then-encoded sequence length. |
| 510 | For example, for GPT2Tokenizer: |
| 511 | [6880, 6881] -> ['Ġcalls', 'here'] -> |
| 512 | [1650, 939, 486] -> ['Ġcall', 'sh', 'ere'] |
| 513 | |
| 514 | Returns a tuple of the final prompt string, the adjusted token sequence, |
| 515 | and the token mismatch (final_len - target_token_len) if the retry budget |
| 516 | is exhausted. |
| 517 | """ |
| 518 | remain_num_try = max_retry |
| 519 | token_mismatch = 0 |
| 520 | while True: |
| 521 | prompt = tokenizer.decode(token_sequence) |
| 522 | token_sequence = tokenizer.encode(prompt, add_special_tokens=add_special_tokens) |
| 523 | if remain_num_try <= 0: |
| 524 | if len(token_sequence) != target_token_len: |
| 525 | token_mismatch = len(token_sequence) - target_token_len |
| 526 | break |
| 527 | |
| 528 | if len(token_sequence) == target_token_len: |
| 529 | break |
| 530 | elif len(token_sequence) < target_token_len: |
| 531 | if rng is not None: |
| 532 | extra_tokens = rng.integers( |
| 533 | 0, |
| 534 | tokenizer.vocab_size, |
| 535 | size=target_token_len - len(token_sequence), |
| 536 | ).tolist() |
| 537 | else: |
| 538 | extra_tokens = np.random.randint( |
| 539 | 0, |
| 540 | tokenizer.vocab_size, |
| 541 | size=target_token_len - len(token_sequence), |
| 542 | ).tolist() |
| 543 | token_sequence.extend(extra_tokens) |
| 544 | elif len(token_sequence) > target_token_len: |
| 545 | token_sequence = token_sequence[:target_token_len] |
| 546 | |
| 547 | remain_num_try -= 1 |
| 548 | |
| 549 | return prompt, token_sequence, token_mismatch |
| 550 | |
| 551 | |
| 552 | # ----------------------------------------------------------------------------- |
| 553 | # Random Dataset Implementation (Synthetic Data) |
| 554 | # ----------------------------------------------------------------------------- |
| 555 | |
| 556 | |
| 557 | class RandomDataset(BenchmarkDataset): |
| 558 | """Synthetic text-only dataset for serving/throughput benchmarks. |
| 559 | |
| 560 | Strategy: |
| 561 | - Sample input/output token lengths per request from integer-uniform ranges |
| 562 | around configured means (controlled by range_ratio). |
| 563 | - Prepend a fixed random prefix of length prefix_len. |
| 564 | - Generate the remaining tokens as a reproducible sequence: |
| 565 | (offset + index + arange(input_len)) % vocab_size. |
| 566 | - Decode then re-encode/truncate to ensure prompt token counts match. |
| 567 | - Uses numpy.default_rng seeded with random_seed for reproducible sampling. |
| 568 | """ |
| 569 | |
| 570 | # Default values copied from benchmark_serving.py for the random dataset. |
| 571 | DEFAULT_PREFIX_LEN = 0 |
| 572 | DEFAULT_RANGE_RATIO = 0.0 |
| 573 | DEFAULT_INPUT_LEN = 1024 |
| 574 | DEFAULT_OUTPUT_LEN = 128 |
| 575 | |
| 576 | def __init__(self, **kwargs) -> None: |
| 577 | super().__init__(**kwargs) |
| 578 | # Use numpy's default_rng for deterministic sampling |
| 579 | # Do not use random.seed() or np.random.seed() elsewhere in this class. |
| 580 | # This ensures that the RNG is isolated from global RNG state. |
| 581 | self._rng = np.random.default_rng(self.random_seed) |
| 582 | |
| 583 | def sample( |
| 584 | self, |
| 585 | tokenizer: TokenizerLike, |
| 586 | num_requests: int, |
| 587 | request_id_prefix: str = "", |
| 588 | no_oversample: bool = False, |
| 589 | prefix_len: int = DEFAULT_PREFIX_LEN, |
| 590 | range_ratio: RangeRatio = DEFAULT_RANGE_RATIO, |
| 591 | input_len: int = DEFAULT_INPUT_LEN, |
| 592 | output_len: int = DEFAULT_OUTPUT_LEN, |
| 593 | batchsize: int = 1, |
| 594 | max_loras: int | None = None, |
| 595 | lora_path: str | None = None, |
| 596 | lora_assignment: str = "random", |
| 597 | **kwargs, |
| 598 | ) -> list[SampleRequest]: |
| 599 | resolved_input_rr, _ = _resolve_range_ratios(range_ratio) |
| 600 | |
| 601 | num_special = int(tokenizer.num_special_tokens_to_add()) |
| 602 | real_input_len = max(0, int(input_len) - num_special) |
| 603 | min_sampled_input = math.floor( |
| 604 | real_input_len * (1.0 - float(resolved_input_rr)) |
| 605 | ) |
| 606 | min_total_input = int(prefix_len) + min_sampled_input |
| 607 | if min_total_input < 1: |
| 608 | raise ValueError( |
| 609 | "--random-input-len is too small: with tokenizer special " |
| 610 | f"tokens {num_special} and " |
| 611 | f"input range ratio {resolved_input_rr}, " |
| 612 | "the minimum possible total input tokens (prefix + sampled) is " |
| 613 | f"{min_total_input}. Increase --random-input-len and/or " |
| 614 | "--random-prefix-len, or decrease the input range ratio " |
| 615 | "so that prefix_len + floor(max(0, random_input_len - " |
| 616 | "num_special)) * (1 - input_range_ratio) >= 1." |
| 617 | ) |
| 618 | |
| 619 | input_lens, output_lens, offsets = get_sampling_params( |
| 620 | self._rng, |
| 621 | num_requests, |
| 622 | range_ratio, |
| 623 | input_len, |
| 624 | output_len, |
| 625 | tokenizer, |
| 626 | ) |
| 627 | |
| 628 | vocab_size = tokenizer.vocab_size |
| 629 | prohibited_tokens = tokenizer.all_special_ids |
| 630 | all_tokens = np.arange(vocab_size) |
| 631 | allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens))) |
| 632 | |
| 633 | # Generate prefix once |
| 634 | prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len) |
| 635 | |
| 636 | requests: list[SampleRequest] = [] |
| 637 | token_mismatch_total = 0 |
| 638 | for i in range(num_requests): |
| 639 | prompt, total_input_len, token_mismatch = self.generate_token_sequence( # noqa: E501 |
| 640 | tokenizer=tokenizer, |
| 641 | prefix_token_ids=prefix_token_ids, |
| 642 | prefix_len=prefix_len, |
| 643 | vocab_size=vocab_size, |
| 644 | input_len=int(input_lens[i]), |
| 645 | offset=int(offsets[i]), |
| 646 | index=i, |
| 647 | allowed_tokens=allowed_tokens, |
| 648 | ) |
| 649 | token_mismatch_total += token_mismatch |
| 650 | lora_req = self.get_lora_request( |
| 651 | index=i, |
| 652 | max_loras=max_loras, |
| 653 | lora_path=lora_path, |
| 654 | lora_assignment=lora_assignment, |
| 655 | ) |
| 656 | requests.append( |
| 657 | SampleRequest( |
| 658 | prompt=prompt, |
| 659 | prompt_len=total_input_len, |
| 660 | expected_output_len=int(output_lens[i]), |
| 661 | lora_request=lora_req, |
| 662 | request_id=request_id_prefix + str(i), |
| 663 | ) |
| 664 | ) |
| 665 | # only used for embeddings benchmark. |
| 666 | if batchsize > 1: |
| 667 | batch_requests: list[SampleRequest] = [] |
| 668 | # Create batched requests |
| 669 | for i in range(0, num_requests, batchsize): |
| 670 | batch = requests[i : i + batchsize] |
| 671 | batch_requests.append( |
| 672 | SampleRequest( |
| 673 | prompt=[req.prompt for req in batch], # type: ignore[arg-type] |
| 674 | prompt_len=sum(req.prompt_len for req in batch), |
| 675 | request_id=request_id_prefix + str(i // batchsize), |
| 676 | ) |
| 677 | ) |
| 678 | requests = batch_requests |
| 679 | |
| 680 | if token_mismatch_total != 0: |
| 681 | sign = "more" if token_mismatch_total > 0 else "fewer" |
| 682 | logger.warning( |
| 683 | "Across all generated prompts, there were %d %s tokens " |
| 684 | "than expected after decoding and re-encoding. This is " |
| 685 | "expected due to the imperfect nature of the sampling " |
| 686 | "procedure.", |
| 687 | abs(token_mismatch_total), |
| 688 | sign, |
| 689 | ) |
| 690 | |
| 691 | return requests |
| 692 | |
| 693 | def get_prefix( |
| 694 | self, |
| 695 | tokenizer: TokenizerLike, |
| 696 | allowed_tokens: np.ndarray, |
| 697 | prefix_len: int, |
| 698 | ) -> list[int]: |
| 699 | """Get the prefix for the dataset.""" |
| 700 | if prefix_len <= 0: |
| 701 | return [] |
| 702 | |
| 703 | prefix_tokens = allowed_tokens[ |
| 704 | self._rng.integers(0, len(allowed_tokens), size=prefix_len) |
| 705 | ].tolist() |
| 706 | _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len( |
| 707 | tokenizer=tokenizer, |
| 708 | token_sequence=prefix_tokens, |
| 709 | target_token_len=prefix_len, |
| 710 | add_special_tokens=False, |
| 711 | rng=self._rng, |
| 712 | ) |
| 713 | if token_mismatch != 0: |
| 714 | sign = "more" if token_mismatch > 0 else "fewer" |
| 715 | logger.warning( |
| 716 | "Prefix tokenization produced %d %s tokens than expected " |
| 717 | "after decoding and re-encoding. This is expected due to " |
| 718 | "the imperfect nature of the sampling procedure", |
| 719 | abs(token_mismatch), |
| 720 | sign, |
| 721 | ) |
| 722 | return adjusted_tokens |
| 723 | |
| 724 | def generate_token_sequence( |
| 725 | self, |
| 726 | *, |
| 727 | tokenizer: TokenizerLike, |
| 728 | prefix_token_ids: list[int], |
| 729 | prefix_len: int, |
| 730 | vocab_size: int, |
| 731 | input_len: int, |
| 732 | offset: int, |
| 733 | index: int, |
| 734 | allowed_tokens: np.ndarray, |
| 735 | ) -> tuple[str, int, int]: |
| 736 | """Returns (prompt, total_input_len). |
| 737 | |
| 738 | NOTE: After decoding the prompt we have to encode and decode it again. |
| 739 | This is done because in some cases N consecutive tokens |
| 740 | give a string tokenized into != N number of tokens. |
| 741 | For example for GPT2Tokenizer: |
| 742 | [6880, 6881] -> ['Ġcalls', 'here'] -> |
| 743 | [1650, 939, 486] -> ['Ġcall', 'sh', 'ere'] |
| 744 | To avoid uncontrolled change of the prompt length, |
| 745 | the encoded sequence is truncated before being decoded again. |
| 746 | """ |
| 747 | # Build the inner sequence by sampling |
| 748 | # sequentially from the allowed tokens |
| 749 | inner_seq = allowed_tokens[ |
| 750 | (offset + index + np.arange(input_len)) % len(allowed_tokens) |
| 751 | ].tolist() |
| 752 | token_sequence = prefix_token_ids + inner_seq |
| 753 | |
| 754 | # Decode, then re-encode and truncate to preserve token count invariants |
| 755 | total_input_len = prefix_len + int(input_len) |
| 756 | prompt, adjusted_token_sequence, token_mismatch = ( |
| 757 | gen_prompt_decode_to_target_len( |
| 758 | tokenizer=tokenizer, |
| 759 | token_sequence=token_sequence, |
| 760 | target_token_len=total_input_len, |
| 761 | add_special_tokens=False, |
| 762 | rng=self._rng, |
| 763 | ) |
| 764 | ) |
| 765 | total_input_len = len(adjusted_token_sequence) |
| 766 | return prompt, total_input_len, token_mismatch |
| 767 | |
| 768 | |
| 769 | # ----------------------------------------------------------------------------- |
| 770 | # Random Dataset Implementation (Synthetic Data) |
| 771 | # ----------------------------------------------------------------------------- |
| 772 | |
| 773 | |
| 774 | class RandomDatasetForReranking(RandomDataset): |
| 775 | """Random dataset specialized for the needs of scoring: |
| 776 | - Batches of inputs |
| 777 | - Inputs composed of pairs |
| 778 | """ |
| 779 | |
| 780 | def __init__(self, **kwargs) -> None: |
| 781 | super().__init__(**kwargs) |
| 782 | |
| 783 | def sample( |
| 784 | self, |
| 785 | tokenizer: TokenizerLike, |
| 786 | num_requests: int, |
| 787 | request_id_prefix: str = "", |
| 788 | no_oversample: bool = False, |
| 789 | prefix_len: int = RandomDataset.DEFAULT_PREFIX_LEN, |
| 790 | range_ratio: RangeRatio = RandomDataset.DEFAULT_RANGE_RATIO, |
| 791 | input_len: int = RandomDataset.DEFAULT_INPUT_LEN, |
| 792 | output_len: int = RandomDataset.DEFAULT_OUTPUT_LEN, |
| 793 | batchsize: int = 1, |
| 794 | max_loras: int | None = None, |
| 795 | lora_path: str | None = None, |
| 796 | lora_assignment: str = "random", |
| 797 | is_reranker: bool = True, |
| 798 | **kwargs, |
| 799 | ) -> list[SampleRequest]: |
| 800 | n_sep_tokens = int(is_reranker) |
| 801 | |
| 802 | query_len_param = (input_len // 2) - n_sep_tokens if is_reranker else input_len |
| 803 | |
| 804 | query_lens, _, query_offsets = get_sampling_params( |
| 805 | self._rng, |
| 806 | 1, |
| 807 | range_ratio, |
| 808 | query_len_param, |
| 809 | 0, |
| 810 | tokenizer, |
| 811 | ) |
| 812 | |
| 813 | query_len = int(query_lens[0]) |
| 814 | |
| 815 | if not is_reranker: |
| 816 | assert num_requests > 1 and batchsize > 1 |
| 817 | num_requests -= 1 |
| 818 | batchsize -= 1 |
| 819 | doc_len_param = input_len |
| 820 | else: |
| 821 | doc_len_param = input_len - query_len - n_sep_tokens |
| 822 | |
| 823 | doc_lens, _, doc_offsets = get_sampling_params( |
| 824 | self._rng, |
| 825 | num_requests, |
| 826 | range_ratio, |
| 827 | doc_len_param, |
| 828 | 0, |
| 829 | tokenizer, |
| 830 | ) |
| 831 | |
| 832 | vocab_size = tokenizer.vocab_size |
| 833 | prohibited_tokens = tokenizer.all_special_ids |
| 834 | all_tokens = np.arange(vocab_size) |
| 835 | allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens))) |
| 836 | |
| 837 | query_prompt, query_input_len, token_mismatch_total = ( |
| 838 | self.generate_token_sequence( |
| 839 | tokenizer=tokenizer, |
| 840 | prefix_token_ids=[], |
| 841 | prefix_len=0, |
| 842 | vocab_size=vocab_size, |
| 843 | input_len=query_len, |
| 844 | offset=int(query_offsets[0]), |
| 845 | index=0, |
| 846 | allowed_tokens=allowed_tokens, |
| 847 | ) |
| 848 | ) |
| 849 | |
| 850 | requests = [] |
| 851 | for i in range(num_requests): |
| 852 | prompt, total_input_len, token_mismatch = self.generate_token_sequence( # noqa: E501 |
| 853 | tokenizer=tokenizer, |
| 854 | prefix_token_ids=[], |
| 855 | prefix_len=0, |
| 856 | vocab_size=vocab_size, |
| 857 | input_len=int(doc_lens[i]), |
| 858 | offset=int(doc_offsets[i]), |
| 859 | index=i + 1, |
| 860 | allowed_tokens=allowed_tokens, |
| 861 | ) |
| 862 | token_mismatch_total += token_mismatch |
| 863 | requests.append((prompt, total_input_len)) |
| 864 | |
| 865 | batch_requests = [] |
| 866 | # Create batched requests |
| 867 | for i in range(0, num_requests, batchsize): |
| 868 | batch = requests[i : i + batchsize] |
| 869 | query_contrib = ( |
| 870 | (query_input_len + n_sep_tokens) * len(batch) |
| 871 | if is_reranker |
| 872 | else query_input_len |
| 873 | ) |
| 874 | batch_requests.append( |
| 875 | SampleRequest( |
| 876 | prompt=[query_prompt] + [req[0] for req in batch], |
| 877 | prompt_len=query_contrib + sum(req[1] for req in batch), |
| 878 | expected_output_len=0, |
| 879 | request_id=request_id_prefix + str(i // batchsize), |
| 880 | ) |
| 881 | ) |
| 882 | |
| 883 | if token_mismatch_total != 0: |
| 884 | logger.warning( |
| 885 | "Across all generated prompts, there were %d %s tokens " |
| 886 | "than expected after decoding and re-encoding. This is " |
| 887 | "expected due to the imperfect nature of the sampling " |
| 888 | "procedure.", |
| 889 | abs(token_mismatch_total), |
| 890 | "more" if token_mismatch_total > 0 else "fewer", |
| 891 | ) |
| 892 | |
| 893 | return batch_requests |
| 894 | |
| 895 | |
| 896 | # ----------------------------------------------------------------------------- |
| 897 | # MultiModalDataset Implementation |
| 898 | # ----------------------------------------------------------------------------- |
| 899 | |
| 900 | |
| 901 | class RandomMultiModalDataset(RandomDataset): |
| 902 | """Synthetic multimodal dataset (text + images) that extends RandomDataset. |
| 903 | |
| 904 | Status: |
| 905 | - Images: supported via synthetic RGB data. |
| 906 | - Video: supported via synthetic RGB data. |
| 907 | - Audio: not yet supported. |
| 908 | |
| 909 | Sampling overview: |
| 910 | 1) Number of items per request is sampled uniformly from the integer range |
| 911 | [floor(n·(1−r)), ceil(n·(1+r))], where n is the base count and r is |
| 912 | `num_mm_items_range_ratio` in [0, 1]. r=0 keeps it fixed; r=1 allows 0. |
| 913 | The maximum is further clamped to the sum of per-modality limits. |
| 914 | 2) Each item’s modality and shape is sampled from `bucket_config`, a dict |
| 915 | mapping (height, width, num_frames) → probability. We treat |
| 916 | `num_frames`=1 as image and `num_frames` > 1 as video. |
| 917 | Entries with zero probability are removed and the rest are renormalized |
| 918 | to sum to 1. |
| 919 | 3) Per-modality hard caps are enforced via `limit_mm_per_prompt`. |
| 920 | When a modality reaches its cap, all of its buckets are excluded and the |
| 921 | remaining probabilities are renormalized. |
| 922 | |
| 923 | Example bucket configuration: |
| 924 | {(256, 256, 1): 0.5, (720, 1280, 1): 0.4, (720, 1280, 16): 0.1} |
| 925 | - Two image buckets (`num_frames`=1) and one video bucket |
| 926 | (`num_frames`=16). |
| 927 | OBS.: Only image sampling is supported for now. |
| 928 | """ |
| 929 | |
| 930 | IS_MULTIMODAL = True |
| 931 | DEFAULT_LIMIT_MM_PER_PROMPT = {"image": 255, "video": 1} |
| 932 | |
| 933 | DEFAULT_BASE_ITEMS_PER_REQUEST = 1 |
| 934 | DEFAULT_NUM_MM_ITEMS_RANGE_RATIO = 0.0 |
| 935 | DEFAULT_MM_ITEM_BUCKET_CONFIG = { |
| 936 | (256, 256, 1): 0.5, |
| 937 | (720, 1280, 1): 0.5, |
| 938 | (720, 1280, 16): 0.0, |
| 939 | } |
| 940 | DEFAULT_ENABLE_MULTIMODAL_CHAT = False |
| 941 | |
| 942 | def __init__(self, **kwargs) -> None: |
| 943 | super().__init__(**kwargs) |
| 944 | |
| 945 | def generate_synthetic_image(self, width: int, height: int) -> Image.Image: |
| 946 | """Generate synthetic PIL image with random RGB values. |
| 947 | |
| 948 | NOTE: iid pixel sampling results in worst-case compression |
| 949 | (good for stressing I/O), but very unlike real photos. |
| 950 | We could consider a “low-freq” mode (e.g., noise blur) |
| 951 | to emulate network realism instead of max stress. |
| 952 | """ |
| 953 | random_pixels = self._rng.integers( |
| 954 | 0, |
| 955 | 256, |
| 956 | (height, width, 3), |
| 957 | dtype=np.uint8, |
| 958 | ) |
| 959 | return Image.fromarray(random_pixels) |
| 960 | |
| 961 | def generate_synthetic_video( |
| 962 | self, width: int, height: int, num_frames: int |
| 963 | ) -> dict: |
| 964 | """Generate synthetic video with random values. |
| 965 | |
| 966 | Creates a video with random pixel values, encodes it to MP4 format, |
| 967 | and returns the content as bytes. |
| 968 | """ |
| 969 | import cv2 |
| 970 | |
| 971 | random_pixels = self._rng.integers( |
| 972 | 0, |
| 973 | 256, |
| 974 | (num_frames, height, width, 3), |
| 975 | dtype=np.uint8, |
| 976 | ) |
| 977 | |
| 978 | fourcc = cv2.VideoWriter_fourcc(*"mp4v") |
| 979 | fps = 30 # frames per second |
| 980 | |
| 981 | with TemporaryDirectory() as temp_dir: |
| 982 | temp_path = Path(temp_dir) / "video.mp4" |
| 983 | |
| 984 | video_writer = cv2.VideoWriter( |
| 985 | str(temp_path), fourcc=fourcc, fps=fps, frameSize=(width, height) |
| 986 | ) |
| 987 | |
| 988 | try: |
| 989 | if not video_writer.isOpened(): |
| 990 | raise RuntimeError("Failed to create video writer") |
| 991 | |
| 992 | for frame in random_pixels: |
| 993 | video_writer.write(frame) |
| 994 | finally: |
| 995 | video_writer.release() |
| 996 | |
| 997 | return {"bytes": temp_path.read_bytes()} |
| 998 | |
| 999 | def map_config_to_modality(self, config: tuple[int, int, int]) -> str: |
| 1000 | """Map the configuration to the modality.""" |
| 1001 | if config[-1] == 1: |
| 1002 | return "image" |
| 1003 | elif config[-1] > 1: |
| 1004 | return "video" |
| 1005 | else: |
| 1006 | raise ValueError(f"Invalid multimodal item configuration: {config}") |
| 1007 | |
| 1008 | def normalize_bucket_config( |
| 1009 | self, bucket_config: dict[tuple[int, int, int], float] |
| 1010 | ) -> dict[tuple[int, int, int], float]: |
| 1011 | """Remove zero probability entries |
| 1012 | and normalize the bucket config to sum to 1. |
| 1013 | """ |
| 1014 | # Raise error if value is negative |
| 1015 | if any(v < 0 for v in bucket_config.values()): |
| 1016 | raise ValueError("Bucket config values must be non-negative.") |
| 1017 | # Remove zero probability entries |
| 1018 | bucket_config = {k: v for k, v in bucket_config.items() if v > 0} |
| 1019 | # if bucket config is empty, raise error |
| 1020 | if not bucket_config: |
| 1021 | raise ValueError( |
| 1022 | "Got invalid bucket config. Bucket config values must be non-zero." |
| 1023 | ) |
| 1024 | # Normalize the remaining bucket config to sum to 1 |
| 1025 | total = sum(bucket_config.values()) |
| 1026 | return {k: v / total for k, v in bucket_config.items()} |
| 1027 | |
| 1028 | def generate_mm_item( |
| 1029 | self, |
| 1030 | mm_item_config: tuple[int, int, int], |
| 1031 | ) -> Mapping[str, Any]: |
| 1032 | """Create synthetic images and videos and |
| 1033 | apply process_image/process_video respectively. |
| 1034 | This follows the OpenAI API chat completions |
| 1035 | https://github.com/openai/openai-python |
| 1036 | """ |
| 1037 | if self.map_config_to_modality(mm_item_config) == "image": |
| 1038 | return process_image( |
| 1039 | self.generate_synthetic_image(mm_item_config[1], mm_item_config[0]) |
| 1040 | ) |
| 1041 | elif self.map_config_to_modality(mm_item_config) == "video": |
| 1042 | return process_video( |
| 1043 | self.generate_synthetic_video( |
| 1044 | mm_item_config[1], mm_item_config[0], mm_item_config[2] |
| 1045 | ) |
| 1046 | ) |
| 1047 | else: |
| 1048 | raise ValueError(f"Invalid multimodal item configuration: {mm_item_config}") |
| 1049 | |
| 1050 | def get_mm_item_sampling_params( |
| 1051 | self, |
| 1052 | base_items_per_request: int, |
| 1053 | num_mm_items_range_ratio: float, |
| 1054 | limit_mm_per_prompt: dict[str, int], |
| 1055 | bucket_config: dict[tuple[int, int, int], float], |
| 1056 | ) -> tuple[int, int, dict[str, int], dict[tuple[int, int, int], float]]: |
| 1057 | """Get the sampling parameters for the multimodal items.""" |
| 1058 | # Enforce num_mm_items_range_ratio <= 1 |
| 1059 | if not (0.0 <= num_mm_items_range_ratio <= 1.0): |
| 1060 | raise ValueError("num_mm_items_range_ratio must be in [0, 1].") |
| 1061 | |
| 1062 | # Ensure modalities to sample are in limit_mm_per_prompt |
| 1063 | for k, v in bucket_config.items(): |
| 1064 | # get modality from bucket config |
| 1065 | modality = self.map_config_to_modality(k) |
| 1066 | if modality not in limit_mm_per_prompt: |
| 1067 | raise ValueError( |
| 1068 | f"Modality {modality} is not in " |
| 1069 | f"limit_mm_per_prompt: " |
| 1070 | f"{limit_mm_per_prompt.keys()}" |
| 1071 | ) |
| 1072 | |
| 1073 | # Remove zero probability entries |
| 1074 | # and normalize bucket config to sum to 1 |
| 1075 | bucket_config = self.normalize_bucket_config(bucket_config) |
| 1076 | logger.info( |
| 1077 | "Normalized bucket config: %s", |
| 1078 | bucket_config, |
| 1079 | ) |
| 1080 | # Only consider limit per prompt for modalities in bucket config |
| 1081 | allowed_modalities = {self.map_config_to_modality(cfg) for cfg in bucket_config} |
| 1082 | limit_mm_per_prompt = { |
| 1083 | k: v for k, v in limit_mm_per_prompt.items() if k in allowed_modalities |
| 1084 | } |
| 1085 | if not limit_mm_per_prompt: |
| 1086 | raise ValueError("No valid limits for modalities present in bucket_config.") |
| 1087 | |
| 1088 | logger.info( |
| 1089 | "Updated mm-limit-per-prompt: %s", |
| 1090 | limit_mm_per_prompt, |
| 1091 | ) |
| 1092 | |
| 1093 | # Get max and min num mm items and ensure |
| 1094 | # it is at most the sum of limit_mm_per_prompt for all modalities |
| 1095 | max_num_mm_items = min( |
| 1096 | sum(limit_mm_per_prompt.values()), |
| 1097 | math.ceil(base_items_per_request * (1 + num_mm_items_range_ratio)), |
| 1098 | ) |
| 1099 | # Ensure min num mm items is at least 0 |
| 1100 | min_num_mm_items = max( |
| 1101 | 0, math.floor(base_items_per_request * (1 - num_mm_items_range_ratio)) |
| 1102 | ) |
| 1103 | # Raise error if min num mm items is greater than max num mm items |
| 1104 | if min_num_mm_items > max_num_mm_items: |
| 1105 | raise ValueError( |
| 1106 | f"Min num mm items is greater than max mm items: " |
| 1107 | f"{min_num_mm_items} > {max_num_mm_items}" |
| 1108 | ) |
| 1109 | |
| 1110 | logger.info( |
| 1111 | "Sampling number of multimodal items from [%s, %s]", |
| 1112 | min_num_mm_items, |
| 1113 | max_num_mm_items, |
| 1114 | ) |
| 1115 | |
| 1116 | return ( |
| 1117 | min_num_mm_items, |
| 1118 | max_num_mm_items, |
| 1119 | limit_mm_per_prompt, |
| 1120 | bucket_config, |
| 1121 | ) |
| 1122 | |
| 1123 | def get_mm_item_iterator( |
| 1124 | self, |
| 1125 | min_num_mm_items: int, |
| 1126 | max_num_mm_items: int, |
| 1127 | bucket_config: dict[tuple[int, int, int], float], |
| 1128 | limit_mm_per_prompt: dict[str, int], |
| 1129 | ) -> Iterator[tuple[int, int, int]]: |
| 1130 | """Iterator over the multimodal items for each request |
| 1131 | whose size is between min_num_mm_items and max_num_mm_items. |
| 1132 | |
| 1133 | Loop over the bucket config and sample a multimodal item. |
| 1134 | Loop until the number of multimodal items sampled is equal to |
| 1135 | request_num_mm_items or limit of multimodal items per prompt |
| 1136 | for all modalities is reached. |
| 1137 | |
| 1138 | Note: |
| 1139 | - This function operates on a per-request shallow copy of |
| 1140 | `bucket_config` (tuple->float). The original dict passed to |
| 1141 | `sample` is not mutated. If this ever changes, a test |
| 1142 | is implemented and will fail. |
| 1143 | |
| 1144 | """ |
| 1145 | # Get the number of multimodal items to sample |
| 1146 | request_num_mm_items = int( |
| 1147 | self._rng.integers(min_num_mm_items, max_num_mm_items + 1) |
| 1148 | ) |
| 1149 | # If request_num_mm_items is 0, yield an empty iterator |
| 1150 | if request_num_mm_items == 0: |
| 1151 | return |
| 1152 | # Initialize modality counters |
| 1153 | modality_counter = {self.map_config_to_modality(k): 0 for k in bucket_config} |
| 1154 | # Copy the bucket config to avoid modifying the original |
| 1155 | bucket_config_copy = bucket_config.copy() |
| 1156 | # Loop over the number of multimodal items to sample |
| 1157 | while sum(modality_counter.values()) < request_num_mm_items: |
| 1158 | # Sample a multimodal item config |
| 1159 | mm_item_config = self._rng.choice( |
| 1160 | list(bucket_config_copy.keys()), p=list(bucket_config_copy.values()) |
| 1161 | ) |
| 1162 | modality = self.map_config_to_modality(mm_item_config) |
| 1163 | # Check that modality count is less than limit per prompt |
| 1164 | if modality_counter[modality] < limit_mm_per_prompt[modality]: |
| 1165 | modality_counter[modality] += 1 |
| 1166 | yield (mm_item_config) |
| 1167 | else: |
| 1168 | # If the counter is greater than the limit per prompt |
| 1169 | # set all multimodal items of this modality to 0 |
| 1170 | for k, v in bucket_config_copy.items(): |
| 1171 | if self.map_config_to_modality(k) == modality: |
| 1172 | bucket_config_copy[k] = 0 |
| 1173 | # If all configs are 0, break the loop |
| 1174 | # This should not happen as request_num_mm_items is at most |
| 1175 | # the sum of limit_mm_per_prompt for all modalities |
| 1176 | if all(v == 0 for v in bucket_config_copy.values()): |
| 1177 | logger.warning( |
| 1178 | "Exhausted all multimodal items of modality %s", modality |
| 1179 | ) |
| 1180 | break |
| 1181 | # Renormalize the bucket config |
| 1182 | bucket_config_copy = self.normalize_bucket_config(bucket_config_copy) |
| 1183 | |
| 1184 | def sample( |
| 1185 | self, |
| 1186 | tokenizer: TokenizerLike, |
| 1187 | num_requests: int, |
| 1188 | request_id_prefix: str = "", |
| 1189 | no_oversample: bool = False, |
| 1190 | prefix_len: int = RandomDataset.DEFAULT_PREFIX_LEN, |
| 1191 | range_ratio: RangeRatio = RandomDataset.DEFAULT_RANGE_RATIO, |
| 1192 | input_len: int = RandomDataset.DEFAULT_INPUT_LEN, |
| 1193 | output_len: int = RandomDataset.DEFAULT_OUTPUT_LEN, |
| 1194 | batchsize: int = 1, |
| 1195 | max_loras: int | None = None, |
| 1196 | lora_path: str | None = None, |
| 1197 | lora_assignment: str = "random", |
| 1198 | limit_mm_per_prompt: dict[str, int] = DEFAULT_LIMIT_MM_PER_PROMPT, |
| 1199 | base_items_per_request: int = DEFAULT_BASE_ITEMS_PER_REQUEST, |
| 1200 | num_mm_items_range_ratio: float = DEFAULT_NUM_MM_ITEMS_RANGE_RATIO, |
| 1201 | bucket_config: dict[ |
| 1202 | tuple[int, int, int], float |
| 1203 | ] = DEFAULT_MM_ITEM_BUCKET_CONFIG, |
| 1204 | enable_multimodal_chat: bool = DEFAULT_ENABLE_MULTIMODAL_CHAT, |
| 1205 | **kwargs, |
| 1206 | ) -> list[SampleRequest]: |
| 1207 | if batchsize != 1: |
| 1208 | raise NotImplementedError( |
| 1209 | "batchsize > 1 is not supported for RandomMultiModalDataset." |
| 1210 | ) |
| 1211 | |
| 1212 | input_lens, output_lens, offsets = get_sampling_params( |
| 1213 | self._rng, |
| 1214 | num_requests, |
| 1215 | range_ratio, |
| 1216 | input_len, |
| 1217 | output_len, |
| 1218 | tokenizer, |
| 1219 | ) |
| 1220 | |
| 1221 | ( |
| 1222 | min_num_mm_items, |
| 1223 | max_num_mm_items, |
| 1224 | limit_mm_per_prompt, |
| 1225 | bucket_config, |
| 1226 | ) = self.get_mm_item_sampling_params( |
| 1227 | base_items_per_request, |
| 1228 | num_mm_items_range_ratio, |
| 1229 | limit_mm_per_prompt, |
| 1230 | bucket_config, |
| 1231 | ) |
| 1232 | |
| 1233 | vocab_size = tokenizer.vocab_size |
| 1234 | if is_mistral_tokenizer(tokenizer): |
| 1235 | prohibited_tokens = tokenizer.all_special_ids |
| 1236 | else: |
| 1237 | # Can't use tokenizer.all_special_ids since |
| 1238 | # it returns ONLY ids from special_tokens_map.json |
| 1239 | # We want to exclude placeholder tokens and all |
| 1240 | # tokens that indicate start/end of image as it |
| 1241 | # may break prompt replacement logic. |
| 1242 | assert hasattr(tokenizer, "added_tokens_decoder") |
| 1243 | prohibited_tokens = list( |
| 1244 | tok_id |
| 1245 | for tok_id, token in tokenizer.added_tokens_decoder.items() |
| 1246 | if token.special |
| 1247 | ) |
| 1248 | all_tokens = np.arange(vocab_size) |
| 1249 | allowed_tokens = np.array(list(set(all_tokens) - set(prohibited_tokens))) |
| 1250 | logger.debug( |
| 1251 | "Sampling from %d out of %d (vocab size)", len(allowed_tokens), vocab_size |
| 1252 | ) |
| 1253 | # Generate prefix once |
| 1254 | prefix_token_ids = self.get_prefix(tokenizer, allowed_tokens, prefix_len) |
| 1255 | # Add synthetic multimodal items to each request |
| 1256 | mm_requests = [] |
| 1257 | token_mismatch_total = 0 |
| 1258 | for i in range(num_requests): |
| 1259 | prompt, total_input_len, token_mismatch = self.generate_token_sequence( # noqa: E501 |
| 1260 | tokenizer=tokenizer, |
| 1261 | prefix_token_ids=prefix_token_ids, |
| 1262 | prefix_len=prefix_len, |
| 1263 | vocab_size=vocab_size, |
| 1264 | input_len=int(input_lens[i]), |
| 1265 | offset=int(offsets[i]), |
| 1266 | index=i, |
| 1267 | allowed_tokens=allowed_tokens, |
| 1268 | ) |
| 1269 | token_mismatch_total += token_mismatch |
| 1270 | # Get multimodal item iterator for a given request |
| 1271 | mm_item_iterator = self.get_mm_item_iterator( |
| 1272 | min_num_mm_items, |
| 1273 | max_num_mm_items, |
| 1274 | bucket_config, |
| 1275 | limit_mm_per_prompt, |
| 1276 | ) |
| 1277 | |
| 1278 | mm_content = cast( |
| 1279 | list[dict[str, Any]], |
| 1280 | [ |
| 1281 | self.generate_mm_item(mm_item_config) |
| 1282 | for mm_item_config in mm_item_iterator |
| 1283 | ], |
| 1284 | ) |
| 1285 | |
| 1286 | if enable_multimodal_chat: |
| 1287 | # NOTE: For now this option is only provided for completeness |
| 1288 | # given that the serve.py benchmark currently does not use it. |
| 1289 | mm_chat_prompt: Any = prompt |
| 1290 | mm_chat_prompt = self.apply_multimodal_chat_transformation( |
| 1291 | prompt, mm_content |
| 1292 | ) |
| 1293 | sample_request = SampleRequest( |
| 1294 | prompt=mm_chat_prompt, |
| 1295 | prompt_len=total_input_len, |
| 1296 | expected_output_len=int(output_lens[i]), |
| 1297 | multi_modal_data=None, |
| 1298 | request_id=request_id_prefix + str(i), |
| 1299 | ) |
| 1300 | else: |
| 1301 | sample_request = SampleRequest( |
| 1302 | prompt=prompt, |
| 1303 | prompt_len=total_input_len, |
| 1304 | expected_output_len=int(output_lens[i]), |
| 1305 | multi_modal_data=mm_content, |
| 1306 | request_id=request_id_prefix + str(i), |
| 1307 | ) |
| 1308 | mm_requests.append(sample_request) |
| 1309 | |
| 1310 | if token_mismatch_total != 0: |
| 1311 | sign = "more" if token_mismatch_total > 0 else "fewer" |
| 1312 | logger.warning( |
| 1313 | "Across all generated prompts, there were %d %s tokens " |
| 1314 | "than expected after decoding and re-encoding. This is " |
| 1315 | "expected due to the imperfect nature of the sampling " |
| 1316 | "procedure.", |
| 1317 | abs(token_mismatch_total), |
| 1318 | sign, |
| 1319 | ) |
| 1320 | |
| 1321 | return mm_requests |
| 1322 | |
| 1323 | |
| 1324 | # ----------------------------------------------------------------------------- |
| 1325 | # ShareGPT Dataset Implementation |
| 1326 | # ----------------------------------------------------------------------------- |
| 1327 | |
| 1328 | |
| 1329 | class ShareGPTDataset(BenchmarkDataset): |
| 1330 | """Implements the ShareGPT dataset. Loads data from a JSON file and generates |
| 1331 | sample requests based on conversation turns. |
| 1332 | """ |
| 1333 | |
| 1334 | def __init__(self, **kwargs) -> None: |
| 1335 | super().__init__(**kwargs) |
| 1336 | self.load_data() |
| 1337 | |
| 1338 | def load_data(self) -> None: |
| 1339 | if self.dataset_path is None: |
| 1340 | raise ValueError("dataset_path must be provided for loading data.") |
| 1341 | |
| 1342 | with open(self.dataset_path, encoding="utf-8") as f: |
| 1343 | self.data = json.load(f) |
| 1344 | # Filter entries with at least two conversation turns. |
| 1345 | self.data = [ |
| 1346 | entry |
| 1347 | for entry in self.data |
| 1348 | if "conversations" in entry and len(entry["conversations"]) >= 2 |
| 1349 | ] |
| 1350 | random.seed(self.random_seed) |
| 1351 | if not getattr(self, "disable_shuffle", False): |
| 1352 | random.shuffle(self.data) |
| 1353 | |
| 1354 | def sample( |
| 1355 | self, |
| 1356 | tokenizer: TokenizerLike, |
| 1357 | num_requests: int, |
| 1358 | request_id_prefix: str = "", |
| 1359 | no_oversample: bool = False, |
| 1360 | lora_path: str | None = None, |
| 1361 | max_loras: int | None = None, |
| 1362 | output_len: int | None = None, |
| 1363 | enable_multimodal_chat: bool = False, |
| 1364 | lora_assignment: str = "random", |
| 1365 | **kwargs, |
| 1366 | ) -> list[SampleRequest]: |
| 1367 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 1368 | samples: list[SampleRequest] = [] |
| 1369 | ind = 0 |
| 1370 | for entry in self.data: |
| 1371 | if len(samples) >= num_requests: |
| 1372 | break |
| 1373 | prompt, completion = ( |
| 1374 | entry["conversations"][0]["value"], |
| 1375 | entry["conversations"][1]["value"], |
| 1376 | ) |
| 1377 | |
| 1378 | lora_request = self.get_lora_request( |
| 1379 | index=ind, |
| 1380 | max_loras=max_loras, |
| 1381 | lora_path=lora_path, |
| 1382 | lora_assignment=lora_assignment, |
| 1383 | ) |
| 1384 | prompt_ids = tokenizer(prompt).input_ids |
| 1385 | completion_ids = tokenizer(completion).input_ids |
| 1386 | prompt_len = len(prompt_ids) |
| 1387 | new_output_len = len(completion_ids) if output_len is None else output_len |
| 1388 | if not is_valid_sequence( |
| 1389 | prompt_len, |
| 1390 | new_output_len, |
| 1391 | skip_min_output_len_check=output_len is not None, |
| 1392 | ): |
| 1393 | continue |
| 1394 | if image_path := entry.get("image"): |
| 1395 | mm_content = process_image(image_path) |
| 1396 | elif video_path := entry.get("video"): |
| 1397 | mm_content = process_video(video_path) |
| 1398 | else: |
| 1399 | mm_content = None |
| 1400 | if enable_multimodal_chat: |
| 1401 | prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) |
| 1402 | samples.append( |
| 1403 | SampleRequest( |
| 1404 | prompt=prompt, |
| 1405 | prompt_len=prompt_len, |
| 1406 | expected_output_len=new_output_len, |
| 1407 | lora_request=lora_request, |
| 1408 | multi_modal_data=mm_content, |
| 1409 | request_id=request_id_prefix + str(ind), |
| 1410 | ) |
| 1411 | ) |
| 1412 | ind += 1 |
| 1413 | self.maybe_oversample_requests( |
| 1414 | samples, num_requests, request_id_prefix, no_oversample |
| 1415 | ) |
| 1416 | return samples |
| 1417 | |
| 1418 | |
| 1419 | class TimedTrace(BenchmarkDataset): |
| 1420 | """Implements a base class to replay various timed traces. |
| 1421 | Loads data from a JSON file and generates sample requests |
| 1422 | based on the timing information in the traces. |
| 1423 | """ |
| 1424 | |
| 1425 | def __init__(self, **kwargs) -> None: |
| 1426 | super().__init__(**kwargs) |
| 1427 | random.seed(self.random_seed) |
| 1428 | np.random.seed(self.random_seed) |
| 1429 | |
| 1430 | # Set parameters with defaults from timed_trace_group arguments |
| 1431 | self.chunk_size = int(kwargs.get("timed_trace_chunk_hash_size", 16)) |
| 1432 | self.sec_multiplier = float(kwargs.get("timed_trace_sec_multiplier", 1)) |
| 1433 | self.label_ts = str(kwargs.get("timed_trace_label_timestamp")) |
| 1434 | self.label_input_length = str(kwargs.get("timed_trace_label_input_length")) |
| 1435 | self.label_output_length = str(kwargs.get("timed_trace_label_output_length")) |
| 1436 | self.label_hash_ids = str(kwargs.get("timed_trace_label_hash_ids")) |
| 1437 | print( |
| 1438 | f"timed-trace: chunk_size: {self.chunk_size}, " |
| 1439 | f"sec_multiplier: {self.sec_multiplier}, " |
| 1440 | f'label_ts: "{self.label_ts}", ' |
| 1441 | f'label_input_length: "{self.label_input_length}", ' |
| 1442 | f'label_output_length: "{self.label_output_length}", ' |
| 1443 | f'label_hash_ids: "{self.label_hash_ids}"' |
| 1444 | ) |
| 1445 | self._expanded_generated_prompts: dict[str, Any] = {} |
| 1446 | self.load_data() |
| 1447 | |
| 1448 | def load_data(self) -> None: |
| 1449 | # check if the file is there |
| 1450 | if self.dataset_path is None: |
| 1451 | raise ValueError("dataset_path must be provided for loading data.") |
| 1452 | |
| 1453 | # load and we will do transformation once we have the Tokenizer available |
| 1454 | # this is jsonl data format |
| 1455 | with open(self.dataset_path) as f: |
| 1456 | self.data = f.readlines() |
| 1457 | |
| 1458 | def _sample_token( |
| 1459 | self, num_tokens: int, tokenizer: TokenizerLike, seed: int | None = None |
| 1460 | ) -> list[int]: |
| 1461 | # Initialize vocab only if it doesn't exist yet |
| 1462 | if not hasattr(self, "vocab"): |
| 1463 | self.vocab = tokenizer.get_vocab() |
| 1464 | # Remove the special tokens. |
| 1465 | self.vocab = { |
| 1466 | k: v |
| 1467 | for k, v in self.vocab.items() |
| 1468 | if v not in tokenizer.all_special_ids |
| 1469 | } |
| 1470 | # Create a sorted list of vocab values for deterministic sampling |
| 1471 | self.vocab_values_sorted = sorted(self.vocab.values()) |
| 1472 | |
| 1473 | # Use the provided seed if given, otherwise use global random state |
| 1474 | if seed is not None: |
| 1475 | rng = random.Random(seed) |
| 1476 | sampled_token_ids = rng.choices(self.vocab_values_sorted, k=num_tokens) |
| 1477 | else: |
| 1478 | sampled_token_ids = random.choices(self.vocab_values_sorted, k=num_tokens) |
| 1479 | |
| 1480 | return sampled_token_ids |
| 1481 | |
| 1482 | def _expand_prompt( |
| 1483 | self, |
| 1484 | chunked_hashes: list[int], |
| 1485 | target_input_size: int, |
| 1486 | tokenizer: TokenizerLike, |
| 1487 | ) -> list[int]: |
| 1488 | raw_tokenized_prompt = [] |
| 1489 | for h in chunked_hashes: |
| 1490 | # Calculate how many tokens to expand for this chunk |
| 1491 | expanded_size = ( |
| 1492 | self.chunk_size |
| 1493 | if target_input_size >= self.chunk_size |
| 1494 | else target_input_size |
| 1495 | ) |
| 1496 | |
| 1497 | # Cache key includes size for partial chunks at the end |
| 1498 | key = f"{h}:{expanded_size}" |
| 1499 | |
| 1500 | if key not in self._expanded_generated_prompts: |
| 1501 | # Convert key to a deterministic seed |
| 1502 | key_seed = hash(key) & 0xFFFFFFFF # Convert to 32-bit int |
| 1503 | self._expanded_generated_prompts[key] = self._sample_token( |
| 1504 | expanded_size, tokenizer, seed=key_seed |
| 1505 | ) |
| 1506 | # once inserted get the tokenized prompt and append to the list |
| 1507 | raw_tokenized_prompt.extend(self._expanded_generated_prompts[key]) |
| 1508 | target_input_size -= expanded_size |
| 1509 | |
| 1510 | if target_input_size <= 0: |
| 1511 | break |
| 1512 | |
| 1513 | return raw_tokenized_prompt |
| 1514 | |
| 1515 | def sample( |
| 1516 | self, |
| 1517 | tokenizer: TokenizerLike, |
| 1518 | num_requests: int, |
| 1519 | request_id_prefix: str = "", |
| 1520 | no_oversample: bool = False, |
| 1521 | **kwargs, |
| 1522 | ) -> list[SampleRequest]: |
| 1523 | samples: list[SampleRequest] = [] |
| 1524 | assert tokenizer is not None, "Tokenizer must be provided, now is Null" |
| 1525 | assert self.data is not None, "Data must be loaded before sampling" |
| 1526 | |
| 1527 | for ind, entry in enumerate(self.data): |
| 1528 | if len(samples) >= num_requests: |
| 1529 | break |
| 1530 | |
| 1531 | # now we create the SampleRequest with timing info |
| 1532 | entry = json.loads(entry.strip()) |
| 1533 | input_length = entry.get(self.label_input_length) |
| 1534 | if input_length is None: |
| 1535 | raise ValueError( |
| 1536 | f"Input length field '{self.label_input_length}' " |
| 1537 | f"not found in trace entry. " |
| 1538 | f"Available fields: {list(entry.keys())}. " |
| 1539 | f"Use --label-input-length to specify the correct " |
| 1540 | f"field name." |
| 1541 | ) |
| 1542 | new_output_len = entry.get(self.label_output_length) |
| 1543 | if new_output_len is None: |
| 1544 | raise ValueError( |
| 1545 | f"Output length field '{self.label_output_length}' " |
| 1546 | f"not found in trace entry. " |
| 1547 | f"Available fields: {list(entry.keys())}. " |
| 1548 | f"Use --label-output-length to specify the correct " |
| 1549 | f"field name." |
| 1550 | ) |
| 1551 | prompt_ids = self._expand_prompt( |
| 1552 | entry.get(self.label_hash_ids, []), input_length, tokenizer |
| 1553 | ) |
| 1554 | |
| 1555 | # Get timestamp with proper error handling |
| 1556 | ts_value = entry.get(self.label_ts) |
| 1557 | if ts_value is None: |
| 1558 | raise ValueError( |
| 1559 | f"Timestamp field '{self.label_ts}' not found in trace entry. " |
| 1560 | f"Available fields: {list(entry.keys())}. " |
| 1561 | f"Use --label-timestamp to specify the correct field name." |
| 1562 | ) |
| 1563 | timestamp = float(ts_value) * self.sec_multiplier |
| 1564 | |
| 1565 | prompt_len = len(prompt_ids) |
| 1566 | |
| 1567 | samples.append( |
| 1568 | SampleRequest( |
| 1569 | prompt=prompt_ids, |
| 1570 | prompt_len=prompt_len, |
| 1571 | expected_output_len=new_output_len, |
| 1572 | lora_request=None, |
| 1573 | multi_modal_data=None, |
| 1574 | request_id=request_id_prefix + str(ind), |
| 1575 | timestamp=timestamp, |
| 1576 | ) |
| 1577 | ) |
| 1578 | return samples |
| 1579 | |
| 1580 | |
| 1581 | def add_dataset_parser(parser: FlexibleArgumentParser): |
| 1582 | parser.add_argument( |
| 1583 | "--trust-remote-code", |
| 1584 | action="store_true", |
| 1585 | help="Trust remote code from huggingface", |
| 1586 | ) |
| 1587 | parser.add_argument("--seed", type=int, default=0) |
| 1588 | parser.add_argument( |
| 1589 | "--num-prompts", |
| 1590 | type=int, |
| 1591 | default=DEFAULT_NUM_PROMPTS, |
| 1592 | help="Number of prompts to process.", |
| 1593 | ) |
| 1594 | parser.add_argument( |
| 1595 | "--dataset-name", |
| 1596 | type=str, |
| 1597 | default="random", |
| 1598 | choices=[ |
| 1599 | "sharegpt", |
| 1600 | "burstgpt", |
| 1601 | "sonnet", |
| 1602 | "random", |
| 1603 | "random-mm", |
| 1604 | "random-rerank", |
| 1605 | "hf", |
| 1606 | "custom", |
| 1607 | "custom_audio", |
| 1608 | "custom_image", |
| 1609 | "prefix_repetition", |
| 1610 | "spec_bench", |
| 1611 | "speed_bench", |
| 1612 | "timed_trace", |
| 1613 | ], |
| 1614 | help="Name of the dataset to benchmark on.", |
| 1615 | ) |
| 1616 | parser.add_argument( |
| 1617 | "--no-stream", |
| 1618 | action="store_true", |
| 1619 | help="Do not load the dataset in streaming mode.", |
| 1620 | ) |
| 1621 | parser.add_argument( |
| 1622 | "--dataset-path", |
| 1623 | type=str, |
| 1624 | default=None, |
| 1625 | help="Path to the sharegpt/sonnet dataset or the HF dataset ID if " |
| 1626 | "using HF dataset.", |
| 1627 | ) |
| 1628 | parser.add_argument( |
| 1629 | "--no-oversample", |
| 1630 | action="store_true", |
| 1631 | help="Do not oversample if the dataset has fewer samples than num-prompts.", |
| 1632 | ) |
| 1633 | parser.add_argument( |
| 1634 | "--skip-chat-template", |
| 1635 | action="store_true", |
| 1636 | help="Skip applying chat template to prompt for datasets that support it.", |
| 1637 | ) |
| 1638 | parser.add_argument( |
| 1639 | "--enable-multimodal-chat", |
| 1640 | action="store_true", |
| 1641 | help="Enable multimodal chat transformation for datasets that support it.", |
| 1642 | ) |
| 1643 | parser.add_argument( |
| 1644 | "--disable-shuffle", |
| 1645 | action="store_true", |
| 1646 | help="Disable shuffling of dataset samples for deterministic ordering.", |
| 1647 | ) |
| 1648 | |
| 1649 | # group for dataset specific arguments |
| 1650 | custom_group = parser.add_argument_group("custom dataset options") |
| 1651 | custom_group.add_argument( |
| 1652 | "--custom-output-len", |
| 1653 | type=int, |
| 1654 | default=256, |
| 1655 | help="Number of output tokens per request. Unless it is set to -1, the " |
| 1656 | "value overrides potential output length loaded from the dataset. It is " |
| 1657 | "used only for custom dataset.", |
| 1658 | ) |
| 1659 | custom_group.add_argument( |
| 1660 | "--custom-ensure-client-side-data", |
| 1661 | action="store_true", |
| 1662 | help=( |
| 1663 | "Ensure custom dataset media is sent as client-side data instead " |
| 1664 | "of references. For custom_image datasets, this loads local and " |
| 1665 | "HTTP(S) images on the benchmark client and encodes them as " |
| 1666 | "base64 data URLs. Existing data:image URLs are kept unchanged." |
| 1667 | ), |
| 1668 | ) |
| 1669 | |
| 1670 | spec_bench_group = parser.add_argument_group("spec bench dataset options") |
| 1671 | spec_bench_group.add_argument( |
| 1672 | "--spec-bench-output-len", |
| 1673 | type=int, |
| 1674 | default=256, |
| 1675 | help="Num of output tokens per request, used only for spec bench dataset.", |
| 1676 | ) |
| 1677 | spec_bench_group.add_argument( |
| 1678 | "--spec-bench-category", |
| 1679 | type=str, |
| 1680 | default=None, |
| 1681 | help="Category for spec bench dataset. If None, use all categories.", |
| 1682 | ) |
| 1683 | |
| 1684 | sonnet_group = parser.add_argument_group("sonnet dataset options") |
| 1685 | sonnet_group.add_argument( |
| 1686 | "--sonnet-input-len", |
| 1687 | type=int, |
| 1688 | default=550, |
| 1689 | help="Number of input tokens per request, used only for sonnet dataset.", |
| 1690 | ) |
| 1691 | sonnet_group.add_argument( |
| 1692 | "--sonnet-output-len", |
| 1693 | type=int, |
| 1694 | default=150, |
| 1695 | help="Number of output tokens per request, used only for sonnet dataset.", |
| 1696 | ) |
| 1697 | sonnet_group.add_argument( |
| 1698 | "--sonnet-prefix-len", |
| 1699 | type=int, |
| 1700 | default=200, |
| 1701 | help="Number of prefix tokens per request, used only for sonnet dataset.", |
| 1702 | ) |
| 1703 | |
| 1704 | sharegpt_group = parser.add_argument_group("sharegpt dataset options") |
| 1705 | sharegpt_group.add_argument( |
| 1706 | "--sharegpt-output-len", |
| 1707 | type=int, |
| 1708 | default=None, |
| 1709 | help="Output length for each request. Overrides the output length " |
| 1710 | "from the ShareGPT dataset.", |
| 1711 | ) |
| 1712 | |
| 1713 | timed_trace_group = parser.add_argument_group("timed-trace dataset options") |
| 1714 | timed_trace_group.add_argument( |
| 1715 | "--timed-trace-chunk-hash-size", |
| 1716 | type=int, |
| 1717 | default=16, |
| 1718 | help=( |
| 1719 | "Each hash tokens, if present, represent how many token " |
| 1720 | "hashes. For example in the Moonshot traces it is 512, while " |
| 1721 | "the Qwen/Alibaba has 16." |
| 1722 | ), |
| 1723 | ) |
| 1724 | timed_trace_group.add_argument( |
| 1725 | "--timed-trace-sec-multiplier", |
| 1726 | type=float, |
| 1727 | default=1, |
| 1728 | help=( |
| 1729 | "What multiplier to use when converting timestamps to " |
| 1730 | "seconds. We will multiply timestamps by this. For example" |
| 1731 | "if the timestamps are in milliseconds, then pass 0.001." |
| 1732 | "If they are already in seconds, then the default 1 is sufficient." |
| 1733 | ), |
| 1734 | ) |
| 1735 | timed_trace_group.add_argument( |
| 1736 | "--timed-trace-label-timestamp", |
| 1737 | type=str, |
| 1738 | default="timestamp", |
| 1739 | help="What json label to use to index the timestamp in the trace.", |
| 1740 | ) |
| 1741 | timed_trace_group.add_argument( |
| 1742 | "--timed-trace-label-input-length", |
| 1743 | type=str, |
| 1744 | default="input_length", |
| 1745 | help=("What json label to use to index the input length field in the trace."), |
| 1746 | ) |
| 1747 | timed_trace_group.add_argument( |
| 1748 | "--timed-trace-label-output-length", |
| 1749 | type=str, |
| 1750 | default="output_length", |
| 1751 | help=("What json label to use to index the output length field in the trace."), |
| 1752 | ) |
| 1753 | timed_trace_group.add_argument( |
| 1754 | "--timed-trace-label-hash-ids", |
| 1755 | type=str, |
| 1756 | default="hash_ids", |
| 1757 | help=("What json label to use to index the hash ids for the input prompts."), |
| 1758 | ) |
| 1759 | |
| 1760 | blazedit_group = parser.add_argument_group("blazedit dataset options") |
| 1761 | blazedit_group.add_argument( |
| 1762 | "--blazedit-min-distance", |
| 1763 | type=float, |
| 1764 | default=0.0, |
| 1765 | help="Minimum distance for blazedit dataset. Min: 0, Max: 1.0", |
| 1766 | ) |
| 1767 | blazedit_group.add_argument( |
| 1768 | "--blazedit-max-distance", |
| 1769 | type=float, |
| 1770 | default=1.0, |
| 1771 | help="Maximum distance for blazedit dataset. Min: 0, Max: 1.0", |
| 1772 | ) |
| 1773 | |
| 1774 | asr_group = parser.add_argument_group("asr dataset options") |
| 1775 | asr_group.add_argument( |
| 1776 | "--asr-max-audio-len-sec", |
| 1777 | type=float, |
| 1778 | default=float("inf"), |
| 1779 | help="Maximum audio length in seconds for ASR dataset.", |
| 1780 | ) |
| 1781 | asr_group.add_argument( |
| 1782 | "--asr-min-audio-len-sec", |
| 1783 | type=float, |
| 1784 | default=0.0, |
| 1785 | help="Minimum audio length in seconds for ASR dataset.", |
| 1786 | ) |
| 1787 | |
| 1788 | random_group = parser.add_argument_group("random dataset options") |
| 1789 | add_random_dataset_base_args(random_group) |
| 1790 | |
| 1791 | random_mm_group = parser.add_argument_group( |
| 1792 | "random multimodal dataset options extended from random dataset" |
| 1793 | ) |
| 1794 | add_random_multimodal_dataset_args(random_mm_group) |
| 1795 | |
| 1796 | hf_group = parser.add_argument_group("hf dataset options") |
| 1797 | hf_group.add_argument( |
| 1798 | "--hf-subset", type=str, default=None, help="Subset of the HF dataset." |
| 1799 | ) |
| 1800 | hf_group.add_argument( |
| 1801 | "--hf-split", type=str, default=None, help="Split of the HF dataset." |
| 1802 | ) |
| 1803 | hf_group.add_argument( |
| 1804 | "--hf-name", |
| 1805 | type=str, |
| 1806 | default=None, |
| 1807 | help=( |
| 1808 | "Name of the dataset on HuggingFace " |
| 1809 | "(e.g., 'lmarena-ai/VisionArena-Chat'). " |
| 1810 | "Specify this if your dataset-path is a local path." |
| 1811 | ), |
| 1812 | ) |
| 1813 | hf_group.add_argument( |
| 1814 | "--hf-output-len", |
| 1815 | type=int, |
| 1816 | default=None, |
| 1817 | help="Output length for each request. Overrides the output lengths " |
| 1818 | "from the sampled HF dataset.", |
| 1819 | ) |
| 1820 | |
| 1821 | bfcl_group = parser.add_argument_group( |
| 1822 | "BFCL dataset options", description=BFCLDataset.__doc__ |
| 1823 | ) |
| 1824 | bfcl_group.add_argument( |
| 1825 | "--bfcl-categories", |
| 1826 | type=lambda s: [c.strip() for c in s.split(",") if c.strip()], |
| 1827 | default=None, |
| 1828 | help="Comma-separated list of BFCL v3 category names (without the " |
| 1829 | "'BFCL_v3_' prefix or '.json' suffix) to sample from, e.g. " |
| 1830 | "'simple,live_simple,multiple'. Defaults to " |
| 1831 | f"'{','.join(BFCLDataset.DEFAULT_CATEGORIES)}'.", |
| 1832 | ) |
| 1833 | |
| 1834 | prefix_repetition_group = parser.add_argument_group( |
| 1835 | "prefix repetition dataset options" |
| 1836 | ) |
| 1837 | prefix_repetition_group.add_argument( |
| 1838 | "--prefix-repetition-prefix-len", |
| 1839 | type=int, |
| 1840 | default=256, |
| 1841 | help="Number of prefix tokens per request, used only for prefix " |
| 1842 | "repetition dataset.", |
| 1843 | ) |
| 1844 | prefix_repetition_group.add_argument( |
| 1845 | "--prefix-repetition-suffix-len", |
| 1846 | type=int, |
| 1847 | default=256, |
| 1848 | help="Number of suffix tokens per request, used only for prefix " |
| 1849 | "repetition dataset. Total input length is prefix_len + suffix_len.", |
| 1850 | ) |
| 1851 | prefix_repetition_group.add_argument( |
| 1852 | "--prefix-repetition-num-prefixes", |
| 1853 | type=int, |
| 1854 | default=10, |
| 1855 | help="Number of prefixes to generate, used only for prefix repetition " |
| 1856 | "dataset. Prompts per prefix is num_requests // num_prefixes.", |
| 1857 | ) |
| 1858 | prefix_repetition_group.add_argument( |
| 1859 | "--prefix-repetition-output-len", |
| 1860 | type=int, |
| 1861 | default=128, |
| 1862 | help="Number of output tokens per request, used only for prefix " |
| 1863 | "repetition dataset.", |
| 1864 | ) |
| 1865 | |
| 1866 | speed_bench_group = parser.add_argument_group( |
| 1867 | "speed bench dataset options", description=SpeedBench.__doc__ |
| 1868 | ) |
| 1869 | speed_bench_group.add_argument( |
| 1870 | "--speed-bench-dataset-subset", |
| 1871 | type=str, |
| 1872 | default="qualitative", |
| 1873 | choices={ |
| 1874 | "qualitative", |
| 1875 | "throughput_1k", |
| 1876 | "throughput_2k", |
| 1877 | "throughput_8k", |
| 1878 | "throughput_16k", |
| 1879 | "throughput_32k", |
| 1880 | }, |
| 1881 | help="Subset of the SPEED-Bench dataset.", |
| 1882 | ) |
| 1883 | speed_bench_group.add_argument( |
| 1884 | "--speed-bench-output-len", |
| 1885 | type=int, |
| 1886 | default=4096, |
| 1887 | help="Num of output tokens per request, used only for speed bench dataset.", |
| 1888 | ) |
| 1889 | speed_bench_group.add_argument( |
| 1890 | "--speed-bench-category", |
| 1891 | type=str, |
| 1892 | default=None, |
| 1893 | help="Category for speed bench dataset. If None, use all categories.", |
| 1894 | ) |
| 1895 | |
| 1896 | |
| 1897 | def add_random_dataset_base_args( |
| 1898 | parser_or_group: FlexibleArgumentParser | argparse._ArgumentGroup, |
| 1899 | ) -> None: |
| 1900 | """Add CLI arguments for base random dataset options. |
| 1901 | |
| 1902 | This function adds arguments needed for: |
| 1903 | - random (random dataset) |
| 1904 | - random-mm (random multimodal dataset) |
| 1905 | - random-rerank (random dataset for reranking) |
| 1906 | |
| 1907 | Args: |
| 1908 | parser_or_group: Either a parser or an argument group to add arguments to. |
| 1909 | |
| 1910 | """ |
| 1911 | parser_or_group.add_argument( |
| 1912 | "--random-input-len", |
| 1913 | type=int, |
| 1914 | default=1024, |
| 1915 | help="Number of input tokens per request, used only for random sampling.", |
| 1916 | ) |
| 1917 | parser_or_group.add_argument( |
| 1918 | "--random-output-len", |
| 1919 | type=int, |
| 1920 | default=128, |
| 1921 | help="Number of output tokens per request, used only for random sampling.", |
| 1922 | ) |
| 1923 | parser_or_group.add_argument( |
| 1924 | "--random-range-ratio", |
| 1925 | type=str, |
| 1926 | default="0.0", |
| 1927 | help="Range ratio for sampling input/output length, " |
| 1928 | "used only for random sampling. A single float applies to both " |
| 1929 | 'ISL and OSL. A JSON dict like \'{"input": 0.3, "output": 0.5}\' ' |
| 1930 | "sets them independently. Values must be in [0, 1).", |
| 1931 | ) |
| 1932 | parser_or_group.add_argument( |
| 1933 | "--random-prefix-len", |
| 1934 | type=int, |
| 1935 | default=0, |
| 1936 | help=( |
| 1937 | "Number of fixed prefix tokens before the random context " |
| 1938 | "in a request. " |
| 1939 | "The total input length is the sum of `random-prefix-len` and " |
| 1940 | "a random " |
| 1941 | "context length sampled from [input_len * (1 - range_ratio), " |
| 1942 | "input_len * (1 + range_ratio)]." |
| 1943 | ), |
| 1944 | ) |
| 1945 | parser_or_group.add_argument( |
| 1946 | "--random-batch-size", |
| 1947 | type=int, |
| 1948 | default=1, |
| 1949 | help=("Batch size for random sampling. Only used for embeddings benchmark."), |
| 1950 | ) |
| 1951 | parser_or_group.add_argument( |
| 1952 | "--no-reranker", |
| 1953 | action="store_true", |
| 1954 | help=( |
| 1955 | "Whether the model supports reranking natively." |
| 1956 | " Only used for reranker benchmark." |
| 1957 | ), |
| 1958 | ) |
| 1959 | |
| 1960 | |
| 1961 | def add_random_multimodal_dataset_args( |
| 1962 | parser_or_group: FlexibleArgumentParser | argparse._ArgumentGroup, |
| 1963 | ) -> None: |
| 1964 | """Add CLI arguments for random multimodal dataset options. |
| 1965 | |
| 1966 | This function adds arguments needed for: |
| 1967 | - random-mm (random multimodal dataset) |
| 1968 | |
| 1969 | Args: |
| 1970 | parser_or_group: Either a parser or an argument group to add arguments to. |
| 1971 | |
| 1972 | """ |
| 1973 | parser_or_group.add_argument( |
| 1974 | "--random-mm-base-items-per-request", |
| 1975 | type=int, |
| 1976 | default=RandomMultiModalDataset.DEFAULT_BASE_ITEMS_PER_REQUEST, |
| 1977 | help=( |
| 1978 | "Base number of multimodal items per request for random-mm. " |
| 1979 | "Actual per-request count is sampled around this base using " |
| 1980 | "--random-mm-num-mm-items-range-ratio." |
| 1981 | ), |
| 1982 | ) |
| 1983 | parser_or_group.add_argument( |
| 1984 | "--random-mm-num-mm-items-range-ratio", |
| 1985 | type=float, |
| 1986 | default=RandomMultiModalDataset.DEFAULT_NUM_MM_ITEMS_RANGE_RATIO, |
| 1987 | help=( |
| 1988 | "Range ratio r in [0, 1] for sampling items per request. " |
| 1989 | "We sample uniformly from the closed integer range " |
| 1990 | "[floor(n*(1-r)), ceil(n*(1+r))] " |
| 1991 | "where n is the base items per request. " |
| 1992 | "r=0 keeps it fixed; r=1 allows 0 items. The maximum is clamped " |
| 1993 | "to the sum of per-modality limits from " |
| 1994 | "--random-mm-limit-mm-per-prompt. " |
| 1995 | "An error is raised if the computed min exceeds the max." |
| 1996 | ), |
| 1997 | ) |
| 1998 | parser_or_group.add_argument( |
| 1999 | "--random-mm-limit-mm-per-prompt", |
| 2000 | type=json.loads, |
| 2001 | default=RandomMultiModalDataset.DEFAULT_LIMIT_MM_PER_PROMPT, |
| 2002 | help=( |
| 2003 | "Per-modality hard caps for items attached per request, e.g. " |
| 2004 | '\'{"image": 3, "video": 0}\'. The sampled per-request item ' |
| 2005 | "count is clamped to the sum of these limits. When a modality " |
| 2006 | "reaches its cap, its buckets are excluded and probabilities are " |
| 2007 | "renormalized." |
| 2008 | "OBS.: Only image sampling is supported for now." |
| 2009 | ), |
| 2010 | ) |
| 2011 | |
| 2012 | def _parse_mm_bucket_config(v: object) -> dict[tuple[int, int, int], float]: |
| 2013 | # If already a dict (e.g., programmatic call), normalize keys |
| 2014 | def normalize(d: dict) -> dict[tuple[int, int, int], float]: |
| 2015 | out: dict[tuple[int, int, int], float] = {} |
| 2016 | for k, val in d.items(): |
| 2017 | key = k |
| 2018 | if isinstance(key, str): |
| 2019 | with suppress(Exception): |
| 2020 | key = ast.literal_eval(key) |
| 2021 | if not ( |
| 2022 | isinstance(key, tuple) |
| 2023 | and len(key) == 3 |
| 2024 | and all(isinstance(x, int) for x in key) |
| 2025 | ): |
| 2026 | raise ValueError( |
| 2027 | f"Invalid bucket key {k!r}. Expected tuple (H, W, T)." |
| 2028 | ) |
| 2029 | out[(int(key[0]), int(key[1]), int(key[2]))] = float(val) |
| 2030 | return out |
| 2031 | |
| 2032 | if isinstance(v, dict): |
| 2033 | return normalize(v) |
| 2034 | if isinstance(v, str): |
| 2035 | # Python literal (supports tuple keys) |
| 2036 | parsed = ast.literal_eval(v) |
| 2037 | if not isinstance(parsed, dict): |
| 2038 | raise ValueError("Bucket config must parse to a dict.") |
| 2039 | return normalize(parsed) |
| 2040 | raise ValueError("Unsupported value for --random-mm-bucket-config.") |
| 2041 | |
| 2042 | parser_or_group.add_argument( |
| 2043 | "--random-mm-bucket-config", |
| 2044 | type=_parse_mm_bucket_config, |
| 2045 | default=RandomMultiModalDataset.DEFAULT_MM_ITEM_BUCKET_CONFIG, |
| 2046 | help=( |
| 2047 | "The bucket config is a dictionary mapping a multimodal item" |
| 2048 | "sampling configuration to a probability." |
| 2049 | "Currently allows for 2 modalities: images and videos. " |
| 2050 | "An bucket key is a tuple of (height, width, num_frames)" |
| 2051 | "The value is the probability of sampling that specific item. " |
| 2052 | "Example: " |
| 2053 | "--random-mm-bucket-config " |
| 2054 | "{(256, 256, 1): 0.5, (720, 1280, 1): 0.4, (720, 1280, 16): 0.10} " |
| 2055 | "First item: images with resolution 256x256 w.p. 0.5" |
| 2056 | "Second item: images with resolution 720x1280 w.p. 0.4 " |
| 2057 | "Third item: videos with resolution 720x1280 and 16 frames w.p. 0.1" |
| 2058 | "OBS.: If the probabilities do not sum to 1, they are normalized." |
| 2059 | "OBS bis.: Only image sampling is supported for now." |
| 2060 | ), |
| 2061 | ) |
| 2062 | |
| 2063 | |
| 2064 | def _parse_range_ratio(value: str) -> RangeRatio: |
| 2065 | """Parse a ``--random-range-ratio`` CLI string. |
| 2066 | |
| 2067 | Accepts either a plain float (``"0.3"``) or a JSON dict |
| 2068 | (``'{"input": 0.3, "output": 0.5}'``). |
| 2069 | """ |
| 2070 | try: |
| 2071 | return float(value) |
| 2072 | except ValueError: |
| 2073 | return json.loads(value) |
| 2074 | |
| 2075 | |
| 2076 | def get_samples( |
| 2077 | args, |
| 2078 | tokenizer: TokenizerLike | None, |
| 2079 | *, |
| 2080 | multimodal_backends: tuple[str, ...] = ("openai-chat", "openai-audio"), |
| 2081 | ) -> list[SampleRequest]: |
| 2082 | if not hasattr(args, "request_id_prefix"): |
| 2083 | args.request_id_prefix = "" |
| 2084 | |
| 2085 | if hasattr(args, "random_range_ratio") and isinstance(args.random_range_ratio, str): |
| 2086 | args.random_range_ratio = _parse_range_ratio(args.random_range_ratio) |
| 2087 | |
| 2088 | if args.dataset_name == "custom": |
| 2089 | dataset = CustomDataset( |
| 2090 | dataset_path=args.dataset_path, |
| 2091 | disable_shuffle=args.disable_shuffle, |
| 2092 | random_seed=args.seed, |
| 2093 | ) |
| 2094 | input_requests = dataset.sample( |
| 2095 | num_requests=args.num_prompts, |
| 2096 | tokenizer=tokenizer, |
| 2097 | output_len=args.custom_output_len, |
| 2098 | skip_chat_template=args.skip_chat_template, |
| 2099 | chat_template_kwargs=getattr(args, "chat_template_kwargs", None), |
| 2100 | request_id_prefix=args.request_id_prefix, |
| 2101 | no_oversample=args.no_oversample, |
| 2102 | ) |
| 2103 | |
| 2104 | elif args.dataset_name == "custom_image": |
| 2105 | dataset = CustomImageDataset( |
| 2106 | dataset_path=args.dataset_path, |
| 2107 | disable_shuffle=args.disable_shuffle, |
| 2108 | random_seed=args.seed, |
| 2109 | ) |
| 2110 | input_requests = dataset.sample( |
| 2111 | num_requests=args.num_prompts, |
| 2112 | tokenizer=tokenizer, |
| 2113 | output_len=args.custom_output_len, |
| 2114 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2115 | ensure_client_side_data=getattr( |
| 2116 | args, "custom_ensure_client_side_data", False |
| 2117 | ), |
| 2118 | request_id_prefix=args.request_id_prefix, |
| 2119 | no_oversample=args.no_oversample, |
| 2120 | ) |
| 2121 | |
| 2122 | elif args.dataset_name == "custom_audio": |
| 2123 | dataset = CustomAudioDataset( |
| 2124 | dataset_path=args.dataset_path, |
| 2125 | disable_shuffle=args.disable_shuffle, |
| 2126 | random_seed=args.seed, |
| 2127 | ) |
| 2128 | input_requests = dataset.sample( |
| 2129 | num_requests=args.num_prompts, |
| 2130 | tokenizer=tokenizer, |
| 2131 | output_len=args.custom_output_len, |
| 2132 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2133 | request_id_prefix=args.request_id_prefix, |
| 2134 | no_oversample=args.no_oversample, |
| 2135 | ) |
| 2136 | |
| 2137 | elif args.dataset_name == "sonnet": |
| 2138 | assert tokenizer is not None, ( |
| 2139 | "Tokenizer must be initialized for the 'sonnet' dataset." |
| 2140 | ) |
| 2141 | sonnet_dataset = SonnetDataset( |
| 2142 | dataset_path=args.dataset_path, disable_shuffle=args.disable_shuffle |
| 2143 | ) |
| 2144 | # For the "sonnet" dataset, formatting depends on the backend. |
| 2145 | # Chat-style backends leave templating to the server; completions |
| 2146 | # backends need the prompt rendered client side. |
| 2147 | if args.backend in ("openai-chat", "openai-responses"): |
| 2148 | input_requests = sonnet_dataset.sample( |
| 2149 | num_requests=args.num_prompts, |
| 2150 | input_len=args.sonnet_input_len, |
| 2151 | output_len=args.sonnet_output_len, |
| 2152 | prefix_len=args.sonnet_prefix_len, |
| 2153 | tokenizer=tokenizer, |
| 2154 | return_prompt_formatted=False, |
| 2155 | request_id_prefix=args.request_id_prefix, |
| 2156 | no_oversample=args.no_oversample, |
| 2157 | ) |
| 2158 | else: |
| 2159 | assert ( |
| 2160 | hasattr(tokenizer, "chat_template") and tokenizer.chat_template |
| 2161 | ) or ( |
| 2162 | hasattr(tokenizer, "default_chat_template") |
| 2163 | and tokenizer.default_chat_template |
| 2164 | ), "Tokenizer/model must have chat template for sonnet dataset." |
| 2165 | input_requests = sonnet_dataset.sample( |
| 2166 | num_requests=args.num_prompts, |
| 2167 | input_len=args.sonnet_input_len, |
| 2168 | output_len=args.sonnet_output_len, |
| 2169 | prefix_len=args.sonnet_prefix_len, |
| 2170 | tokenizer=tokenizer, |
| 2171 | return_prompt_formatted=True, |
| 2172 | request_id_prefix=args.request_id_prefix, |
| 2173 | no_oversample=args.no_oversample, |
| 2174 | ) |
| 2175 | |
| 2176 | elif args.dataset_name == "hf": |
| 2177 | # all following datasets are implemented from the |
| 2178 | # HuggingFaceDataset base class |
| 2179 | hf_kwargs = {} |
| 2180 | dataset_class: type[BenchmarkDataset] |
| 2181 | if ( |
| 2182 | args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS |
| 2183 | or args.hf_name in VisionArenaDataset.SUPPORTED_DATASET_PATHS |
| 2184 | ): |
| 2185 | dataset_class = VisionArenaDataset |
| 2186 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2187 | args.hf_subset = None |
| 2188 | elif ( |
| 2189 | args.dataset_path in MMVUDataset.SUPPORTED_DATASET_PATHS |
| 2190 | or args.hf_name in MMVUDataset.SUPPORTED_DATASET_PATHS |
| 2191 | ): |
| 2192 | dataset_class = MMVUDataset |
| 2193 | args.hf_split = args.hf_split if args.hf_split else "validation" |
| 2194 | args.hf_subset = None |
| 2195 | elif ( |
| 2196 | args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS |
| 2197 | or args.hf_name in InstructCoderDataset.SUPPORTED_DATASET_PATHS |
| 2198 | ): |
| 2199 | dataset_class = InstructCoderDataset |
| 2200 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2201 | elif ( |
| 2202 | args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS |
| 2203 | or args.hf_name in MTBenchDataset.SUPPORTED_DATASET_PATHS |
| 2204 | ): |
| 2205 | dataset_class = MTBenchDataset |
| 2206 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2207 | elif ( |
| 2208 | args.dataset_path in HumanEvalDataset.SUPPORTED_DATASET_PATHS |
| 2209 | or args.hf_name in HumanEvalDataset.SUPPORTED_DATASET_PATHS |
| 2210 | ): |
| 2211 | dataset_class = HumanEvalDataset |
| 2212 | args.hf_split = args.hf_split if args.hf_split else "test" |
| 2213 | elif ( |
| 2214 | args.dataset_path in GSM8KDataset.SUPPORTED_DATASET_PATHS |
| 2215 | or args.hf_name in GSM8KDataset.SUPPORTED_DATASET_PATHS |
| 2216 | ): |
| 2217 | dataset_class = GSM8KDataset |
| 2218 | args.hf_subset = args.hf_subset if args.hf_subset else "main" |
| 2219 | args.hf_split = args.hf_split if args.hf_split else "test" |
| 2220 | elif ( |
| 2221 | args.dataset_path in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS |
| 2222 | or args.hf_name in MultiModalConversationDataset.SUPPORTED_DATASET_PATHS |
| 2223 | ): |
| 2224 | dataset_class = MultiModalConversationDataset |
| 2225 | elif ( |
| 2226 | args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS |
| 2227 | or args.hf_name in ConversationDataset.SUPPORTED_DATASET_PATHS |
| 2228 | ): |
| 2229 | dataset_class = ConversationDataset |
| 2230 | elif ( |
| 2231 | args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS |
| 2232 | or args.hf_name in AIMODataset.SUPPORTED_DATASET_PATHS |
| 2233 | ): |
| 2234 | dataset_class = AIMODataset |
| 2235 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2236 | elif ( |
| 2237 | args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS # noqa: E501 |
| 2238 | or args.hf_name in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS |
| 2239 | ): |
| 2240 | dataset_class = NextEditPredictionDataset |
| 2241 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2242 | elif ( |
| 2243 | args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS |
| 2244 | or args.hf_name in ASRDataset.SUPPORTED_DATASET_PATHS |
| 2245 | ): |
| 2246 | dataset_class = ASRDataset |
| 2247 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2248 | hf_kwargs = { |
| 2249 | "asr_min_audio_len_sec": args.asr_min_audio_len_sec, |
| 2250 | "asr_max_audio_len_sec": args.asr_max_audio_len_sec, |
| 2251 | } |
| 2252 | elif args.dataset_path in BlazeditDataset.SUPPORTED_DATASET_PATHS: |
| 2253 | dataset_class = BlazeditDataset |
| 2254 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2255 | hf_kwargs = { |
| 2256 | "min_distance": args.blazedit_min_distance, |
| 2257 | "max_distance": args.blazedit_max_distance, |
| 2258 | } |
| 2259 | elif ( |
| 2260 | args.dataset_path in MLPerfDataset.SUPPORTED_DATASET_PATHS |
| 2261 | or args.hf_name in MLPerfDataset.SUPPORTED_DATASET_PATHS |
| 2262 | ): |
| 2263 | dataset_class = MLPerfDataset |
| 2264 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2265 | elif ( |
| 2266 | args.dataset_path in MMStarDataset.SUPPORTED_DATASET_PATHS |
| 2267 | or args.hf_name in MMStarDataset.SUPPORTED_DATASET_PATHS |
| 2268 | ): |
| 2269 | dataset_class = MMStarDataset |
| 2270 | args.hf_split = args.hf_split if args.hf_split else "val" |
| 2271 | args.hf_subset = None |
| 2272 | elif ( |
| 2273 | args.dataset_path in BFCLDataset.SUPPORTED_DATASET_PATHS |
| 2274 | or args.hf_name in BFCLDataset.SUPPORTED_DATASET_PATHS |
| 2275 | ): |
| 2276 | if args.backend != "openai-chat": |
| 2277 | raise ValueError( |
| 2278 | "BFCL dataset requires the 'openai-chat' backend because " |
| 2279 | "it sends per-request tool schemas via chat completions." |
| 2280 | ) |
| 2281 | dataset_class = BFCLDataset |
| 2282 | # BFCL does not use HF splits/subsets; stub values for base init. |
| 2283 | args.hf_split = args.hf_split if args.hf_split else "train" |
| 2284 | args.hf_subset = None |
| 2285 | hf_kwargs = {"categories": args.bfcl_categories} |
| 2286 | else: |
| 2287 | supported_datasets = set( |
| 2288 | [ |
| 2289 | dataset_name |
| 2290 | for cls in HuggingFaceDataset.__subclasses__() |
| 2291 | for dataset_name in cls.SUPPORTED_DATASET_PATHS |
| 2292 | ] |
| 2293 | ) |
| 2294 | raise ValueError( |
| 2295 | f"Unsupported dataset path: {args.dataset_path}. " |
| 2296 | "Huggingface dataset only supports dataset_path" |
| 2297 | f" from one of following: {supported_datasets}. " |
| 2298 | "Please consider contributing if you would " |
| 2299 | "like to add support for additional dataset formats." |
| 2300 | ) |
| 2301 | |
| 2302 | if ( |
| 2303 | dataset_class.IS_MULTIMODAL |
| 2304 | and args.backend not in multimodal_backends |
| 2305 | and "embeddings-" not in args.backend |
| 2306 | ): |
| 2307 | # multi-modal benchmark is only available on chat-style backends; |
| 2308 | # which ones are allowed is caller-controlled (serve uses the |
| 2309 | # OpenAI endpoints, throughput uses vllm-chat). |
| 2310 | raise ValueError( |
| 2311 | f"Multi-modal content is not supported on backend " |
| 2312 | f"{args.backend!r}; use one of {sorted(multimodal_backends)}." |
| 2313 | ) |
| 2314 | assert tokenizer is not None, ( |
| 2315 | "Tokenizer must be initialized for the 'hf' dataset." |
| 2316 | ) |
| 2317 | input_requests = dataset_class( |
| 2318 | dataset_path=args.dataset_path, |
| 2319 | dataset_subset=args.hf_subset, |
| 2320 | dataset_split=args.hf_split, |
| 2321 | random_seed=args.seed, |
| 2322 | no_stream=args.no_stream, |
| 2323 | hf_name=args.hf_name, |
| 2324 | disable_shuffle=args.disable_shuffle, |
| 2325 | trust_remote_code=args.trust_remote_code, |
| 2326 | ).sample( |
| 2327 | num_requests=args.num_prompts, |
| 2328 | tokenizer=tokenizer, |
| 2329 | output_len=args.hf_output_len, |
| 2330 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2331 | request_id_prefix=args.request_id_prefix, |
| 2332 | no_oversample=args.no_oversample, |
| 2333 | skip_chat_template=args.skip_chat_template, |
| 2334 | **hf_kwargs, |
| 2335 | ) |
| 2336 | |
| 2337 | elif args.dataset_name == "timed_trace": |
| 2338 | assert tokenizer is not None, ( |
| 2339 | "Tokenizer must be initialized for the 'timed_trace' dataset." |
| 2340 | ) |
| 2341 | dataloader = TimedTrace(**vars(args)) |
| 2342 | input_requests = dataloader.sample( |
| 2343 | num_requests=args.num_prompts, |
| 2344 | tokenizer=tokenizer, |
| 2345 | request_id_prefix=args.request_id_prefix, |
| 2346 | ) |
| 2347 | |
| 2348 | else: |
| 2349 | # For datasets that follow a similar structure, use a mapping. |
| 2350 | assert tokenizer is not None, ( |
| 2351 | f"Tokenizer must be initialized for the '{args.dataset_name}' dataset." |
| 2352 | ) |
| 2353 | dataset_mapping = { |
| 2354 | "spec_bench": lambda: SpecBench( |
| 2355 | dataset_path=args.dataset_path, |
| 2356 | category=args.spec_bench_category, |
| 2357 | disable_shuffle=args.disable_shuffle, |
| 2358 | ).sample( |
| 2359 | num_requests=args.num_prompts, |
| 2360 | tokenizer=tokenizer, |
| 2361 | output_len=args.spec_bench_output_len, |
| 2362 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2363 | request_id_prefix=args.request_id_prefix, |
| 2364 | no_oversample=args.no_oversample, |
| 2365 | ), |
| 2366 | "sharegpt": lambda: ShareGPTDataset( |
| 2367 | random_seed=args.seed, |
| 2368 | dataset_path=args.dataset_path, |
| 2369 | disable_shuffle=args.disable_shuffle, |
| 2370 | ).sample( |
| 2371 | tokenizer=tokenizer, |
| 2372 | num_requests=args.num_prompts, |
| 2373 | output_len=args.sharegpt_output_len, |
| 2374 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2375 | request_id_prefix=args.request_id_prefix, |
| 2376 | no_oversample=args.no_oversample, |
| 2377 | ), |
| 2378 | "burstgpt": lambda: BurstGPTDataset( |
| 2379 | random_seed=args.seed, |
| 2380 | dataset_path=args.dataset_path, |
| 2381 | disable_shuffle=args.disable_shuffle, |
| 2382 | ).sample( |
| 2383 | tokenizer=tokenizer, |
| 2384 | num_requests=args.num_prompts, |
| 2385 | request_id_prefix=args.request_id_prefix, |
| 2386 | no_oversample=args.no_oversample, |
| 2387 | ), |
| 2388 | "random": lambda: RandomDataset( |
| 2389 | random_seed=args.seed, |
| 2390 | dataset_path=args.dataset_path, |
| 2391 | disable_shuffle=args.disable_shuffle, |
| 2392 | ).sample( |
| 2393 | tokenizer=tokenizer, |
| 2394 | num_requests=args.num_prompts, |
| 2395 | prefix_len=args.random_prefix_len, |
| 2396 | input_len=args.random_input_len, |
| 2397 | output_len=args.random_output_len, |
| 2398 | range_ratio=args.random_range_ratio, |
| 2399 | request_id_prefix=args.request_id_prefix, |
| 2400 | batchsize=args.random_batch_size, |
| 2401 | no_oversample=args.no_oversample, |
| 2402 | ), |
| 2403 | "random-mm": lambda: RandomMultiModalDataset( |
| 2404 | random_seed=args.seed, |
| 2405 | dataset_path=args.dataset_path, |
| 2406 | disable_shuffle=args.disable_shuffle, |
| 2407 | ).sample( |
| 2408 | tokenizer=tokenizer, |
| 2409 | num_requests=args.num_prompts, |
| 2410 | prefix_len=args.random_prefix_len, |
| 2411 | range_ratio=args.random_range_ratio, |
| 2412 | input_len=args.random_input_len, |
| 2413 | output_len=args.random_output_len, |
| 2414 | base_items_per_request=args.random_mm_base_items_per_request, |
| 2415 | limit_mm_per_prompt=args.random_mm_limit_mm_per_prompt, |
| 2416 | num_mm_items_range_ratio=args.random_mm_num_mm_items_range_ratio, |
| 2417 | bucket_config=args.random_mm_bucket_config, |
| 2418 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2419 | request_id_prefix=args.request_id_prefix, |
| 2420 | no_oversample=args.no_oversample, |
| 2421 | ), |
| 2422 | "random-rerank": lambda: RandomDatasetForReranking( |
| 2423 | random_seed=args.seed, |
| 2424 | dataset_path=args.dataset_path, |
| 2425 | disable_shuffle=args.disable_shuffle, |
| 2426 | ).sample( |
| 2427 | tokenizer=tokenizer, |
| 2428 | num_requests=args.num_prompts, |
| 2429 | input_len=args.random_input_len, |
| 2430 | range_ratio=args.random_range_ratio, |
| 2431 | request_id_prefix=args.request_id_prefix, |
| 2432 | batchsize=args.random_batch_size, |
| 2433 | is_reranker=not args.no_reranker, |
| 2434 | ), |
| 2435 | "prefix_repetition": lambda: PrefixRepetitionRandomDataset( |
| 2436 | random_seed=args.seed, |
| 2437 | dataset_path=args.dataset_path, |
| 2438 | disable_shuffle=args.disable_shuffle, |
| 2439 | ).sample( |
| 2440 | tokenizer=tokenizer, |
| 2441 | num_requests=args.num_prompts, |
| 2442 | prefix_len=args.prefix_repetition_prefix_len, |
| 2443 | suffix_len=args.prefix_repetition_suffix_len, |
| 2444 | num_prefixes=args.prefix_repetition_num_prefixes, |
| 2445 | output_len=args.prefix_repetition_output_len, |
| 2446 | request_id_prefix=args.request_id_prefix, |
| 2447 | no_oversample=args.no_oversample, |
| 2448 | ), |
| 2449 | "speed_bench": lambda: SpeedBench( |
| 2450 | dataset_path=args.dataset_path, |
| 2451 | dataset_subset=args.speed_bench_dataset_subset, |
| 2452 | category=args.speed_bench_category, |
| 2453 | disable_shuffle=args.disable_shuffle, |
| 2454 | ).sample( |
| 2455 | num_requests=args.num_prompts, |
| 2456 | tokenizer=tokenizer, |
| 2457 | output_len=args.speed_bench_output_len, |
| 2458 | skip_chat_template=args.skip_chat_template, |
| 2459 | chat_template_kwargs=getattr(args, "chat_template_kwargs", None), |
| 2460 | enable_multimodal_chat=args.enable_multimodal_chat, |
| 2461 | request_id_prefix=args.request_id_prefix, |
| 2462 | no_oversample=args.no_oversample, |
| 2463 | ), |
| 2464 | } |
| 2465 | |
| 2466 | try: |
| 2467 | # Enforce endpoint compatibility for multimodal datasets. |
| 2468 | if ( |
| 2469 | args.dataset_name == "random-mm" |
| 2470 | and args.backend not in multimodal_backends |
| 2471 | ): |
| 2472 | raise ValueError( |
| 2473 | f"Multi-modal content (images) is not supported on backend " |
| 2474 | f"{args.backend!r}; use one of {sorted(multimodal_backends)}." |
| 2475 | ) |
| 2476 | input_requests = dataset_mapping[args.dataset_name]() |
| 2477 | except KeyError as err: |
| 2478 | raise ValueError(f"Unknown dataset: {args.dataset_name}") from err |
| 2479 | |
| 2480 | return input_requests |
| 2481 | |
| 2482 | |
| 2483 | # ----------------------------------------------------------------------------- |
| 2484 | # Custom Dataset Implementation |
| 2485 | # ----------------------------------------------------------------------------- |
| 2486 | |
| 2487 | |
| 2488 | class CustomDataset(BenchmarkDataset): |
| 2489 | """Implements the Custom dataset. Loads data from a JSONL file and generates |
| 2490 | sample requests based on conversation turns. E.g., |
| 2491 | ``` |
| 2492 | {"prompt": "What is the capital of India?", "output_tokens": 10} |
| 2493 | {"prompt": "What is the capital of Iran?", "output_tokens": 1520} |
| 2494 | {"prompt": "What is the capital of China?", "output_tokens": 819} |
| 2495 | ``` |
| 2496 | Note that 'output_tokens' column is optional and has to be provided only if |
| 2497 | 'custom-output-len' argument is None or -1. |
| 2498 | """ |
| 2499 | |
| 2500 | def __init__(self, **kwargs) -> None: |
| 2501 | super().__init__(**kwargs) |
| 2502 | self.load_data() |
| 2503 | |
| 2504 | def load_data(self) -> None: |
| 2505 | if self.dataset_path is None: |
| 2506 | raise ValueError("dataset_path must be provided for loading data.") |
| 2507 | |
| 2508 | # self.data will be a list of dictionaries |
| 2509 | # e.g., [{"prompt": "What is the capital of India?"}, ...] |
| 2510 | # This will be the standardized format which load_data() |
| 2511 | # has to convert into depending on the filetype of dataset_path. |
| 2512 | # sample() will assume this standardized format of self.data |
| 2513 | self.data: list[dict] = [] |
| 2514 | |
| 2515 | # Load the JSONL file |
| 2516 | if self.dataset_path.endswith(".jsonl"): |
| 2517 | jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True) |
| 2518 | |
| 2519 | # check if the JSONL file has a 'prompt' column |
| 2520 | if "prompt" not in jsonl_data.columns: |
| 2521 | raise ValueError("JSONL file must contain a 'prompt' column.") |
| 2522 | |
| 2523 | # Convert each row to a dictionary and append to self.data |
| 2524 | # This will convert the DataFrame to a list of dictionaries |
| 2525 | # where each dictionary corresponds to a row in the DataFrame. |
| 2526 | # This is the standardized format we want for self.data |
| 2527 | for _, row in jsonl_data.iterrows(): |
| 2528 | self.data.append(row.to_dict()) |
| 2529 | else: |
| 2530 | raise NotImplementedError( |
| 2531 | "Only JSONL format is supported for CustomDataset." |
| 2532 | ) |
| 2533 | |
| 2534 | random.seed(self.random_seed) |
| 2535 | if not getattr(self, "disable_shuffle", False): |
| 2536 | random.shuffle(self.data) |
| 2537 | |
| 2538 | def sample( |
| 2539 | self, |
| 2540 | tokenizer: TokenizerLike | None, |
| 2541 | num_requests: int, |
| 2542 | request_id_prefix: str = "", |
| 2543 | no_oversample: bool = False, |
| 2544 | lora_path: str | None = None, |
| 2545 | max_loras: int | None = None, |
| 2546 | output_len: int | None = None, |
| 2547 | enable_multimodal_chat: bool = False, |
| 2548 | skip_chat_template: bool = False, |
| 2549 | chat_template_kwargs: dict | None = None, |
| 2550 | **kwargs, |
| 2551 | ) -> list[SampleRequest]: |
| 2552 | # load all data if needed |
| 2553 | self.num_available_samples = len(self.data) |
| 2554 | if num_requests <= 0: |
| 2555 | num_requests = self.num_available_samples |
| 2556 | logger.info( |
| 2557 | "num_requests is set to 0 or negative, " |
| 2558 | "so using all available samples: %d", |
| 2559 | num_requests, |
| 2560 | ) |
| 2561 | |
| 2562 | sampled_requests: list[SampleRequest] = [] |
| 2563 | for i, item in enumerate(self.data): |
| 2564 | if len(sampled_requests) >= num_requests: |
| 2565 | break |
| 2566 | prompt = item["prompt"] |
| 2567 | |
| 2568 | if tokenizer is None: |
| 2569 | new_output_len = 1 |
| 2570 | elif output_len is None or output_len == -1: |
| 2571 | # check that the request has an 'output_tokens' field |
| 2572 | if "output_tokens" not in item: |
| 2573 | raise ValueError( |
| 2574 | "If no output length is provided the " |
| 2575 | "custom dataset must contain an 'output_tokens' field." |
| 2576 | ) |
| 2577 | # Use number of output tokens from the request data |
| 2578 | try: |
| 2579 | new_output_len = int(item["output_tokens"]) |
| 2580 | except (ValueError, TypeError) as e: |
| 2581 | raise ValueError( |
| 2582 | f"Invalid value for 'output_tokens' in custom dataset: " |
| 2583 | f"'{item['output_tokens']}'. Must be an integer." |
| 2584 | ) from e |
| 2585 | else: |
| 2586 | new_output_len = output_len |
| 2587 | |
| 2588 | if tokenizer is None: |
| 2589 | prompt_len = 1 |
| 2590 | else: |
| 2591 | # apply template |
| 2592 | if not skip_chat_template: |
| 2593 | prompt = tokenizer.apply_chat_template( |
| 2594 | [{"role": "user", "content": prompt}], |
| 2595 | add_generation_prompt=True, |
| 2596 | tokenize=False, |
| 2597 | **(chat_template_kwargs or {}), |
| 2598 | ) |
| 2599 | |
| 2600 | prompt_len = len(tokenizer(prompt).input_ids) |
| 2601 | sampled_requests.append( |
| 2602 | SampleRequest( |
| 2603 | prompt=prompt, |
| 2604 | prompt_len=prompt_len, |
| 2605 | expected_output_len=new_output_len, |
| 2606 | request_id=request_id_prefix + str(i), |
| 2607 | ) |
| 2608 | ) |
| 2609 | self.maybe_oversample_requests( |
| 2610 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 2611 | ) |
| 2612 | |
| 2613 | return sampled_requests |
| 2614 | |
| 2615 | |
| 2616 | class CustomImageDataset(CustomDataset): |
| 2617 | """Implements the Custom image dataset. Loads data from a JSONL file and generates |
| 2618 | sample requests based on conversation turns. E.g., |
| 2619 | ``` |
| 2620 | { |
| 2621 | "prompt": "How many red blocks in the given images?", |
| 2622 | "image_files": ["path/to/image1.png", "path/to/image2.png"], |
| 2623 | } |
| 2624 | { |
| 2625 | "prompt": "Which country has the most pokemons based on the given graphs?", |
| 2626 | "image_files": ["path/to/image.png"], |
| 2627 | } |
| 2628 | { |
| 2629 | "content": [ |
| 2630 | {"type": "text", "text": "Compare these images: "}, |
| 2631 | {"type": "image", "image": "path/to/image1.png"}, |
| 2632 | {"type": "text", "text": " and "}, |
| 2633 | {"type": "image_url", "image_url": {"url": "path/to/image2.png"}}, |
| 2634 | ], |
| 2635 | } |
| 2636 | ``` |
| 2637 | |
| 2638 | This is used to benchmark multimodal LLMs on arbitrary datasets. |
| 2639 | """ |
| 2640 | |
| 2641 | IS_MULTIMODAL = True |
| 2642 | |
| 2643 | def load_data(self) -> None: |
| 2644 | if self.dataset_path is None: |
| 2645 | raise ValueError("dataset_path must be provided for loading data.") |
| 2646 | |
| 2647 | self.data: list[dict] = [] |
| 2648 | |
| 2649 | if not self.dataset_path.endswith(".jsonl"): |
| 2650 | raise NotImplementedError( |
| 2651 | "Only JSONL format is supported for CustomImageDataset." |
| 2652 | ) |
| 2653 | |
| 2654 | with open(self.dataset_path, encoding="utf-8") as f: |
| 2655 | for line_number, line in enumerate(f, start=1): |
| 2656 | line = line.strip() |
| 2657 | if not line: |
| 2658 | continue |
| 2659 | |
| 2660 | try: |
| 2661 | item = json.loads(line) |
| 2662 | except json.JSONDecodeError as e: |
| 2663 | raise ValueError( |
| 2664 | f"Invalid JSON in custom image dataset line {line_number}: {e}" |
| 2665 | ) from e |
| 2666 | |
| 2667 | if not isinstance(item, dict): |
| 2668 | raise ValueError( |
| 2669 | "Each custom image dataset line must contain a JSON object. " |
| 2670 | f"Found {type(item)} on line {line_number}." |
| 2671 | ) |
| 2672 | |
| 2673 | has_legacy_fields = "prompt" in item and "image_files" in item |
| 2674 | has_interleaved_content = "content" in item |
| 2675 | if not has_legacy_fields and not has_interleaved_content: |
| 2676 | raise ValueError( |
| 2677 | "Each custom image dataset line must contain either " |
| 2678 | "'prompt' and 'image_files' fields, or a 'content' field. " |
| 2679 | f"Invalid line: {line_number}." |
| 2680 | ) |
| 2681 | |
| 2682 | self.data.append(item) |
| 2683 | |
| 2684 | random.seed(self.random_seed) |
| 2685 | if not getattr(self, "disable_shuffle", False): |
| 2686 | random.shuffle(self.data) |
| 2687 | |
| 2688 | @staticmethod |
| 2689 | def _validate_content_parts(content: Any) -> list[dict[str, Any]]: |
| 2690 | if not isinstance(content, list): |
| 2691 | raise ValueError( |
| 2692 | "'content' must be a list of text and image content dictionaries." |
| 2693 | ) |
| 2694 | |
| 2695 | if not content: |
| 2696 | raise ValueError("'content' must contain at least one item.") |
| 2697 | |
| 2698 | parts: list[dict[str, Any]] = [] |
| 2699 | for part in content: |
| 2700 | if not isinstance(part, dict): |
| 2701 | raise ValueError( |
| 2702 | f"Each item in 'content' must be a dictionary. Found {type(part)}." |
| 2703 | ) |
| 2704 | parts.append(part) |
| 2705 | |
| 2706 | return parts |
| 2707 | |
| 2708 | @classmethod |
| 2709 | def _process_content_part( |
| 2710 | cls, |
| 2711 | part: dict[str, Any], |
| 2712 | *, |
| 2713 | ensure_client_side_data: bool = False, |
| 2714 | ) -> dict[str, Any]: |
| 2715 | content_type = part.get("type") |
| 2716 | if content_type == "text": |
| 2717 | text = part.get("text") |
| 2718 | if not isinstance(text, str): |
| 2719 | raise ValueError("Text content parts must contain a string 'text'.") |
| 2720 | return {"type": "text", "text": text} |
| 2721 | |
| 2722 | if content_type == "image": |
| 2723 | if "image" not in part: |
| 2724 | raise ValueError("Image content parts must contain an 'image' field.") |
| 2725 | return dict( |
| 2726 | process_image( |
| 2727 | part["image"], |
| 2728 | ensure_client_side_data=ensure_client_side_data, |
| 2729 | ) |
| 2730 | ) |
| 2731 | |
| 2732 | if content_type == "image_url": |
| 2733 | image_url = part.get("image_url") |
| 2734 | if isinstance(image_url, str): |
| 2735 | return dict( |
| 2736 | process_image( |
| 2737 | image_url, |
| 2738 | ensure_client_side_data=ensure_client_side_data, |
| 2739 | ) |
| 2740 | ) |
| 2741 | |
| 2742 | if isinstance(image_url, dict): |
| 2743 | url = image_url.get("url") |
| 2744 | if not isinstance(url, str): |
| 2745 | raise ValueError( |
| 2746 | "Image URL content parts must contain a string 'image_url.url'." |
| 2747 | ) |
| 2748 | |
| 2749 | processed_part = dict( |
| 2750 | process_image( |
| 2751 | url, |
| 2752 | ensure_client_side_data=ensure_client_side_data, |
| 2753 | ) |
| 2754 | ) |
| 2755 | processed_image_url = dict(processed_part["image_url"]) |
| 2756 | processed_image_url.update( |
| 2757 | {key: value for key, value in image_url.items() if key != "url"} |
| 2758 | ) |
| 2759 | processed_part["image_url"] = processed_image_url |
| 2760 | return processed_part |
| 2761 | |
| 2762 | raise ValueError( |
| 2763 | "Image URL content parts must contain an 'image_url' string " |
| 2764 | "or dictionary." |
| 2765 | ) |
| 2766 | |
| 2767 | raise ValueError( |
| 2768 | "Content parts must have type 'text', 'image', or 'image_url'. " |
| 2769 | f"Found: {content_type!r}." |
| 2770 | ) |
| 2771 | |
| 2772 | @classmethod |
| 2773 | def _process_interleaved_content( |
| 2774 | cls, |
| 2775 | content: Any, |
| 2776 | *, |
| 2777 | ensure_client_side_data: bool = False, |
| 2778 | ) -> list[dict[str, Any]]: |
| 2779 | return [ |
| 2780 | cls._process_content_part( |
| 2781 | part, |
| 2782 | ensure_client_side_data=ensure_client_side_data, |
| 2783 | ) |
| 2784 | for part in cls._validate_content_parts(content) |
| 2785 | ] |
| 2786 | |
| 2787 | @staticmethod |
| 2788 | def _get_text_from_content(content: list[dict[str, Any]]) -> str: |
| 2789 | return "".join(part["text"] for part in content if part.get("type") == "text") |
| 2790 | |
| 2791 | @staticmethod |
| 2792 | def _process_image_files( |
| 2793 | images: Any, |
| 2794 | *, |
| 2795 | ensure_client_side_data: bool = False, |
| 2796 | ) -> dict[str, Any] | list[dict[str, Any]]: |
| 2797 | if not isinstance(images, list) or not images: |
| 2798 | raise ValueError("'image_files' must be a non-empty list.") |
| 2799 | |
| 2800 | mm_content = [ |
| 2801 | dict( |
| 2802 | process_image( |
| 2803 | image, |
| 2804 | ensure_client_side_data=ensure_client_side_data, |
| 2805 | ) |
| 2806 | ) |
| 2807 | for image in images |
| 2808 | ] |
| 2809 | if len(mm_content) == 1: |
| 2810 | return mm_content[0] |
| 2811 | |
| 2812 | return mm_content |
| 2813 | |
| 2814 | def sample( |
| 2815 | self, |
| 2816 | tokenizer: TokenizerLike | None, |
| 2817 | num_requests: int, |
| 2818 | request_id_prefix: str = "", |
| 2819 | no_oversample: bool = False, |
| 2820 | lora_path: str | None = None, |
| 2821 | max_loras: int | None = None, |
| 2822 | output_len: int | None = None, |
| 2823 | enable_multimodal_chat: bool = False, |
| 2824 | skip_chat_template: bool = False, |
| 2825 | chat_template_kwargs: dict | None = None, |
| 2826 | ensure_client_side_data: bool = False, |
| 2827 | **kwargs, |
| 2828 | ) -> list[SampleRequest]: |
| 2829 | # load all data if needed |
| 2830 | self.num_available_samples = len(self.data) |
| 2831 | if num_requests <= 0: |
| 2832 | num_requests = self.num_available_samples |
| 2833 | logger.info( |
| 2834 | "num_requests is set to 0 or negative, " |
| 2835 | "so using all available samples: %d", |
| 2836 | num_requests, |
| 2837 | ) |
| 2838 | |
| 2839 | sampled_requests: list[SampleRequest] = [] |
| 2840 | for i, item in enumerate(self.data): |
| 2841 | if len(sampled_requests) >= num_requests: |
| 2842 | break |
| 2843 | |
| 2844 | if "content" in item: |
| 2845 | content = self._process_interleaved_content( |
| 2846 | item["content"], |
| 2847 | ensure_client_side_data=ensure_client_side_data, |
| 2848 | ) |
| 2849 | text_prompt = self._get_text_from_content(content) |
| 2850 | prompt_len = ( |
| 2851 | 1 if tokenizer is None else len(tokenizer(text_prompt).input_ids) |
| 2852 | ) |
| 2853 | prompt = ( |
| 2854 | [{"role": "user", "content": content}] |
| 2855 | if enable_multimodal_chat |
| 2856 | else content |
| 2857 | ) |
| 2858 | sampled_requests.append( |
| 2859 | SampleRequest( |
| 2860 | prompt=prompt, |
| 2861 | prompt_len=prompt_len, |
| 2862 | expected_output_len=output_len or 0, |
| 2863 | multi_modal_data=None, |
| 2864 | request_id=request_id_prefix + str(i), |
| 2865 | ) |
| 2866 | ) |
| 2867 | continue |
| 2868 | |
| 2869 | prompt = item["prompt"] |
| 2870 | if not isinstance(prompt, str): |
| 2871 | raise ValueError("'prompt' must be a string.") |
| 2872 | |
| 2873 | prompt_len = 1 if tokenizer is None else len(tokenizer(prompt).input_ids) |
| 2874 | mm_content = self._process_image_files( |
| 2875 | item["image_files"], |
| 2876 | ensure_client_side_data=ensure_client_side_data, |
| 2877 | ) |
| 2878 | if enable_multimodal_chat: |
| 2879 | # Note: when chat is enabled the request prompt_len is no longer |
| 2880 | # accurate and we will be using request output to count the |
| 2881 | # actual prompt len |
| 2882 | prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) |
| 2883 | |
| 2884 | sampled_requests.append( |
| 2885 | SampleRequest( |
| 2886 | prompt=prompt, |
| 2887 | prompt_len=prompt_len, |
| 2888 | expected_output_len=output_len or 0, |
| 2889 | multi_modal_data=mm_content, |
| 2890 | request_id=request_id_prefix + str(i), |
| 2891 | ) |
| 2892 | ) |
| 2893 | self.maybe_oversample_requests( |
| 2894 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 2895 | ) |
| 2896 | |
| 2897 | return sampled_requests |
| 2898 | |
| 2899 | |
| 2900 | class CustomAudioDataset(CustomDataset): |
| 2901 | """Custom dataset for audio benchmarking. Loads data from a JSONL file. E.g., |
| 2902 | {"prompt": "Transcribe the audio.", "audio": "/path/to/audio.wav"} |
| 2903 | |
| 2904 | Supports both: |
| 2905 | - Dedicated ASR models (e.g. Whisper) via openai-audio & /v1/audio/transcriptions |
| 2906 | - Chat-based audio models (e.g. Qwen2-Audio) via openai-chat & /v1/chat/completions |
| 2907 | """ |
| 2908 | |
| 2909 | IS_MULTIMODAL = True |
| 2910 | |
| 2911 | def sample( |
| 2912 | self, |
| 2913 | tokenizer: TokenizerLike | None, |
| 2914 | num_requests: int, |
| 2915 | request_id_prefix: str = "", |
| 2916 | no_oversample: bool = False, |
| 2917 | lora_path: str | None = None, |
| 2918 | max_loras: int | None = None, |
| 2919 | output_len: int | None = None, |
| 2920 | enable_multimodal_chat: bool = False, |
| 2921 | skip_chat_template: bool = False, |
| 2922 | chat_template_kwargs: dict | None = None, |
| 2923 | **kwargs, |
| 2924 | ) -> list[SampleRequest]: |
| 2925 | self.num_available_samples = len(self.data) |
| 2926 | if num_requests <= 0: |
| 2927 | num_requests = self.num_available_samples |
| 2928 | sampled_requests: list[SampleRequest] = [] |
| 2929 | for i, item in enumerate(self.data): |
| 2930 | if len(sampled_requests) >= num_requests: |
| 2931 | break |
| 2932 | prompt = item.get("prompt", "") |
| 2933 | if tokenizer is None: |
| 2934 | prompt_len = 1 |
| 2935 | new_output_len = ( |
| 2936 | output_len if (output_len is not None and output_len != -1) else 256 |
| 2937 | ) |
| 2938 | mm_content = None |
| 2939 | else: |
| 2940 | use_chat_template = ( |
| 2941 | not skip_chat_template |
| 2942 | and hasattr(tokenizer, "chat_template") |
| 2943 | and tokenizer.chat_template is not None |
| 2944 | ) |
| 2945 | if enable_multimodal_chat: |
| 2946 | # Chat-based audio models (e.g., Qwen2-Audio): |
| 2947 | # encode audio as base64; serve.py assembles the chat message |
| 2948 | # as: {"role": "user", "content": [ |
| 2949 | # {"type": "text", "text": prompt}, |
| 2950 | # {"type": "input_audio", "input_audio": {...}} |
| 2951 | # ]} |
| 2952 | y, sr = process_audio(item["audio"]) |
| 2953 | buf = io.BytesIO() |
| 2954 | sf.write(buf, y, sr, format="WAV") |
| 2955 | audio_base64 = base64.b64encode(buf.getvalue()).decode("utf-8") |
| 2956 | mm_content = { |
| 2957 | "type": "input_audio", |
| 2958 | "input_audio": { |
| 2959 | "data": audio_base64, |
| 2960 | "format": "wav", |
| 2961 | }, |
| 2962 | } |
| 2963 | # prompt stays as plain string; serve.py handles wrapping |
| 2964 | else: |
| 2965 | # Whisper-style models: load audio array locally |
| 2966 | y, sr = process_audio(item["audio"]) |
| 2967 | mm_content = {"audio": (y, sr)} |
| 2968 | if use_chat_template: |
| 2969 | # ASR models with a chat template but not multimodal chat |
| 2970 | prompt = tokenizer.apply_chat_template( |
| 2971 | [{"role": "user", "content": prompt}], |
| 2972 | add_generation_prompt=True, |
| 2973 | tokenize=False, |
| 2974 | ) |
| 2975 | # else: plain prompt for Whisper-style models |
| 2976 | prompt_len = ( |
| 2977 | len(tokenizer(prompt).input_ids) if isinstance(prompt, str) else 1 |
| 2978 | ) |
| 2979 | if output_len is None or output_len == -1: |
| 2980 | if "output_tokens" not in item: |
| 2981 | raise ValueError( |
| 2982 | "If no output length is provided the " |
| 2983 | "custom dataset must contain an 'output_tokens' field." |
| 2984 | ) |
| 2985 | new_output_len = int(item["output_tokens"]) |
| 2986 | else: |
| 2987 | assert output_len is not None |
| 2988 | new_output_len = output_len |
| 2989 | sampled_requests.append( |
| 2990 | SampleRequest( |
| 2991 | prompt=prompt, |
| 2992 | prompt_len=prompt_len, |
| 2993 | expected_output_len=new_output_len, |
| 2994 | multi_modal_data=mm_content, |
| 2995 | request_id=request_id_prefix + str(i), |
| 2996 | ) |
| 2997 | ) |
| 2998 | self.maybe_oversample_requests( |
| 2999 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3000 | ) |
| 3001 | return sampled_requests |
| 3002 | |
| 3003 | |
| 3004 | # ----------------------------------------------------------------------------- |
| 3005 | # Spec Bench Dataset Implementation |
| 3006 | # ----------------------------------------------------------------------------- |
| 3007 | |
| 3008 | |
| 3009 | class SpecBench(CustomDataset): |
| 3010 | """Implements the SpecBench dataset: https://github.com/hemingkx/Spec-Bench |
| 3011 | Download the dataset using: |
| 3012 | wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl |
| 3013 | """ # noqa: E501 |
| 3014 | |
| 3015 | def __init__(self, **kwargs) -> None: |
| 3016 | self.category = kwargs.pop("category", None) |
| 3017 | super().__init__(**kwargs) |
| 3018 | self.load_data() |
| 3019 | |
| 3020 | def load_data(self) -> None: |
| 3021 | if self.dataset_path is None: |
| 3022 | raise ValueError("dataset_path must be provided for loading data.") |
| 3023 | |
| 3024 | self.data = [] |
| 3025 | |
| 3026 | # Load the JSONL file |
| 3027 | jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True) |
| 3028 | |
| 3029 | # check if the JSONL file has a 'turns' column |
| 3030 | if "turns" not in jsonl_data.columns: |
| 3031 | raise ValueError("JSONL file must contain a 'turns' column.") |
| 3032 | |
| 3033 | for _, row in jsonl_data.iterrows(): |
| 3034 | # sample only from a specific category if specified |
| 3035 | if (not self.category) or (self.category == row["category"]): |
| 3036 | prompt = row["turns"][0] |
| 3037 | self.data.append({"prompt": prompt}) |
| 3038 | |
| 3039 | random.seed(self.random_seed) |
| 3040 | if not getattr(self, "disable_shuffle", False): |
| 3041 | random.shuffle(self.data) |
| 3042 | |
| 3043 | def sample( |
| 3044 | self, |
| 3045 | tokenizer: TokenizerLike | None, |
| 3046 | num_requests: int, |
| 3047 | request_id_prefix: str = "", |
| 3048 | no_oversample: bool = False, |
| 3049 | lora_path: str | None = None, |
| 3050 | max_loras: int | None = None, |
| 3051 | output_len: int | None = None, |
| 3052 | enable_multimodal_chat: bool = False, |
| 3053 | skip_chat_template: bool = False, |
| 3054 | chat_template_kwargs: dict | None = None, |
| 3055 | **kwargs, |
| 3056 | ) -> list[SampleRequest]: |
| 3057 | # leverage CustomDataset sample |
| 3058 | return super().sample( |
| 3059 | tokenizer=tokenizer, |
| 3060 | num_requests=num_requests, |
| 3061 | request_id_prefix=request_id_prefix, |
| 3062 | no_oversample=no_oversample, |
| 3063 | lora_path=lora_path, |
| 3064 | max_loras=max_loras, |
| 3065 | output_len=output_len, |
| 3066 | enable_multimodal_chat=enable_multimodal_chat, |
| 3067 | skip_chat_template=skip_chat_template, |
| 3068 | chat_template_kwargs=chat_template_kwargs, |
| 3069 | **kwargs, |
| 3070 | ) |
| 3071 | |
| 3072 | |
| 3073 | # ----------------------------------------------------------------------------- |
| 3074 | # Sonnet Dataset Implementation |
| 3075 | # ----------------------------------------------------------------------------- |
| 3076 | |
| 3077 | |
| 3078 | @deprecated( |
| 3079 | "SonnetDataset is deprecated and will be removed in a future version.", |
| 3080 | ) |
| 3081 | class SonnetDataset(BenchmarkDataset): |
| 3082 | """Simplified implementation of the Sonnet dataset. Loads poem lines from a |
| 3083 | text file and generates sample requests. Default values here copied from |
| 3084 | `benchmark_serving.py` for the sonnet dataset. |
| 3085 | """ |
| 3086 | |
| 3087 | DEFAULT_PREFIX_LEN = 200 |
| 3088 | DEFAULT_INPUT_LEN = 550 |
| 3089 | DEFAULT_OUTPUT_LEN = 150 |
| 3090 | |
| 3091 | def __init__( |
| 3092 | self, |
| 3093 | **kwargs, |
| 3094 | ) -> None: |
| 3095 | super().__init__(**kwargs) |
| 3096 | self.load_data() |
| 3097 | |
| 3098 | def load_data(self) -> None: |
| 3099 | if not self.dataset_path: |
| 3100 | raise ValueError("dataset_path must be provided.") |
| 3101 | with open(self.dataset_path, encoding="utf-8") as f: |
| 3102 | self.data = f.readlines() |
| 3103 | |
| 3104 | def sample( |
| 3105 | self, |
| 3106 | tokenizer: TokenizerLike, |
| 3107 | num_requests: int, |
| 3108 | request_id_prefix: str = "", |
| 3109 | no_oversample: bool = False, |
| 3110 | prefix_len: int = DEFAULT_PREFIX_LEN, |
| 3111 | input_len: int = DEFAULT_INPUT_LEN, |
| 3112 | output_len: int = DEFAULT_OUTPUT_LEN, |
| 3113 | return_prompt_formatted: bool = False, |
| 3114 | **kwargs, |
| 3115 | ) -> list[SampleRequest]: |
| 3116 | poem_lines = self.data |
| 3117 | assert poem_lines is not None |
| 3118 | # Calculate average token length for a poem line. |
| 3119 | tokenized_lines = [tokenizer(line).input_ids for line in poem_lines] |
| 3120 | avg_len = sum(len(tokens) for tokens in tokenized_lines) / len(tokenized_lines) |
| 3121 | |
| 3122 | # Build the base prompt. |
| 3123 | base_prompt = "Pick as many lines as you can from these poem lines:\n" |
| 3124 | base_msg = [{"role": "user", "content": base_prompt}] |
| 3125 | base_fmt = tokenizer.apply_chat_template( |
| 3126 | base_msg, # type: ignore[arg-type] |
| 3127 | add_generation_prompt=True, |
| 3128 | tokenize=False, |
| 3129 | ) |
| 3130 | assert isinstance(base_fmt, str) |
| 3131 | base_offset = len(tokenizer(base_fmt).input_ids) |
| 3132 | if input_len <= base_offset: |
| 3133 | raise ValueError( |
| 3134 | f"'input_len' must be higher than the base prompt length " |
| 3135 | f"({base_offset})." |
| 3136 | ) |
| 3137 | |
| 3138 | # Determine how many poem lines to use. |
| 3139 | num_input_lines = max(round((input_len - base_offset) / avg_len), 1) |
| 3140 | num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0) |
| 3141 | prefix_lines = poem_lines[:num_prefix_lines] |
| 3142 | |
| 3143 | samples: list[SampleRequest] = [] |
| 3144 | ind = 0 |
| 3145 | while len(samples) < num_requests: |
| 3146 | extra_lines = random.choices( |
| 3147 | poem_lines, k=num_input_lines - num_prefix_lines |
| 3148 | ) |
| 3149 | prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}" |
| 3150 | msg = [{"role": "user", "content": prompt}] |
| 3151 | prompt_formatted = tokenizer.apply_chat_template( |
| 3152 | msg, # type: ignore[arg-type] |
| 3153 | add_generation_prompt=True, |
| 3154 | tokenize=False, |
| 3155 | ) |
| 3156 | assert isinstance(prompt_formatted, str) |
| 3157 | prompt_len = len(tokenizer(prompt_formatted).input_ids) |
| 3158 | if prompt_len <= input_len: |
| 3159 | samples.append( |
| 3160 | SampleRequest( |
| 3161 | prompt=( |
| 3162 | prompt_formatted if return_prompt_formatted else prompt |
| 3163 | ), |
| 3164 | prompt_len=prompt_len, |
| 3165 | expected_output_len=output_len, |
| 3166 | request_id=request_id_prefix + str(ind), |
| 3167 | ) |
| 3168 | ) |
| 3169 | ind += 1 |
| 3170 | return samples |
| 3171 | |
| 3172 | |
| 3173 | # ----------------------------------------------------------------------------- |
| 3174 | # BurstGPT Dataset Implementation |
| 3175 | # ----------------------------------------------------------------------------- |
| 3176 | |
| 3177 | |
| 3178 | class BurstGPTDataset(BenchmarkDataset): |
| 3179 | """Implements the BurstGPT dataset. Loads data from a CSV file and generates |
| 3180 | sample requests based on synthetic prompt generation. Only rows with Model |
| 3181 | "GPT-4" and positive response tokens are used. |
| 3182 | """ |
| 3183 | |
| 3184 | def __init__(self, **kwargs) -> None: |
| 3185 | super().__init__(**kwargs) |
| 3186 | self.load_data() |
| 3187 | |
| 3188 | def load_data( |
| 3189 | self, |
| 3190 | ): |
| 3191 | if self.dataset_path is None: |
| 3192 | raise ValueError("dataset_path must be provided for loading data.") |
| 3193 | |
| 3194 | df = pd.read_csv(self.dataset_path) |
| 3195 | # Filter to keep only GPT-4 rows. |
| 3196 | gpt4_df = df[df["Model"] == "GPT-4"] |
| 3197 | # Remove failed requests (where Response tokens is 0 or less). |
| 3198 | gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0] |
| 3199 | # Sample the desired number of rows. |
| 3200 | self.data = gpt4_df |
| 3201 | |
| 3202 | def _sample_loaded_data(self, num_requests: int) -> list: |
| 3203 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3204 | if num_requests <= len(self.data): |
| 3205 | data = self.data.sample(n=num_requests, random_state=self.random_seed) |
| 3206 | else: |
| 3207 | data = self.data.sample( |
| 3208 | n=num_requests, |
| 3209 | random_state=self.random_seed, |
| 3210 | replace=True, |
| 3211 | ) |
| 3212 | # Convert the dataframe to a list of lists. |
| 3213 | return data.values.tolist() |
| 3214 | |
| 3215 | def sample( |
| 3216 | self, |
| 3217 | tokenizer: TokenizerLike, |
| 3218 | num_requests: int, |
| 3219 | request_id_prefix: str = "", |
| 3220 | no_oversample: bool = False, |
| 3221 | lora_assignment: str = "random", |
| 3222 | max_loras: int | None = None, |
| 3223 | lora_path: str | None = None, |
| 3224 | **kwargs, |
| 3225 | ) -> list[SampleRequest]: |
| 3226 | samples = [] |
| 3227 | data = self._sample_loaded_data(num_requests=num_requests) |
| 3228 | for i in range(num_requests): |
| 3229 | input_len = int(data[i][2]) |
| 3230 | output_len = int(data[i][3]) |
| 3231 | lora_req = self.get_lora_request( |
| 3232 | index=i, |
| 3233 | max_loras=max_loras, |
| 3234 | lora_path=lora_path, |
| 3235 | lora_assignment=lora_assignment, |
| 3236 | ) |
| 3237 | vocab_size = tokenizer.vocab_size |
| 3238 | # Generate a synthetic prompt: a list of token IDs computed as (i + |
| 3239 | # j) modulo vocab_size. |
| 3240 | token_ids = [(i + j) % vocab_size for j in range(input_len)] |
| 3241 | prompt = tokenizer.decode(token_ids) |
| 3242 | samples.append( |
| 3243 | SampleRequest( |
| 3244 | prompt=prompt, |
| 3245 | prompt_len=input_len, |
| 3246 | expected_output_len=output_len, |
| 3247 | lora_request=lora_req, |
| 3248 | request_id=request_id_prefix + str(i), |
| 3249 | ) |
| 3250 | ) |
| 3251 | return samples |
| 3252 | |
| 3253 | |
| 3254 | # ----------------------------------------------------------------------------- |
| 3255 | # HuggingFace Dataset Base Implementation |
| 3256 | # ----------------------------------------------------------------------------- |
| 3257 | class HuggingFaceDataset(BenchmarkDataset): |
| 3258 | """Base class for datasets hosted on HuggingFace.""" |
| 3259 | |
| 3260 | SUPPORTED_DATASET_PATHS: set[str] | dict[str, Callable] = set() |
| 3261 | |
| 3262 | def __init__( |
| 3263 | self, |
| 3264 | dataset_path: str, |
| 3265 | dataset_split: str, |
| 3266 | no_stream: bool = False, |
| 3267 | dataset_subset: str | None = None, |
| 3268 | hf_name: str | None = None, |
| 3269 | trust_remote_code: bool = False, |
| 3270 | **kwargs, |
| 3271 | ) -> None: |
| 3272 | super().__init__(dataset_path=dataset_path, **kwargs) |
| 3273 | |
| 3274 | self.dataset_split = dataset_split |
| 3275 | self.dataset_subset = dataset_subset |
| 3276 | self.load_stream = not no_stream |
| 3277 | self.hf_name = hf_name or dataset_path |
| 3278 | self.trust_remote_code = trust_remote_code |
| 3279 | self.load_data() |
| 3280 | |
| 3281 | def load_data(self) -> None: |
| 3282 | """Load data from HuggingFace datasets.""" |
| 3283 | self.data = load_dataset( |
| 3284 | self.dataset_path, |
| 3285 | name=self.dataset_subset, |
| 3286 | split=self.dataset_split, |
| 3287 | streaming=self.load_stream, |
| 3288 | trust_remote_code=self.trust_remote_code, |
| 3289 | ) |
| 3290 | if not getattr(self, "disable_shuffle", False): |
| 3291 | self.data = self.data.shuffle(seed=self.random_seed) |
| 3292 | |
| 3293 | |
| 3294 | # ----------------------------------------------------------------------------- |
| 3295 | # Conversation Dataset Implementation |
| 3296 | # ----------------------------------------------------------------------------- |
| 3297 | |
| 3298 | |
| 3299 | class ConversationDataset(HuggingFaceDataset): |
| 3300 | """Dataset for text-only conversation data.""" |
| 3301 | |
| 3302 | SUPPORTED_DATASET_PATHS = { |
| 3303 | "Aeala/ShareGPT_Vicuna_unfiltered", |
| 3304 | } |
| 3305 | IS_MULTIMODAL = False |
| 3306 | |
| 3307 | def sample( |
| 3308 | self, |
| 3309 | tokenizer: TokenizerLike, |
| 3310 | num_requests: int, |
| 3311 | request_id_prefix: str = "", |
| 3312 | no_oversample: bool = False, |
| 3313 | output_len: int | None = None, |
| 3314 | enable_multimodal_chat: bool = False, |
| 3315 | **kwargs, |
| 3316 | ) -> list[SampleRequest]: |
| 3317 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3318 | # Filter examples with at least 2 conversations |
| 3319 | filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2) |
| 3320 | sampled_requests: list[SampleRequest] = [] |
| 3321 | ind = 0 |
| 3322 | dynamic_output = output_len is None |
| 3323 | |
| 3324 | for item in filtered_data: |
| 3325 | if len(sampled_requests) >= num_requests: |
| 3326 | break |
| 3327 | conv = item["conversations"] |
| 3328 | prompt, completion = conv[0]["value"], conv[1]["value"] |
| 3329 | |
| 3330 | prompt_ids = tokenizer(prompt).input_ids |
| 3331 | completion_ids = tokenizer(completion).input_ids |
| 3332 | prompt_len = len(prompt_ids) |
| 3333 | completion_len = len(completion_ids) |
| 3334 | output_len = completion_len if dynamic_output else output_len |
| 3335 | assert isinstance(output_len, int) and output_len > 0 |
| 3336 | if dynamic_output and not is_valid_sequence(prompt_len, completion_len): |
| 3337 | continue |
| 3338 | mm_content = process_image(item["image"]) if "image" in item else None |
| 3339 | if enable_multimodal_chat: |
| 3340 | # Note: when chat is enabled the request prompt_len is no longer |
| 3341 | # accurate and we will be using request output to count the |
| 3342 | # actual prompt len and output len |
| 3343 | prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) |
| 3344 | sampled_requests.append( |
| 3345 | SampleRequest( |
| 3346 | prompt=prompt, |
| 3347 | prompt_len=prompt_len, |
| 3348 | expected_output_len=output_len, |
| 3349 | multi_modal_data=mm_content, |
| 3350 | request_id=request_id_prefix + str(ind), |
| 3351 | ) |
| 3352 | ) |
| 3353 | ind += 1 |
| 3354 | self.maybe_oversample_requests( |
| 3355 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3356 | ) |
| 3357 | return sampled_requests |
| 3358 | |
| 3359 | |
| 3360 | class MultiModalConversationDataset(HuggingFaceDataset): |
| 3361 | """Dataset for multimodal conversation data.""" |
| 3362 | |
| 3363 | SUPPORTED_DATASET_PATHS = { |
| 3364 | "lmms-lab/LLaVA-OneVision-Data", |
| 3365 | } |
| 3366 | IS_MULTIMODAL = True |
| 3367 | |
| 3368 | def sample( |
| 3369 | self, |
| 3370 | tokenizer: TokenizerLike, |
| 3371 | num_requests: int, |
| 3372 | request_id_prefix: str = "", |
| 3373 | no_oversample: bool = False, |
| 3374 | output_len: int | None = None, |
| 3375 | enable_multimodal_chat: bool = False, |
| 3376 | **kwargs, |
| 3377 | ) -> list[SampleRequest]: |
| 3378 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3379 | # Filter examples with at least 2 conversations |
| 3380 | filtered_data = self.data.filter(lambda x: len(x["conversations"]) >= 2) |
| 3381 | sampled_requests: list[SampleRequest] = [] |
| 3382 | ind = 0 |
| 3383 | dynamic_output = output_len is None |
| 3384 | |
| 3385 | for item in filtered_data: |
| 3386 | if len(sampled_requests) >= num_requests: |
| 3387 | break |
| 3388 | conv = item["conversations"] |
| 3389 | prompt, completion = conv[0]["value"], conv[1]["value"] |
| 3390 | |
| 3391 | prompt_ids = tokenizer(prompt).input_ids |
| 3392 | completion_ids = tokenizer(completion).input_ids |
| 3393 | prompt_len = len(prompt_ids) |
| 3394 | completion_len = len(completion_ids) |
| 3395 | output_len = completion_len if dynamic_output else output_len |
| 3396 | assert isinstance(output_len, int) and output_len > 0 |
| 3397 | if dynamic_output and not is_valid_sequence(prompt_len, completion_len): |
| 3398 | continue |
| 3399 | mm_content = process_image(item["image"]) if "image" in item else None |
| 3400 | if enable_multimodal_chat: |
| 3401 | # Note: when chat is enabled the request prompt_len is no longer |
| 3402 | # accurate and we will be using request output to count the |
| 3403 | # actual prompt len and output len |
| 3404 | prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) |
| 3405 | sampled_requests.append( |
| 3406 | SampleRequest( |
| 3407 | prompt=prompt, |
| 3408 | prompt_len=prompt_len, |
| 3409 | expected_output_len=output_len, |
| 3410 | multi_modal_data=mm_content, |
| 3411 | request_id=request_id_prefix + str(ind), |
| 3412 | ) |
| 3413 | ) |
| 3414 | ind += 1 |
| 3415 | self.maybe_oversample_requests( |
| 3416 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3417 | ) |
| 3418 | return sampled_requests |
| 3419 | |
| 3420 | |
| 3421 | # ----------------------------------------------------------------------------- |
| 3422 | # Vision Arena Dataset Implementation |
| 3423 | # ----------------------------------------------------------------------------- |
| 3424 | |
| 3425 | |
| 3426 | class VisionArenaDataset(HuggingFaceDataset): |
| 3427 | """Vision Arena Dataset.""" |
| 3428 | |
| 3429 | DEFAULT_OUTPUT_LEN = 128 |
| 3430 | SUPPORTED_DATASET_PATHS = { |
| 3431 | "lmarena-ai/VisionArena-Chat": lambda x: x["conversation"][0][0]["content"], |
| 3432 | "lmarena-ai/vision-arena-bench-v0.1": lambda x: x["turns"][0][0]["content"], |
| 3433 | } |
| 3434 | IS_MULTIMODAL = True |
| 3435 | |
| 3436 | def sample( |
| 3437 | self, |
| 3438 | tokenizer: TokenizerLike, |
| 3439 | num_requests: int, |
| 3440 | request_id_prefix: str = "", |
| 3441 | no_oversample: bool = False, |
| 3442 | output_len: int | None = None, |
| 3443 | enable_multimodal_chat: bool = False, |
| 3444 | **kwargs, |
| 3445 | ) -> list[SampleRequest]: |
| 3446 | parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name) |
| 3447 | if parser_fn is None: |
| 3448 | raise ValueError(f"Unsupported dataset path: {self.hf_name}") |
| 3449 | |
| 3450 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3451 | |
| 3452 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3453 | sampled_requests: list[SampleRequest] = [] |
| 3454 | for i, item in enumerate(self.data): |
| 3455 | if len(sampled_requests) >= num_requests: |
| 3456 | break |
| 3457 | |
| 3458 | prompt = parser_fn(item) |
| 3459 | mm_content = process_image(item["images"][0]) |
| 3460 | prompt_len = len(tokenizer.encode(prompt)) |
| 3461 | if enable_multimodal_chat: |
| 3462 | # Note: when chat is enabled the request prompt_len is no longer |
| 3463 | # accurate and we will be using request output to count the |
| 3464 | # actual prompt len |
| 3465 | prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) |
| 3466 | |
| 3467 | sampled_requests.append( |
| 3468 | SampleRequest( |
| 3469 | prompt=prompt, |
| 3470 | prompt_len=prompt_len, |
| 3471 | expected_output_len=output_len, |
| 3472 | multi_modal_data=mm_content, |
| 3473 | request_id=request_id_prefix + str(i), |
| 3474 | ) |
| 3475 | ) |
| 3476 | |
| 3477 | self.maybe_oversample_requests( |
| 3478 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3479 | ) |
| 3480 | return sampled_requests |
| 3481 | |
| 3482 | |
| 3483 | class MMVUDataset(HuggingFaceDataset): |
| 3484 | """MMVU Dataset. |
| 3485 | https://huggingface.co/datasets/yale-nlp/MMVU |
| 3486 | """ |
| 3487 | |
| 3488 | DEFAULT_OUTPUT_LEN = 128 |
| 3489 | SUPPORTED_DATASET_PATHS = { |
| 3490 | "yale-nlp/MMVU": lambda x: ( |
| 3491 | x["question"] |
| 3492 | + " " |
| 3493 | + (" ".join(f"{k}.{v}" for k, v in x["choices"].items())) |
| 3494 | ), |
| 3495 | } |
| 3496 | |
| 3497 | def __init__(self, **kwargs) -> None: |
| 3498 | super().__init__(**kwargs) |
| 3499 | |
| 3500 | self._remote_path_root = ( |
| 3501 | f"https://huggingface.co/datasets/{self.hf_name}/resolve/main" |
| 3502 | ) |
| 3503 | self._local_path_root = hf_api().snapshot_download( |
| 3504 | self.hf_name, |
| 3505 | repo_type="dataset", |
| 3506 | ) |
| 3507 | |
| 3508 | def sample( |
| 3509 | self, |
| 3510 | tokenizer: TokenizerLike, |
| 3511 | num_requests: int, |
| 3512 | request_id_prefix: str = "", |
| 3513 | no_oversample: bool = False, |
| 3514 | output_len: int | None = None, |
| 3515 | enable_multimodal_chat: bool = False, |
| 3516 | **kwargs, |
| 3517 | ) -> list[SampleRequest]: |
| 3518 | parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.hf_name) |
| 3519 | if parser_fn is None: |
| 3520 | raise ValueError(f"Unsupported dataset path: {self.hf_name}") |
| 3521 | |
| 3522 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3523 | |
| 3524 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3525 | sampled_requests: list[SampleRequest] = [] |
| 3526 | for i, item in enumerate(self.data): |
| 3527 | if len(sampled_requests) >= num_requests: |
| 3528 | break |
| 3529 | |
| 3530 | prompt = parser_fn(item) |
| 3531 | mm_content = process_video( |
| 3532 | item["video"].replace(self._remote_path_root, self._local_path_root) |
| 3533 | ) |
| 3534 | prompt_len = len(tokenizer.encode(prompt)) |
| 3535 | if enable_multimodal_chat: |
| 3536 | # Note: when chat is enabled the request prompt_len is no longer |
| 3537 | # accurate and we will be using request output to count the |
| 3538 | # actual prompt len |
| 3539 | prompt = self.apply_multimodal_chat_transformation(prompt, mm_content) |
| 3540 | |
| 3541 | sampled_requests.append( |
| 3542 | SampleRequest( |
| 3543 | prompt=prompt, |
| 3544 | prompt_len=prompt_len, |
| 3545 | expected_output_len=output_len, |
| 3546 | multi_modal_data=mm_content, |
| 3547 | request_id=request_id_prefix + str(i), |
| 3548 | ) |
| 3549 | ) |
| 3550 | |
| 3551 | self.maybe_oversample_requests( |
| 3552 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3553 | ) |
| 3554 | return sampled_requests |
| 3555 | |
| 3556 | |
| 3557 | # ----------------------------------------------------------------------------- |
| 3558 | # Instruct Coder Dataset Implementation |
| 3559 | # ----------------------------------------------------------------------------- |
| 3560 | |
| 3561 | |
| 3562 | class InstructCoderDataset(HuggingFaceDataset): |
| 3563 | """InstructCoder Dataset. |
| 3564 | https://huggingface.co/datasets/likaixin/InstructCoder |
| 3565 | |
| 3566 | InstructCoder is the dataset designed for general code editing. It consists |
| 3567 | of 114,239 instruction-input-output triplets, and covers multiple distinct |
| 3568 | code editing scenario. |
| 3569 | """ |
| 3570 | |
| 3571 | DEFAULT_OUTPUT_LEN = 200 # this is the average default output length |
| 3572 | SUPPORTED_DATASET_PATHS = { |
| 3573 | "likaixin/InstructCoder", |
| 3574 | } |
| 3575 | |
| 3576 | def sample( |
| 3577 | self, |
| 3578 | tokenizer: TokenizerLike, |
| 3579 | num_requests: int, |
| 3580 | request_id_prefix: str = "", |
| 3581 | no_oversample: bool = False, |
| 3582 | output_len: int | None = None, |
| 3583 | enable_multimodal_chat: bool = False, |
| 3584 | skip_chat_template: bool = False, |
| 3585 | **kwargs, |
| 3586 | ) -> list[SampleRequest]: |
| 3587 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3588 | sampled_requests: list[SampleRequest] = [] |
| 3589 | for i, prompt in enumerate(self.sample_prompts(n=num_requests)): |
| 3590 | prompt_text = prompt |
| 3591 | # apply template |
| 3592 | if not skip_chat_template: |
| 3593 | prompt_text_result = tokenizer.apply_chat_template( |
| 3594 | [{"role": "user", "content": prompt}], |
| 3595 | add_generation_prompt=True, |
| 3596 | tokenize=False, |
| 3597 | ) |
| 3598 | assert isinstance(prompt_text_result, str) |
| 3599 | prompt_text = prompt_text_result |
| 3600 | |
| 3601 | prompt_len = len(tokenizer(prompt_text).input_ids) |
| 3602 | sampled_requests.append( |
| 3603 | SampleRequest( |
| 3604 | prompt=prompt_text, |
| 3605 | prompt_len=prompt_len, |
| 3606 | expected_output_len=output_len, |
| 3607 | request_id=request_id_prefix + str(i), |
| 3608 | ) |
| 3609 | ) |
| 3610 | self.maybe_oversample_requests( |
| 3611 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3612 | ) |
| 3613 | return sampled_requests |
| 3614 | |
| 3615 | def sample_prompts(self, n: int) -> Iterator[str]: |
| 3616 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3617 | for item in self.data.take(n): |
| 3618 | prompt = ( |
| 3619 | f"{item['input']}\n\n{item['instruction']} Just output " |
| 3620 | "the code, do not include any explanation." |
| 3621 | ) |
| 3622 | yield prompt |
| 3623 | |
| 3624 | |
| 3625 | # ----------------------------------------------------------------------------- |
| 3626 | # MT-Bench Dataset Implementation |
| 3627 | # ----------------------------------------------------------------------------- |
| 3628 | |
| 3629 | |
| 3630 | class MTBenchDataset(HuggingFaceDataset): |
| 3631 | """MT-Bench Dataset. |
| 3632 | https://huggingface.co/datasets/philschmid/mt-bench |
| 3633 | |
| 3634 | We create a single turn dataset for MT-Bench. |
| 3635 | This is similar to Spec decoding benchmark setup in vLLM |
| 3636 | https://github.com/vllm-project/vllm/blob/9d98ab5ec/examples/offline_inference/eagle.py#L14-L18 |
| 3637 | """ # noqa: E501 |
| 3638 | |
| 3639 | DEFAULT_OUTPUT_LEN = 256 # avg len used in SD bench in vLLM |
| 3640 | SUPPORTED_DATASET_PATHS = { |
| 3641 | "philschmid/mt-bench", |
| 3642 | } |
| 3643 | |
| 3644 | def sample( |
| 3645 | self, |
| 3646 | tokenizer: TokenizerLike, |
| 3647 | num_requests: int, |
| 3648 | request_id_prefix: str = "", |
| 3649 | no_oversample: bool = False, |
| 3650 | output_len: int | None = None, |
| 3651 | enable_multimodal_chat: bool = False, |
| 3652 | skip_chat_template: bool = False, |
| 3653 | **kwargs, |
| 3654 | ) -> list[SampleRequest]: |
| 3655 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3656 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3657 | sampled_requests: list[SampleRequest] = [] |
| 3658 | |
| 3659 | for i, item in enumerate(self.data): |
| 3660 | if len(sampled_requests) >= num_requests: |
| 3661 | break |
| 3662 | prompt = item["turns"][0] |
| 3663 | |
| 3664 | # apply template |
| 3665 | if not skip_chat_template: |
| 3666 | prompt = tokenizer.apply_chat_template( |
| 3667 | [{"role": "user", "content": prompt}], |
| 3668 | add_generation_prompt=True, |
| 3669 | tokenize=False, |
| 3670 | ) |
| 3671 | |
| 3672 | prompt_len = len(tokenizer(prompt).input_ids) |
| 3673 | sampled_requests.append( |
| 3674 | SampleRequest( |
| 3675 | prompt=prompt, |
| 3676 | prompt_len=prompt_len, |
| 3677 | expected_output_len=output_len, |
| 3678 | request_id=request_id_prefix + str(i), |
| 3679 | ) |
| 3680 | ) |
| 3681 | self.maybe_oversample_requests( |
| 3682 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3683 | ) |
| 3684 | return sampled_requests |
| 3685 | |
| 3686 | |
| 3687 | # ----------------------------------------------------------------------------- |
| 3688 | # HumanEval Dataset Implementation |
| 3689 | # ----------------------------------------------------------------------------- |
| 3690 | |
| 3691 | |
| 3692 | class HumanEvalDataset(HuggingFaceDataset): |
| 3693 | """HumanEvalDataset Dataset. |
| 3694 | https://huggingface.co/datasets/openai/openai_humaneval |
| 3695 | |
| 3696 | We create a single turn dataset for HumanEval. |
| 3697 | """ |
| 3698 | |
| 3699 | DEFAULT_OUTPUT_LEN = 256 |
| 3700 | SUPPORTED_DATASET_PATHS = { |
| 3701 | "openai/openai_humaneval", |
| 3702 | } |
| 3703 | |
| 3704 | def sample( |
| 3705 | self, |
| 3706 | tokenizer: TokenizerLike, |
| 3707 | num_requests: int, |
| 3708 | request_id_prefix: str = "", |
| 3709 | no_oversample: bool = False, |
| 3710 | output_len: int | None = None, |
| 3711 | enable_multimodal_chat: bool = False, |
| 3712 | skip_chat_template: bool = False, |
| 3713 | **kwargs, |
| 3714 | ) -> list[SampleRequest]: |
| 3715 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3716 | sampled_requests: list[SampleRequest] = [] |
| 3717 | assert self.data is not None, "Data must be loaded before sampling" |
| 3718 | |
| 3719 | for i, item in enumerate(self.data): |
| 3720 | if len(sampled_requests) >= num_requests: |
| 3721 | break |
| 3722 | prompt = item["prompt"] |
| 3723 | |
| 3724 | # apply template |
| 3725 | if not skip_chat_template: |
| 3726 | prompt = tokenizer.apply_chat_template( |
| 3727 | [{"role": "user", "content": prompt}], |
| 3728 | add_generation_prompt=True, |
| 3729 | tokenize=False, |
| 3730 | ) |
| 3731 | |
| 3732 | prompt_len = len(tokenizer(prompt).input_ids) |
| 3733 | sampled_requests.append( |
| 3734 | SampleRequest( |
| 3735 | prompt=prompt, |
| 3736 | prompt_len=prompt_len, |
| 3737 | expected_output_len=output_len, |
| 3738 | request_id=request_id_prefix + str(i), |
| 3739 | ) |
| 3740 | ) |
| 3741 | self.maybe_oversample_requests( |
| 3742 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3743 | ) |
| 3744 | return sampled_requests |
| 3745 | |
| 3746 | |
| 3747 | # ----------------------------------------------------------------------------- |
| 3748 | # GSM8K Dataset Implementation |
| 3749 | # ----------------------------------------------------------------------------- |
| 3750 | |
| 3751 | |
| 3752 | class GSM8KDataset(HuggingFaceDataset): |
| 3753 | """GSM8K Dataset. |
| 3754 | https://huggingface.co/datasets/openai/gsm8k |
| 3755 | |
| 3756 | We create a single turn dataset for GSM8K. |
| 3757 | """ |
| 3758 | |
| 3759 | DEFAULT_OUTPUT_LEN = 256 |
| 3760 | SUPPORTED_DATASET_PATHS = { |
| 3761 | "openai/gsm8k", |
| 3762 | } |
| 3763 | |
| 3764 | def sample( |
| 3765 | self, |
| 3766 | tokenizer: TokenizerLike, |
| 3767 | num_requests: int, |
| 3768 | request_id_prefix: str = "", |
| 3769 | no_oversample: bool = False, |
| 3770 | output_len: int | None = None, |
| 3771 | enable_multimodal_chat: bool = False, |
| 3772 | skip_chat_template: bool = False, |
| 3773 | **kwargs, |
| 3774 | ) -> list[SampleRequest]: |
| 3775 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3776 | sampled_requests: list[SampleRequest] = [] |
| 3777 | assert self.data is not None, "Data must be loaded before sampling" |
| 3778 | |
| 3779 | for i, item in enumerate(self.data): |
| 3780 | if len(sampled_requests) >= num_requests: |
| 3781 | break |
| 3782 | prompt = item["question"] |
| 3783 | |
| 3784 | # apply template |
| 3785 | if not skip_chat_template: |
| 3786 | prompt = tokenizer.apply_chat_template( |
| 3787 | [{"role": "user", "content": prompt}], |
| 3788 | add_generation_prompt=True, |
| 3789 | tokenize=False, |
| 3790 | ) |
| 3791 | |
| 3792 | prompt_len = len(tokenizer(prompt).input_ids) |
| 3793 | sampled_requests.append( |
| 3794 | SampleRequest( |
| 3795 | prompt=prompt, |
| 3796 | prompt_len=prompt_len, |
| 3797 | expected_output_len=output_len, |
| 3798 | request_id=request_id_prefix + str(i), |
| 3799 | ) |
| 3800 | ) |
| 3801 | self.maybe_oversample_requests( |
| 3802 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3803 | ) |
| 3804 | return sampled_requests |
| 3805 | |
| 3806 | |
| 3807 | # ----------------------------------------------------------------------------- |
| 3808 | # Blazedit Dataset Implementation |
| 3809 | # ----------------------------------------------------------------------------- |
| 3810 | |
| 3811 | |
| 3812 | class BlazeditDataset(HuggingFaceDataset): |
| 3813 | """Blazedit Dataset. |
| 3814 | https://github.com/ise-uiuc/blazedit |
| 3815 | |
| 3816 | 5k char version: vdaita/edit_5k_char |
| 3817 | 10k char version: vdaita/edit_10k_char |
| 3818 | """ # noqa: E501 |
| 3819 | |
| 3820 | # 5k char version will have output as ~5k chars |
| 3821 | # 10k char version will have output as ~10k chars |
| 3822 | # Assuming 3 char per token, 10k chars will be 3333 tokens |
| 3823 | # We set default to 4000 to be safe |
| 3824 | DEFAULT_OUTPUT_LEN = 4000 |
| 3825 | SUPPORTED_DATASET_PATHS = { |
| 3826 | "vdaita/edit_5k_char", |
| 3827 | "vdaita/edit_10k_char", |
| 3828 | } |
| 3829 | |
| 3830 | def sample( |
| 3831 | self, |
| 3832 | tokenizer: TokenizerLike, |
| 3833 | num_requests: int, |
| 3834 | request_id_prefix: str = "", |
| 3835 | no_oversample: bool = False, |
| 3836 | output_len: int | None = None, |
| 3837 | skip_chat_template: bool = False, |
| 3838 | min_distance: float = 0.0, |
| 3839 | max_distance: float = 1.0, |
| 3840 | **kwargs, |
| 3841 | ) -> list[SampleRequest]: |
| 3842 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 3843 | sampled_requests: list[SampleRequest] = [] |
| 3844 | |
| 3845 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3846 | for i, item in enumerate(self.data): |
| 3847 | if len(sampled_requests) >= num_requests: |
| 3848 | break |
| 3849 | code = item["code"] |
| 3850 | change_request = item["change_request"] |
| 3851 | norm_distance = item["norm_distance"] |
| 3852 | |
| 3853 | # compare the levenshtein distance normalized by code length |
| 3854 | if norm_distance < min_distance or norm_distance > max_distance: |
| 3855 | continue |
| 3856 | |
| 3857 | # template copied from |
| 3858 | # https://github.com/ise-uiuc/blazedit/blob/7765137e656fd62de877422d2e4cf8de51228054/dataset/create_refined_dataset.py#L94-L105 # noqa: E501 |
| 3859 | prompt = f"""Given a code file, please apply the change requests and generate the new file. |
| 3860 | |
| 3861 | Original file: |
| 3862 | ```python |
| 3863 | {code} |
| 3864 | ``` |
| 3865 | |
| 3866 | Change request: |
| 3867 | {change_request} |
| 3868 | |
| 3869 | Please generate the new code file in the "New file" section below.""" # noqa: E501 |
| 3870 | |
| 3871 | # apply template |
| 3872 | if not skip_chat_template: |
| 3873 | prompt_result = tokenizer.apply_chat_template( |
| 3874 | [{"role": "user", "content": prompt}], |
| 3875 | add_generation_prompt=True, |
| 3876 | tokenize=False, |
| 3877 | ) |
| 3878 | assert isinstance(prompt_result, str) |
| 3879 | prompt = prompt_result |
| 3880 | |
| 3881 | prompt_len = len(tokenizer(prompt).input_ids) |
| 3882 | |
| 3883 | sampled_requests.append( |
| 3884 | SampleRequest( |
| 3885 | prompt=prompt, |
| 3886 | prompt_len=prompt_len, |
| 3887 | expected_output_len=output_len, |
| 3888 | request_id=request_id_prefix + str(i), |
| 3889 | ) |
| 3890 | ) |
| 3891 | self.maybe_oversample_requests( |
| 3892 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3893 | ) |
| 3894 | |
| 3895 | return sampled_requests |
| 3896 | |
| 3897 | |
| 3898 | # ----------------------------------------------------------------------------- |
| 3899 | # AIMO Dataset Implementation |
| 3900 | # ----------------------------------------------------------------------------- |
| 3901 | |
| 3902 | |
| 3903 | class AIMODataset(HuggingFaceDataset): |
| 3904 | """Dataset class for processing a AIMO dataset with reasoning questions.""" |
| 3905 | |
| 3906 | SUPPORTED_DATASET_PATHS = { |
| 3907 | "AI-MO/aimo-validation-aime", |
| 3908 | "AI-MO/NuminaMath-1.5", |
| 3909 | "AI-MO/NuminaMath-CoT", |
| 3910 | } |
| 3911 | |
| 3912 | def sample( |
| 3913 | self, |
| 3914 | tokenizer: TokenizerLike, |
| 3915 | num_requests: int, |
| 3916 | request_id_prefix: str = "", |
| 3917 | no_oversample: bool = False, |
| 3918 | output_len: int | None = None, |
| 3919 | **kwargs, |
| 3920 | ) -> list[SampleRequest]: |
| 3921 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 3922 | sampled_requests: list[SampleRequest] = [] |
| 3923 | ind = 0 |
| 3924 | dynamic_output = output_len is None |
| 3925 | |
| 3926 | for item in self.data: |
| 3927 | if len(sampled_requests) >= num_requests: |
| 3928 | break |
| 3929 | prompt, completion = item["problem"], item["solution"] |
| 3930 | |
| 3931 | prompt_ids = tokenizer(prompt).input_ids |
| 3932 | completion_ids = tokenizer(completion).input_ids |
| 3933 | prompt_len = len(prompt_ids) |
| 3934 | completion_len = len(completion_ids) |
| 3935 | output_len = completion_len if dynamic_output else output_len |
| 3936 | assert isinstance(output_len, int) and output_len > 0 |
| 3937 | if dynamic_output and not is_valid_sequence( |
| 3938 | prompt_len, completion_len, max_prompt_len=2048, max_total_len=32000 |
| 3939 | ): |
| 3940 | continue |
| 3941 | sampled_requests.append( |
| 3942 | SampleRequest( |
| 3943 | prompt=prompt, |
| 3944 | prompt_len=prompt_len, |
| 3945 | expected_output_len=output_len, |
| 3946 | multi_modal_data=None, |
| 3947 | request_id=request_id_prefix + str(ind), |
| 3948 | ) |
| 3949 | ) |
| 3950 | ind += 1 |
| 3951 | self.maybe_oversample_requests( |
| 3952 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 3953 | ) |
| 3954 | return sampled_requests |
| 3955 | |
| 3956 | |
| 3957 | # ----------------------------------------------------------------------------- |
| 3958 | # Next Edit Prediction Dataset Implementation |
| 3959 | # ----------------------------------------------------------------------------- |
| 3960 | |
| 3961 | |
| 3962 | zeta_prompt = """### Instruction: |
| 3963 | 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. |
| 3964 | |
| 3965 | ### User Edits: |
| 3966 | |
| 3967 | {} |
| 3968 | |
| 3969 | ### User Excerpt: |
| 3970 | |
| 3971 | {} |
| 3972 | |
| 3973 | ### Response: |
| 3974 | |
| 3975 | """ # noqa: E501 |
| 3976 | |
| 3977 | |
| 3978 | def _format_zeta_prompt( |
| 3979 | sample: dict, original_start_marker: str = "<|editable_region_start|>" |
| 3980 | ) -> dict: |
| 3981 | """Format the zeta prompt for the Next Edit Prediction (NEP) dataset. |
| 3982 | |
| 3983 | This function formats examples from the NEP dataset |
| 3984 | into prompts and expected outputs. It could be |
| 3985 | further extended to support more NEP datasets. |
| 3986 | |
| 3987 | Args: |
| 3988 | sample: The dataset sample containing events, |
| 3989 | inputs, and outputs. |
| 3990 | original_start_marker: The marker indicating the |
| 3991 | start of the editable region. Defaults to |
| 3992 | "<|editable_region_start|>". |
| 3993 | |
| 3994 | Returns: |
| 3995 | A dictionary with the formatted prompts and expected outputs. |
| 3996 | |
| 3997 | """ |
| 3998 | events = sample["events"] |
| 3999 | input = sample["input"] |
| 4000 | output = sample["output"] |
| 4001 | prompt = zeta_prompt.format(events, input) |
| 4002 | |
| 4003 | # following the original implementation, extract the focused region |
| 4004 | # from the raw output |
| 4005 | output_start_index = output.find(original_start_marker) |
| 4006 | output_focused_region = output[output_start_index:] |
| 4007 | expected_output = output_focused_region |
| 4008 | |
| 4009 | return {"prompt": prompt, "expected_output": expected_output} |
| 4010 | |
| 4011 | |
| 4012 | class NextEditPredictionDataset(HuggingFaceDataset): |
| 4013 | """Dataset class for processing a Next Edit Prediction dataset.""" |
| 4014 | |
| 4015 | SUPPORTED_DATASET_PATHS = { |
| 4016 | "zed-industries/zeta", |
| 4017 | } |
| 4018 | MAPPING_PROMPT_FUNCS = { |
| 4019 | "zed-industries/zeta": _format_zeta_prompt, |
| 4020 | } |
| 4021 | |
| 4022 | def sample( |
| 4023 | self, |
| 4024 | tokenizer: TokenizerLike, |
| 4025 | num_requests: int, |
| 4026 | request_id_prefix: str = "", |
| 4027 | no_oversample: bool = False, |
| 4028 | **kwargs, |
| 4029 | ): |
| 4030 | formatting_prompt_func = self.MAPPING_PROMPT_FUNCS.get(self.hf_name) |
| 4031 | if formatting_prompt_func is None: |
| 4032 | raise ValueError(f"Unsupported dataset path: {self.hf_name}") |
| 4033 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 4034 | samples = [] |
| 4035 | for i, sample in enumerate(self.data): |
| 4036 | sample = formatting_prompt_func(sample) |
| 4037 | samples.append( |
| 4038 | SampleRequest( |
| 4039 | prompt=sample["prompt"], |
| 4040 | prompt_len=len(tokenizer(sample["prompt"]).input_ids), |
| 4041 | expected_output_len=len( |
| 4042 | tokenizer(sample["expected_output"]).input_ids |
| 4043 | ), |
| 4044 | request_id=request_id_prefix + str(i), |
| 4045 | ) |
| 4046 | ) |
| 4047 | if len(samples) >= num_requests: |
| 4048 | break |
| 4049 | self.maybe_oversample_requests( |
| 4050 | samples, num_requests, request_id_prefix, no_oversample |
| 4051 | ) |
| 4052 | return samples |
| 4053 | |
| 4054 | |
| 4055 | # ----------------------------------------------------------------------------- |
| 4056 | # ASR Dataset Implementation |
| 4057 | # ----------------------------------------------------------------------------- |
| 4058 | |
| 4059 | |
| 4060 | class ASRDataset(HuggingFaceDataset): |
| 4061 | """Dataset class for processing a ASR dataset for transcription. |
| 4062 | Tested on the following set: |
| 4063 | |
| 4064 | +---------------------------+----------------------------------------+--------------------------+-----------------------------+ |
| 4065 | | Dataset | Domain | Speaking Style | hf-subset | |
| 4066 | +---------------------------+----------------------------------------+--------------------------+-----------------------------+ |
| 4067 | | TED-LIUM | TED talks | Oratory | release1, release2, release3| |
| 4068 | | | | | release3-speaker-adaptation | |
| 4069 | | VoxPopuli | European Parliament | Oratory | en, de, it, fr, ... | |
| 4070 | | LibriSpeech | Audiobook | Narrated | "LIUM/tedlium" | |
| 4071 | | GigaSpeech | Audiobook, podcast, YouTube | Narrated, spontaneous | xs, s, m, l, xl, dev, test | |
| 4072 | | SPGISpeech | Financial meetings | Oratory, spontaneous | S, M, L, dev, test | |
| 4073 | | Earnings22-Cleaned-AA | Long form earnings calls | Prepared remarks, Q&A | test | |
| 4074 | | Earnings22-Tiny-Filtered | Earnings calls | Prepared remarks, Q&A | validation | |
| 4075 | | AMI | Meetings | Spontaneous | ihm, sdm | |
| 4076 | +---------------------------+----------------------------------------+--------------------------+-----------------------------+ |
| 4077 | |
| 4078 | """ # noqa: E501 |
| 4079 | |
| 4080 | EARNINGS22_CLEANED_DATASET = "ArtificialAnalysis/Earnings22-Cleaned-AA" |
| 4081 | EARNINGS22_TINY_FILTERED_DATASET = ( |
| 4082 | "D4nt3/esb-datasets-earnings22-validation-tiny-filtered" |
| 4083 | ) |
| 4084 | LIBRISPEECH_DATASET = "openslr/librispeech_asr" |
| 4085 | |
| 4086 | SUPPORTED_DATASET_PATHS = { |
| 4087 | LIBRISPEECH_DATASET, |
| 4088 | "facebook/voxpopuli", |
| 4089 | "LIUM/tedlium", |
| 4090 | "edinburghcstr/ami", |
| 4091 | "speechcolab/gigaspeech", |
| 4092 | "kensho/spgispeech", |
| 4093 | EARNINGS22_CLEANED_DATASET, |
| 4094 | EARNINGS22_TINY_FILTERED_DATASET, |
| 4095 | } |
| 4096 | |
| 4097 | DEFAULT_OUTPUT_LEN = 1024 |
| 4098 | IS_MULTIMODAL = True |
| 4099 | |
| 4100 | def load_data(self) -> None: |
| 4101 | if self.hf_name == self.EARNINGS22_CLEANED_DATASET: |
| 4102 | # This subset stores repo-local MP3 paths instead of a HF `Audio` |
| 4103 | # column, so eagerly materialize it back into the common schema. |
| 4104 | self.data = load_dataset( |
| 4105 | self.dataset_path, |
| 4106 | name=self.dataset_subset, |
| 4107 | split=self.dataset_split, |
| 4108 | streaming=False, |
| 4109 | trust_remote_code=self.trust_remote_code, |
| 4110 | ) |
| 4111 | if not getattr(self, "disable_shuffle", False): |
| 4112 | self.data = self.data.shuffle(seed=self.random_seed) |
| 4113 | self._materialize_local_audio_column() |
| 4114 | return |
| 4115 | if self.hf_name in ( |
| 4116 | self.EARNINGS22_TINY_FILTERED_DATASET, |
| 4117 | self.LIBRISPEECH_DATASET, |
| 4118 | ): |
| 4119 | super().load_data() |
| 4120 | self._disable_audio_decode() |
| 4121 | return |
| 4122 | |
| 4123 | super().load_data() |
| 4124 | |
| 4125 | def _disable_audio_decode(self) -> None: |
| 4126 | from datasets import Audio |
| 4127 | |
| 4128 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 4129 | self.data = self.data.cast_column("audio", Audio(decode=False)) |
| 4130 | |
| 4131 | def _materialize_local_audio_column(self) -> None: |
| 4132 | local_path_root = Path( |
| 4133 | hf_api().snapshot_download(self.hf_name, repo_type="dataset") |
| 4134 | ) |
| 4135 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 4136 | self.data = self.data.map( |
| 4137 | lambda item: { |
| 4138 | "audio": str(local_path_root / item["url"]), |
| 4139 | "text": item["transcript"], |
| 4140 | } |
| 4141 | ) |
| 4142 | self._disable_audio_decode() |
| 4143 | |
| 4144 | def sample( |
| 4145 | self, |
| 4146 | tokenizer: TokenizerLike, |
| 4147 | num_requests: int, |
| 4148 | request_id_prefix: str = "", |
| 4149 | no_oversample: bool = False, |
| 4150 | output_len: int | None = None, |
| 4151 | **kwargs, |
| 4152 | ) -> list[SampleRequest]: |
| 4153 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 4154 | name_or_path = getattr(tokenizer, "name_or_path", "") |
| 4155 | tok_class = type(tokenizer).__name__ |
| 4156 | if "openai" in name_or_path: |
| 4157 | prompt = "<|startoftranscript|><|en|><|transcribe|><|notimestamps|>" |
| 4158 | elif tok_class == "CohereAsrTokenizer" or "cohere" in name_or_path.lower(): |
| 4159 | # CohereAsrTokenizer does not inject a decoder start token, so the |
| 4160 | # decoder prompt must supply the full control-token sequence. |
| 4161 | # Token order: context boundary, transcript start, emotion (default |
| 4162 | # undefined), language (en), transcription directive (en), punctuation |
| 4163 | # enabled, no ITN, no timestamp, no diarization. |
| 4164 | prompt = ( |
| 4165 | "<|startofcontext|><|startoftranscript|>" |
| 4166 | "<|emo:undefined|><|en|><|en|><|pnc|><|noitn|>" |
| 4167 | "<|notimestamp|><|nodiarize|>" |
| 4168 | ) |
| 4169 | else: |
| 4170 | prompt = "" |
| 4171 | prompt_len = len(tokenizer(prompt).input_ids) |
| 4172 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 4173 | sampled_requests: list[SampleRequest] = [] |
| 4174 | ind = 0 |
| 4175 | skipped = 0 |
| 4176 | asr_min_audio_len_sec: float = float(kwargs.get("asr_min_audio_len_sec") or 0.0) |
| 4177 | asr_max_audio_len_sec: float = float( |
| 4178 | kwargs.get("asr_max_audio_len_sec") or float("inf") |
| 4179 | ) |
| 4180 | durations = [] |
| 4181 | for item in self.data: |
| 4182 | if len(sampled_requests) >= num_requests: |
| 4183 | break |
| 4184 | mm_content: dict[str, Any] |
| 4185 | audio = item["audio"] |
| 4186 | if ( |
| 4187 | isinstance(audio, dict) |
| 4188 | and "array" in audio |
| 4189 | and "sampling_rate" in audio |
| 4190 | ): |
| 4191 | y, sr = audio["array"], audio["sampling_rate"] |
| 4192 | duration_s = get_audio_duration(y=y, sr=sr) |
| 4193 | mm_content = {"audio": (y, sr)} |
| 4194 | elif isinstance(audio, str): |
| 4195 | duration_s = sf.info(audio).duration |
| 4196 | mm_content = {"audio_path": audio} |
| 4197 | elif isinstance(audio, dict) and audio.get("bytes") is not None: |
| 4198 | with BytesIO(audio["bytes"]) as audio_buffer: |
| 4199 | y, sr = sf.read(audio_buffer, dtype="float32") |
| 4200 | duration_s = get_audio_duration(y=y, sr=sr) |
| 4201 | mm_content = {"audio": (y, sr)} |
| 4202 | elif isinstance(audio, dict) and audio.get("path"): |
| 4203 | duration_s = sf.info(audio["path"]).duration |
| 4204 | mm_content = {"audio_path": audio["path"]} |
| 4205 | else: |
| 4206 | raise ValueError( |
| 4207 | "ASR samples must provide decoded audio arrays, " |
| 4208 | "embedded audio bytes, or a local audio path." |
| 4209 | ) |
| 4210 | if duration_s < asr_min_audio_len_sec or duration_s > asr_max_audio_len_sec: |
| 4211 | skipped += 1 |
| 4212 | continue |
| 4213 | |
| 4214 | durations.append(duration_s) |
| 4215 | sampled_requests.append( |
| 4216 | SampleRequest( |
| 4217 | prompt=prompt, |
| 4218 | prompt_len=prompt_len, |
| 4219 | expected_output_len=output_len, |
| 4220 | multi_modal_data=mm_content, |
| 4221 | request_id=request_id_prefix + str(ind), |
| 4222 | ) |
| 4223 | ) |
| 4224 | ind += 1 |
| 4225 | if skipped: |
| 4226 | logger.warning( |
| 4227 | "%d samples discarded from dataset due to" |
| 4228 | " their length being greater than" |
| 4229 | " what Whisper supports.", |
| 4230 | skipped, |
| 4231 | ) |
| 4232 | |
| 4233 | logger.info("Number of audio samples: %d", len(durations)) |
| 4234 | avg_duration = sum(durations) / len(durations) if durations else 0 |
| 4235 | min_duration = min(durations) if durations else 0 |
| 4236 | max_duration = max(durations) if durations else 0 |
| 4237 | median_duration = np.median(durations) if durations else 0 |
| 4238 | logger.info( |
| 4239 | "Audio duration statistics (s): avg=%.2f, min=%.2f, max=%.2f, median=%.2f", |
| 4240 | avg_duration, |
| 4241 | min_duration, |
| 4242 | max_duration, |
| 4243 | median_duration, |
| 4244 | ) |
| 4245 | |
| 4246 | self.maybe_oversample_requests( |
| 4247 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 4248 | ) |
| 4249 | return sampled_requests |
| 4250 | |
| 4251 | |
| 4252 | # ----------------------------------------------------------------------------- |
| 4253 | # MLPerf Dataset Implementation |
| 4254 | # ----------------------------------------------------------------------------- |
| 4255 | |
| 4256 | |
| 4257 | class MLPerfDataset(HuggingFaceDataset): |
| 4258 | """MLPerf Inference Dataset. |
| 4259 | |
| 4260 | Dataset on HF: |
| 4261 | https://huggingface.co/datasets/mgoin/mlperf-inference-llama2-data |
| 4262 | https://huggingface.co/datasets/mgoin/mlperf-inference-llama3.1-data |
| 4263 | |
| 4264 | Each record contains: |
| 4265 | - "system_prompt": system role instruction. |
| 4266 | - "question": user question. |
| 4267 | - "output": reference answer. |
| 4268 | |
| 4269 | We combine the system prompt and question into a chat-formatted prompt |
| 4270 | (using the tokenizer's chat template) and set the expected output length to |
| 4271 | the tokenized length of the provided reference answer. |
| 4272 | """ |
| 4273 | |
| 4274 | SUPPORTED_DATASET_PATHS = { |
| 4275 | "mgoin/mlperf-inference-llama2-data", |
| 4276 | "mgoin/mlperf-inference-llama3.1-data", |
| 4277 | } |
| 4278 | |
| 4279 | def sample( |
| 4280 | self, |
| 4281 | tokenizer: TokenizerLike, |
| 4282 | num_requests: int, |
| 4283 | request_id_prefix: str = "", |
| 4284 | no_oversample: bool = False, |
| 4285 | output_len: int | None = None, |
| 4286 | **kwargs, |
| 4287 | ) -> list[SampleRequest]: |
| 4288 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 4289 | sampled_requests: list[SampleRequest] = [] |
| 4290 | ind = 0 |
| 4291 | |
| 4292 | for item in self.data: |
| 4293 | if len(sampled_requests) >= num_requests: |
| 4294 | break |
| 4295 | |
| 4296 | system_prompt = item["system_prompt"] |
| 4297 | question = item["question"] |
| 4298 | reference_answer = item["output"] |
| 4299 | |
| 4300 | # Build chat-style prompt using tokenizer template, if available. |
| 4301 | messages = [ |
| 4302 | {"role": "system", "content": system_prompt}, |
| 4303 | {"role": "user", "content": question}, |
| 4304 | ] |
| 4305 | prompt_formatted = tokenizer.apply_chat_template( |
| 4306 | messages, # type: ignore[arg-type] |
| 4307 | add_generation_prompt=True, |
| 4308 | tokenize=False, |
| 4309 | ) |
| 4310 | assert isinstance(prompt_formatted, str) |
| 4311 | prompt_len = len(tokenizer(prompt_formatted).input_ids) |
| 4312 | |
| 4313 | # Determine output length from reference answer tokens. |
| 4314 | ref_out_len = len( |
| 4315 | tokenizer(reference_answer, add_special_tokens=False).input_ids |
| 4316 | ) |
| 4317 | expected_output_len = ref_out_len if output_len is None else output_len |
| 4318 | |
| 4319 | # Validate sequence lengths. |
| 4320 | if not is_valid_sequence(prompt_len, expected_output_len): |
| 4321 | continue |
| 4322 | |
| 4323 | sampled_requests.append( |
| 4324 | SampleRequest( |
| 4325 | prompt=prompt_formatted, |
| 4326 | prompt_len=prompt_len, |
| 4327 | expected_output_len=expected_output_len, |
| 4328 | request_id=request_id_prefix + str(ind), |
| 4329 | ) |
| 4330 | ) |
| 4331 | ind += 1 |
| 4332 | |
| 4333 | self.maybe_oversample_requests( |
| 4334 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 4335 | ) |
| 4336 | return sampled_requests |
| 4337 | |
| 4338 | |
| 4339 | # ----------------------------------------------------------------------------- |
| 4340 | # Prefix Repetition Dataset Implementation |
| 4341 | # ----------------------------------------------------------------------------- |
| 4342 | |
| 4343 | |
| 4344 | class PrefixRepetitionRandomDataset(BenchmarkDataset): |
| 4345 | # Default values copied from benchmark_serving.py for the repeated prefix |
| 4346 | # dataset. |
| 4347 | DEFAULT_PREFIX_LEN = 256 |
| 4348 | DEFAULT_SUFFIX_LEN = 256 |
| 4349 | DEFAULT_NUM_PREFIXES = 10 |
| 4350 | DEFAULT_OUTPUT_LEN = 128 |
| 4351 | |
| 4352 | def __init__( |
| 4353 | self, |
| 4354 | **kwargs, |
| 4355 | ) -> None: |
| 4356 | super().__init__(**kwargs) |
| 4357 | random.seed(self.random_seed) |
| 4358 | np.random.seed(self.random_seed) |
| 4359 | |
| 4360 | def sample( |
| 4361 | self, |
| 4362 | tokenizer: TokenizerLike, |
| 4363 | num_requests: int, |
| 4364 | request_id_prefix: str = "", |
| 4365 | no_oversample: bool = False, |
| 4366 | prefix_len: int = DEFAULT_PREFIX_LEN, |
| 4367 | suffix_len: int = DEFAULT_SUFFIX_LEN, |
| 4368 | num_prefixes: int = DEFAULT_NUM_PREFIXES, |
| 4369 | output_len: int = DEFAULT_OUTPUT_LEN, |
| 4370 | **kwargs, |
| 4371 | ) -> list[SampleRequest]: |
| 4372 | vocab_size = tokenizer.vocab_size |
| 4373 | prompts_per_prefix = num_requests // num_prefixes |
| 4374 | if prompts_per_prefix == 0: |
| 4375 | raise ValueError( |
| 4376 | f"num_requests ({num_requests}) must be greater than or equal " |
| 4377 | f"to num_prefixes ({num_prefixes})" |
| 4378 | ) |
| 4379 | |
| 4380 | def _generate_exact_length_tokens(target_length: int) -> tuple[list[int], int]: |
| 4381 | """Generate tokens that decode and re-encode to exactly |
| 4382 | target_length.""" |
| 4383 | # Generate random tokens |
| 4384 | tokens = np.random.randint(0, vocab_size, size=target_length).tolist() |
| 4385 | |
| 4386 | _, adjusted_tokens, token_mismatch = gen_prompt_decode_to_target_len( # noqa: E501 |
| 4387 | tokenizer=tokenizer, |
| 4388 | token_sequence=tokens, |
| 4389 | target_token_len=target_length, |
| 4390 | add_special_tokens=False, |
| 4391 | ) |
| 4392 | return adjusted_tokens, token_mismatch |
| 4393 | |
| 4394 | requests = [] |
| 4395 | token_mismatch_total = 0 |
| 4396 | for _ in range(num_prefixes): |
| 4397 | prefix_tokens, prefix_mismatch = _generate_exact_length_tokens(prefix_len) |
| 4398 | token_mismatch_total += prefix_mismatch |
| 4399 | |
| 4400 | for _ in range(prompts_per_prefix): |
| 4401 | suffix_tokens, suffix_mismatch = _generate_exact_length_tokens( |
| 4402 | suffix_len |
| 4403 | ) |
| 4404 | token_mismatch_total += suffix_mismatch |
| 4405 | combined_tokens = prefix_tokens + suffix_tokens |
| 4406 | prompt = tokenizer.decode(combined_tokens) |
| 4407 | prompt_len = len(combined_tokens) |
| 4408 | requests.append( |
| 4409 | SampleRequest( |
| 4410 | prompt=prompt, |
| 4411 | prompt_len=prompt_len, |
| 4412 | expected_output_len=output_len, |
| 4413 | ) |
| 4414 | ) |
| 4415 | |
| 4416 | if token_mismatch_total != 0: |
| 4417 | sign = "more" if token_mismatch_total > 0 else "fewer" |
| 4418 | logger.warning( |
| 4419 | "Across all generated prompts, there were %d %s tokens " |
| 4420 | "than expected after decoding and re-encoding. This is " |
| 4421 | "expected due to the imperfect nature of the sampling " |
| 4422 | "procedure.", |
| 4423 | abs(token_mismatch_total), |
| 4424 | sign, |
| 4425 | ) |
| 4426 | if not getattr(self, "disable_shuffle", False): |
| 4427 | random.shuffle(requests) |
| 4428 | return requests |
| 4429 | |
| 4430 | |
| 4431 | # ----------------------------------------------------------------------------- |
| 4432 | # MMStar Dataset Implementation |
| 4433 | # ----------------------------------------------------------------------------- |
| 4434 | |
| 4435 | |
| 4436 | class MMStarDataset(HuggingFaceDataset): |
| 4437 | """Lin-Chen/MMStar: https://huggingface.co/datasets/Lin-Chen/MMStar |
| 4438 | refer to: https://github.com/sgl-project/SpecForge/pull/106 |
| 4439 | """ |
| 4440 | |
| 4441 | DEFAULT_OUTPUT_LEN = 128 |
| 4442 | SUPPORTED_DATASET_PATHS = {"Lin-Chen/MMStar"} |
| 4443 | IS_MULTIMODAL = True |
| 4444 | |
| 4445 | def sample( |
| 4446 | self, |
| 4447 | tokenizer: TokenizerLike, |
| 4448 | num_requests: int, |
| 4449 | request_id_prefix: str = "", |
| 4450 | no_oversample: bool = False, |
| 4451 | output_len: int | None = None, |
| 4452 | enable_multimodal_chat: bool = False, |
| 4453 | **kwargs, |
| 4454 | ) -> list[SampleRequest]: |
| 4455 | # If --hf-output-len is not set, use the default output length. |
| 4456 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 4457 | assert self.data is not None, "Dataset must be loaded before sampling" |
| 4458 | sampled_requests: list[SampleRequest] = [] |
| 4459 | |
| 4460 | for ind, item in enumerate(self.data): |
| 4461 | if len(sampled_requests) >= num_requests: |
| 4462 | break |
| 4463 | # Split the question text from options |
| 4464 | # (keep only the part before "Options:"). |
| 4465 | full_q: str = item.get("question", "") |
| 4466 | question_text = full_q.split("Options:", 1)[0].strip() |
| 4467 | |
| 4468 | # Multimodal image content. |
| 4469 | mm_content = process_image(item["image"]) |
| 4470 | |
| 4471 | # Compute prompt token length (note: this is plain text length |
| 4472 | # if enable_multimodal_chat is False). |
| 4473 | prompt_len = len(tokenizer(question_text).input_ids) |
| 4474 | |
| 4475 | prompt: str | list[dict] |
| 4476 | if enable_multimodal_chat: |
| 4477 | # If multimodal content should be embedded in the chat message, |
| 4478 | # convert to [{"role":"user","content":[...]}] |
| 4479 | prompt = self.apply_multimodal_chat_transformation( |
| 4480 | question_text, mm_content |
| 4481 | ) |
| 4482 | mm_for_request = None # Already embedded in chat content. |
| 4483 | else: |
| 4484 | # Default: prompt is plain text, |
| 4485 | # image is in mm_content for the bench to assemble. |
| 4486 | prompt = question_text |
| 4487 | mm_for_request = mm_content |
| 4488 | |
| 4489 | sampled_requests.append( |
| 4490 | SampleRequest( |
| 4491 | prompt=prompt, |
| 4492 | prompt_len=prompt_len, |
| 4493 | expected_output_len=output_len, |
| 4494 | multi_modal_data=mm_for_request, |
| 4495 | request_id=request_id_prefix + str(ind), |
| 4496 | ) |
| 4497 | ) |
| 4498 | |
| 4499 | self.maybe_oversample_requests( |
| 4500 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 4501 | ) |
| 4502 | return sampled_requests |
| 4503 | |
| 4504 | |
| 4505 | # ----------------------------------------------------------------------------- |
| 4506 | # BFCL (Berkeley Function Calling Leaderboard) Dataset Implementation |
| 4507 | # ----------------------------------------------------------------------------- |
| 4508 | |
| 4509 | |
| 4510 | class BFCLDataset(HuggingFaceDataset): |
| 4511 | """Berkeley Function Calling Leaderboard dataset. |
| 4512 | |
| 4513 | https://huggingface.co/datasets/gorilla-llm/Berkeley-Function-Calling-Leaderboard |
| 4514 | |
| 4515 | BFCL ships one JSON-lines file per category at the repo root (e.g. |
| 4516 | ``BFCL_v3_simple.json``, ``BFCL_v3_live_simple.json``) rather than a |
| 4517 | single HuggingFace split. Each record has ``{id, question, function}`` |
| 4518 | where ``function`` uses a non-OpenAI schema dialect (``"type": "dict"``). |
| 4519 | |
| 4520 | This dataset loader: |
| 4521 | - downloads the selected per-category files via ``hf_hub_download`` |
| 4522 | and interleaves rows round-robin so sampling is balanced |
| 4523 | - translates BFCL function schemas to OpenAI tool format |
| 4524 | - sets :attr:`SampleRequest.chat_messages` directly and attaches |
| 4525 | ``tools`` / ``tool_choice`` via :attr:`SampleRequest.request_overrides`, |
| 4526 | producing production-alike tool calling traffic when used with an |
| 4527 | ``openai-chat`` backend |
| 4528 | """ |
| 4529 | |
| 4530 | DEFAULT_OUTPUT_LEN = 512 |
| 4531 | DEFAULT_CATEGORIES = ("simple", "live_simple", "multiple") |
| 4532 | SUPPORTED_DATASET_PATHS = { |
| 4533 | "gorilla-llm/Berkeley-Function-Calling-Leaderboard", |
| 4534 | } |
| 4535 | IS_MULTIMODAL = False |
| 4536 | |
| 4537 | # BFCL primitive type names that are not valid JSON Schema types. |
| 4538 | # Map them to the closest JSON Schema equivalent so that grammar |
| 4539 | # backends (xgrammar, outlines) accept the translated tool schema. |
| 4540 | _TYPE_REMAP = { |
| 4541 | "dict": "object", |
| 4542 | "float": "number", |
| 4543 | "tuple": "array", |
| 4544 | "any": "string", |
| 4545 | } |
| 4546 | |
| 4547 | def load_data(self) -> None: |
| 4548 | """Defer loading to :meth:`sample` where categories are known.""" |
| 4549 | self.data = None |
| 4550 | |
| 4551 | def _resolve_categories(self, categories: list[str] | None) -> list[str]: |
| 4552 | if not categories: |
| 4553 | return list(self.DEFAULT_CATEGORIES) |
| 4554 | resolved: list[str] = [] |
| 4555 | for c in categories: |
| 4556 | c = c.strip() |
| 4557 | if not c: |
| 4558 | continue |
| 4559 | resolved.append(c) |
| 4560 | return resolved or list(self.DEFAULT_CATEGORIES) |
| 4561 | |
| 4562 | def _load_category(self, category: str) -> list[dict]: |
| 4563 | # Local import: huggingface_hub.errors is a small module and |
| 4564 | # importing at call site keeps module import cheap for users who |
| 4565 | # never touch BFCL. |
| 4566 | from huggingface_hub.errors import EntryNotFoundError |
| 4567 | |
| 4568 | filename = f"BFCL_v3_{category}.json" |
| 4569 | try: |
| 4570 | path = hf_api().hf_hub_download( |
| 4571 | self.dataset_path, filename, repo_type="dataset" |
| 4572 | ) |
| 4573 | except EntryNotFoundError as e: |
| 4574 | defaults = ", ".join(self.DEFAULT_CATEGORIES) |
| 4575 | raise ValueError( |
| 4576 | f"BFCL category '{category}' not found: file '{filename}' " |
| 4577 | f"does not exist in {self.dataset_path}. Check --bfcl-categories " |
| 4578 | f"(defaults: {defaults})." |
| 4579 | ) from e |
| 4580 | rows: list[dict] = [] |
| 4581 | with open(path) as f: |
| 4582 | for line in f: |
| 4583 | line = line.strip() |
| 4584 | if not line: |
| 4585 | continue |
| 4586 | rows.append(json.loads(line)) |
| 4587 | return rows |
| 4588 | |
| 4589 | @classmethod |
| 4590 | def _translate_schema(cls, node: Any) -> Any: |
| 4591 | """Recursively translate BFCL-flavored JSON schema to strict JSON Schema.""" |
| 4592 | if isinstance(node, dict): |
| 4593 | translated = {k: cls._translate_schema(v) for k, v in node.items()} |
| 4594 | t = translated.get("type") |
| 4595 | if isinstance(t, str) and t in cls._TYPE_REMAP: |
| 4596 | translated["type"] = cls._TYPE_REMAP[t] |
| 4597 | return translated |
| 4598 | if isinstance(node, list): |
| 4599 | return [cls._translate_schema(v) for v in node] |
| 4600 | return node |
| 4601 | |
| 4602 | @classmethod |
| 4603 | def _to_openai_tools(cls, functions: list[dict]) -> list[dict]: |
| 4604 | tools: list[dict] = [] |
| 4605 | for fn in functions: |
| 4606 | translated = cls._translate_schema(fn) |
| 4607 | tools.append({"type": "function", "function": translated}) |
| 4608 | return tools |
| 4609 | |
| 4610 | def sample( |
| 4611 | self, |
| 4612 | tokenizer: TokenizerLike, |
| 4613 | num_requests: int, |
| 4614 | request_id_prefix: str = "", |
| 4615 | no_oversample: bool = False, |
| 4616 | output_len: int | None = None, |
| 4617 | categories: list[str] | None = None, |
| 4618 | **kwargs, |
| 4619 | ) -> list[SampleRequest]: |
| 4620 | output_len = output_len if output_len is not None else self.DEFAULT_OUTPUT_LEN |
| 4621 | categories = self._resolve_categories(categories) |
| 4622 | |
| 4623 | per_category_rows: list[list[dict]] = [ |
| 4624 | self._load_category(c) for c in categories |
| 4625 | ] |
| 4626 | # Round-robin interleave so that when --disable-shuffle is set, |
| 4627 | # taking the first num_requests rows still yields balanced category |
| 4628 | # coverage. When shuffle is on (the default) this ordering is |
| 4629 | # randomized away, which is fine — the subsequent random sample is |
| 4630 | # already balanced in expectation. |
| 4631 | interleaved: list[dict] = [] |
| 4632 | max_len = max((len(rows) for rows in per_category_rows), default=0) |
| 4633 | for i in range(max_len): |
| 4634 | for rows in per_category_rows: |
| 4635 | if i < len(rows): |
| 4636 | interleaved.append(rows[i]) |
| 4637 | |
| 4638 | if not self.disable_shuffle: |
| 4639 | rng = random.Random(self.random_seed) |
| 4640 | rng.shuffle(interleaved) |
| 4641 | |
| 4642 | sampled_requests: list[SampleRequest] = [] |
| 4643 | for row in interleaved: |
| 4644 | if len(sampled_requests) >= num_requests: |
| 4645 | break |
| 4646 | question = row.get("question") |
| 4647 | functions = row.get("function") |
| 4648 | if not question or not functions: |
| 4649 | continue |
| 4650 | # BFCL question is list[list[dict]] — outer is turns. Use the |
| 4651 | # first turn only; skip multi-turn categories in this loader. |
| 4652 | if not isinstance(question, list) or not question: |
| 4653 | continue |
| 4654 | first_turn = question[0] |
| 4655 | if not isinstance(first_turn, list) or not first_turn: |
| 4656 | continue |
| 4657 | messages = first_turn |
| 4658 | if not isinstance(functions, list): |
| 4659 | functions = [functions] |
| 4660 | |
| 4661 | tools = self._to_openai_tools(functions) |
| 4662 | |
| 4663 | # Best-effort prompt length for percentile bucketing. Pass tools= |
| 4664 | # so modern chat templates (Llama 3.1+, Qwen, gpt-oss harmony, |
| 4665 | # Hermes) render the tool schemas — without this, the estimate |
| 4666 | # misses a significant chunk of the true input for BFCL traffic. |
| 4667 | # Older tokenizers reject the kwarg; fall back to tools-free. |
| 4668 | try: |
| 4669 | rendered = tokenizer.apply_chat_template( |
| 4670 | messages, |
| 4671 | tools=tools, |
| 4672 | tokenize=False, |
| 4673 | add_generation_prompt=True, |
| 4674 | ) |
| 4675 | except TypeError: |
| 4676 | rendered = tokenizer.apply_chat_template( |
| 4677 | messages, tokenize=False, add_generation_prompt=True |
| 4678 | ) |
| 4679 | except Exception as e: |
| 4680 | # Unexpected template failure — prompt_len will fall back to a |
| 4681 | # plain-text concatenation. Log so the degraded estimate is |
| 4682 | # visible instead of silently skewing latency buckets. |
| 4683 | logger.warning( |
| 4684 | "BFCL: apply_chat_template failed for a sample, falling " |
| 4685 | "back to plain-text prompt length: %s", |
| 4686 | e, |
| 4687 | exc_info=True, |
| 4688 | ) |
| 4689 | rendered = None |
| 4690 | if rendered is not None and isinstance(rendered, str): |
| 4691 | prompt_len = len(tokenizer(rendered).input_ids) |
| 4692 | else: |
| 4693 | text = "\n".join(m.get("content", "") for m in messages) |
| 4694 | prompt_len = len(tokenizer(text).input_ids) |
| 4695 | |
| 4696 | # The chat backend uses `messages` directly; `prompt` is only |
| 4697 | # kept as a fallback string for display/debug. |
| 4698 | prompt_text = messages[-1].get("content", "") if messages else "" |
| 4699 | |
| 4700 | sampled_requests.append( |
| 4701 | SampleRequest( |
| 4702 | prompt=prompt_text, |
| 4703 | prompt_len=prompt_len, |
| 4704 | expected_output_len=output_len, |
| 4705 | request_id=request_id_prefix + str(len(sampled_requests)), |
| 4706 | chat_messages=messages, |
| 4707 | request_overrides={ |
| 4708 | "tools": tools, |
| 4709 | "tool_choice": "auto", |
| 4710 | }, |
| 4711 | ) |
| 4712 | ) |
| 4713 | |
| 4714 | self.maybe_oversample_requests( |
| 4715 | sampled_requests, num_requests, request_id_prefix, no_oversample |
| 4716 | ) |
| 4717 | return sampled_requests |
| 4718 | |
| 4719 | |
| 4720 | # ----------------------------------------------------------------------------- |
| 4721 | # Speed Bench Dataset Implementation |
| 4722 | # ----------------------------------------------------------------------------- |
| 4723 | |
| 4724 | |
| 4725 | class SpeedBench(CustomDataset): |
| 4726 | """SPEED-Bench dataset: https://huggingface.co/datasets/nvidia/SPEED-Bench. |
| 4727 | |
| 4728 | Download the dataset using: |
| 4729 | |
| 4730 | `curl -LsSf https://raw.githubusercontent.com/NVIDIA-NeMo/Skills/refs/heads/main/nemo_skills/dataset/speed-bench/prepare.py | python3 -` |
| 4731 | """ # noqa: E501 |
| 4732 | |
| 4733 | DOWNLOAD_SCRIPT_URL = "https://raw.githubusercontent.com/NVIDIA-NeMo/Skills/refs/heads/main/nemo_skills/dataset/speed-bench/prepare.py" |
| 4734 | |
| 4735 | def __init__(self, **kwargs) -> None: |
| 4736 | self.dataset_subset = kwargs.pop("dataset_subset", "qualitative") |
| 4737 | self.category = kwargs.pop("category", None) |
| 4738 | super().__init__(**kwargs) |
| 4739 | self.load_data() |
| 4740 | |
| 4741 | def load_data(self) -> None: |
| 4742 | if self.dataset_path is None: |
| 4743 | raise ValueError("dataset_path must be provided for loading data.") |
| 4744 | |
| 4745 | if not Path(self.dataset_path).is_dir(): |
| 4746 | raise ValueError( |
| 4747 | f"dataset_path {self.dataset_path} is not a directory. " |
| 4748 | f"Please make sure to download the dataset from HuggingFace using " |
| 4749 | f"`curl -LsSf {self.DOWNLOAD_SCRIPT_URL} | python3 -`" |
| 4750 | ) |
| 4751 | |
| 4752 | self.data = [] |
| 4753 | |
| 4754 | # Load the JSONL file |
| 4755 | jsonl_data = pd.read_json( |
| 4756 | path_or_buf=Path(self.dataset_path) / f"{self.dataset_subset}.jsonl", |
| 4757 | lines=True, |
| 4758 | ) |
| 4759 | |
| 4760 | # check if the JSONL file has a 'turns' column |
| 4761 | if "messages" not in jsonl_data.columns: |
| 4762 | raise ValueError( |
| 4763 | "JSONL file must contain a 'messages' column. " |
| 4764 | "Please make sure to download the dataset from HuggingFace using " |
| 4765 | f"`curl -LsSf {self.DOWNLOAD_SCRIPT_URL} | python3 -`" |
| 4766 | ) |
| 4767 | |
| 4768 | for _, row in jsonl_data.iterrows(): |
| 4769 | # sample only from a specific category if specified |
| 4770 | if (not self.category) or (self.category == row["category"]): |
| 4771 | prompt = row["messages"][0]["content"] |
| 4772 | self.data.append({"prompt": prompt}) |
| 4773 | |
| 4774 | random.seed(self.random_seed) |
| 4775 | if not getattr(self, "disable_shuffle", False): |
| 4776 | random.shuffle(self.data) |