| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import argparse |
| 5 | import copy |
| 6 | import dataclasses |
| 7 | import functools |
| 8 | import json |
| 9 | import os |
| 10 | import sys |
| 11 | import warnings |
| 12 | from collections.abc import Callable |
| 13 | from dataclasses import MISSING, asdict, dataclass, fields, is_dataclass |
| 14 | from types import UnionType |
| 15 | from typing import ( |
| 16 | TYPE_CHECKING, |
| 17 | Annotated, |
| 18 | Any, |
| 19 | Literal, |
| 20 | TypeAlias, |
| 21 | TypeVar, |
| 22 | Union, |
| 23 | cast, |
| 24 | get_args, |
| 25 | get_origin, |
| 26 | is_typeddict, |
| 27 | ) |
| 28 | |
| 29 | import huggingface_hub |
| 30 | import regex as re |
| 31 | import torch |
| 32 | from pydantic import TypeAdapter, ValidationError |
| 33 | from pydantic.fields import FieldInfo |
| 34 | from typing_extensions import TypeIs |
| 35 | |
| 36 | import vllm.envs as envs |
| 37 | from vllm.config import ( |
| 38 | AttentionConfig, |
| 39 | AuxOutputConfig, |
| 40 | CacheConfig, |
| 41 | CompilationConfig, |
| 42 | ConfigType, |
| 43 | DeviceConfig, |
| 44 | DiffusionConfig, |
| 45 | ECTransferConfig, |
| 46 | EncoderCacheManagerConfig, |
| 47 | EngramConfig, |
| 48 | EPLBConfig, |
| 49 | FaultToleranceConfig, |
| 50 | KernelConfig, |
| 51 | KVEventsConfig, |
| 52 | KVTransferConfig, |
| 53 | LoadConfig, |
| 54 | LoggingConfig, |
| 55 | LoRAConfig, |
| 56 | MambaConfig, |
| 57 | ModelConfig, |
| 58 | MultiModalConfig, |
| 59 | ObservabilityConfig, |
| 60 | OffloadConfig, |
| 61 | ParallelConfig, |
| 62 | PoolerConfig, |
| 63 | PrefetchOffloadConfig, |
| 64 | ProfilerConfig, |
| 65 | ReasoningConfig, |
| 66 | SchedulerConfig, |
| 67 | SpeculativeConfig, |
| 68 | StructuredOutputsConfig, |
| 69 | UVAOffloadConfig, |
| 70 | VllmConfig, |
| 71 | WeightTransferConfig, |
| 72 | get_attr_docs, |
| 73 | ) |
| 74 | from vllm.config.cache import ( |
| 75 | CacheDType, |
| 76 | KVOffloadingBackend, |
| 77 | MambaCacheMode, |
| 78 | MambaDType, |
| 79 | PrefixCachingHashAlgo, |
| 80 | ) |
| 81 | from vllm.config.device import Device |
| 82 | from vllm.config.kernel import ( |
| 83 | IrOpPriorityConfig, |
| 84 | LinearBackend, |
| 85 | MoEBackend, |
| 86 | SparseIndexerTopkBackend, |
| 87 | ) |
| 88 | from vllm.config.load import SafetensorsLoadStrategy |
| 89 | from vllm.config.logging import LogLevel |
| 90 | from vllm.config.lora import MaxLoRARanks |
| 91 | from vllm.config.mamba import MambaBackendEnum, MambaSSUAlgorithm |
| 92 | from vllm.config.model import ( |
| 93 | ConvertOption, |
| 94 | HfOverrides, |
| 95 | LogprobsMode, |
| 96 | ModelDType, |
| 97 | RunnerOption, |
| 98 | TokenizerMode, |
| 99 | ) |
| 100 | from vllm.config.multimodal import ( |
| 101 | MMCacheType, |
| 102 | MMEncoderTPMode, |
| 103 | MMHasherAlgorithm, |
| 104 | MMProcessorDevice, |
| 105 | MMTensorIPC, |
| 106 | ) |
| 107 | from vllm.config.observability import DetailedTraceModules |
| 108 | from vllm.config.parallel import ( |
| 109 | All2AllBackend, |
| 110 | DataParallelBackend, |
| 111 | DCPCommBackend, |
| 112 | DistributedExecutorBackend, |
| 113 | ExpertPlacementStrategy, |
| 114 | ) |
| 115 | from vllm.config.scheduler import SchedulerPolicy |
| 116 | from vllm.config.utils import get_field |
| 117 | from vllm.config.vllm import OptimizationLevel, PerformanceMode |
| 118 | from vllm.config.watermarking import WatermarkConfig |
| 119 | from vllm.logger import configure_logging_if_needed, init_logger, suppress_logging |
| 120 | from vllm.platforms import CpuArchEnum, current_platform |
| 121 | from vllm.plugins import load_general_plugins |
| 122 | from vllm.ray.lazy_utils import is_in_ray_actor, is_ray_initialized |
| 123 | from vllm.transformers_utils.config import maybe_override_with_speculators |
| 124 | from vllm.transformers_utils.repo_utils import get_model_path |
| 125 | from vllm.transformers_utils.utils import is_cloud_storage |
| 126 | from vllm.utils.argparse_utils import ( |
| 127 | FlexibleArgumentParser, |
| 128 | human_readable_int, |
| 129 | human_readable_int_or_auto, |
| 130 | ) |
| 131 | from vllm.utils.mem_constants import GiB_bytes |
| 132 | from vllm.utils.network_utils import get_ip |
| 133 | from vllm.utils.torch_utils import resolve_kv_cache_dtype_string |
| 134 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 135 | from vllm.v1.sample.logits_processor import LogitsProcessor |
| 136 | from vllm.version import __version__ as VLLM_VERSION |
| 137 | |
| 138 | if TYPE_CHECKING: |
| 139 | from vllm.config.quantization import QuantizationConfigArgs |
| 140 | from vllm.model_executor.layers.quantization import QuantizationMethods |
| 141 | from vllm.model_executor.model_loader import LoadFormats |
| 142 | from vllm.usage.usage_lib import UsageContext |
| 143 | from vllm.v1.executor import Executor |
| 144 | else: |
| 145 | Executor = Any |
| 146 | QuantizationMethods = str |
| 147 | LoadFormats = str |
| 148 | UsageContext = Any |
| 149 | |
| 150 | |
| 151 | logger = init_logger(__name__) |
| 152 | |
| 153 | # object is used to allow for special typing forms |
| 154 | T = TypeVar("T") |
| 155 | TypeHint: TypeAlias = type[Any] | object |
| 156 | TypeHintT: TypeAlias = type[T] | object |
| 157 | |
| 158 | |
| 159 | def parse_type(return_type: Callable[[str], T]) -> Callable[[str], T]: |
| 160 | def _parse_type(val: str) -> T: |
| 161 | try: |
| 162 | return return_type(val) |
| 163 | except ValueError as e: |
| 164 | raise argparse.ArgumentTypeError( |
| 165 | f"Value {val} cannot be converted to {return_type}." |
| 166 | ) from e |
| 167 | |
| 168 | return _parse_type |
| 169 | |
| 170 | |
| 171 | def optional_type(return_type: Callable[[str], T]) -> Callable[[str], T | None]: |
| 172 | def _optional_type(val: str) -> T | None: |
| 173 | if val == "" or val == "None": |
| 174 | return None |
| 175 | return parse_type(return_type)(val) |
| 176 | |
| 177 | return _optional_type |
| 178 | |
| 179 | |
| 180 | def union_dict_and_str(val: str) -> str | dict[str, str] | None: |
| 181 | if not re.match(r"(?s)^\s*{.*}\s*$", val): |
| 182 | return str(val) |
| 183 | return optional_type(json.loads)(val) |
| 184 | |
| 185 | |
| 186 | def is_type(type_hint: TypeHint, type: TypeHintT) -> TypeIs[TypeHintT]: |
| 187 | """Check if the type hint is a specific type.""" |
| 188 | return type_hint is type or get_origin(type_hint) is type |
| 189 | |
| 190 | |
| 191 | def is_dict_subclass(type_hint: TypeHint) -> bool: |
| 192 | """Check if the type hint is a subclass of `dict`. |
| 193 | |
| 194 | `TypedDict`s are excluded because they do not support class checks. |
| 195 | """ |
| 196 | return ( |
| 197 | isinstance(type_hint, type) |
| 198 | and not is_typeddict(type_hint) |
| 199 | and issubclass(type_hint, dict) |
| 200 | ) |
| 201 | |
| 202 | |
| 203 | def contains_type(type_hints: set[TypeHint], type: TypeHintT) -> bool: |
| 204 | """Check if the type hints contain a specific type.""" |
| 205 | return any(is_type(type_hint, type) for type_hint in type_hints) |
| 206 | |
| 207 | |
| 208 | def get_type(type_hints: set[TypeHint], type: TypeHintT) -> TypeHintT: |
| 209 | """Get the specific type from the type hints.""" |
| 210 | return next((th for th in type_hints if is_type(th, type)), None) |
| 211 | |
| 212 | |
| 213 | def literal_to_kwargs(type_hints: set[TypeHint]) -> dict[str, Any]: |
| 214 | """Get the `type` and `choices` from a `Literal` type hint in `type_hints`. |
| 215 | |
| 216 | If `type_hints` also contains `str`, we use `metavar` instead of `choices`. |
| 217 | """ |
| 218 | type_hint = get_type(type_hints, Literal) |
| 219 | options = get_args(type_hint) |
| 220 | option_type = type(options[0]) |
| 221 | if not all(isinstance(option, option_type) for option in options): |
| 222 | raise ValueError( |
| 223 | "All options must be of the same type. " |
| 224 | f"Got {options} with types {[type(c) for c in options]}" |
| 225 | ) |
| 226 | kwarg = "metavar" if contains_type(type_hints, str) else "choices" |
| 227 | return {"type": option_type, kwarg: sorted(options)} |
| 228 | |
| 229 | |
| 230 | def collection_to_kwargs(type_hints: set[TypeHint], type: TypeHint) -> dict[str, Any]: |
| 231 | type_hint = get_type(type_hints, type) |
| 232 | types = get_args(type_hint) |
| 233 | elem_type = types[0] |
| 234 | |
| 235 | # Handle Ellipsis |
| 236 | assert all(t is elem_type for t in types if t is not Ellipsis), ( |
| 237 | f"All non-Ellipsis elements must be of the same type. Got {types}." |
| 238 | ) |
| 239 | |
| 240 | # Handle Union types |
| 241 | if get_origin(elem_type) in {Union, UnionType}: |
| 242 | # Union for Union[X, Y] and UnionType for X | Y |
| 243 | assert str in get_args(elem_type), ( |
| 244 | "If element can have multiple types, one must be 'str' " |
| 245 | f"(i.e. 'list[int | str]'). Got {elem_type}." |
| 246 | ) |
| 247 | elem_type = str |
| 248 | |
| 249 | return { |
| 250 | "type": elem_type, |
| 251 | "nargs": "+" if type is not tuple or Ellipsis in types else len(types), |
| 252 | } |
| 253 | |
| 254 | |
| 255 | def is_not_builtin(type_hint: TypeHint) -> bool: |
| 256 | """Check if the class is not a built-in type.""" |
| 257 | return type_hint.__module__ != "builtins" |
| 258 | |
| 259 | |
| 260 | def get_type_hints(type_hint: TypeHint) -> set[TypeHint]: |
| 261 | """Extract type hints from Annotated or Union type hints.""" |
| 262 | type_hints: set[TypeHint] = set() |
| 263 | origin = get_origin(type_hint) |
| 264 | args = get_args(type_hint) |
| 265 | |
| 266 | if origin is Annotated: |
| 267 | type_hints.update(get_type_hints(args[0])) |
| 268 | elif origin in {Union, UnionType}: |
| 269 | # Union for Union[X, Y] and UnionType for X | Y |
| 270 | for arg in args: |
| 271 | type_hints.update(get_type_hints(arg)) |
| 272 | else: |
| 273 | type_hints.add(type_hint) |
| 274 | |
| 275 | return type_hints |
| 276 | |
| 277 | |
| 278 | NEEDS_HELP = ( |
| 279 | any("--help" in arg for arg in sys.argv) # vllm SUBCOMMAND --help |
| 280 | or "mkdocs" in sys.modules # mkdocs SUBCOMMAND |
| 281 | ) |
| 282 | |
| 283 | |
| 284 | def _maybe_add_docs_url(cls: Any) -> str: |
| 285 | """Generate API docs URL for a vllm config class.""" |
| 286 | import vllm.config |
| 287 | |
| 288 | name = cls.__name__ |
| 289 | if getattr(vllm.config, name, None) is not cls: |
| 290 | return "" |
| 291 | version = f"v{VLLM_VERSION}" if "dev" not in VLLM_VERSION else "latest" |
| 292 | return f"\n\nAPI docs: https://docs.vllm.ai/en/{version}/api/vllm/config/#vllm.config.{name}" |
| 293 | |
| 294 | |
| 295 | def _expand_json_human_readable_numbers(val: str) -> str: |
| 296 | """Expand human-readable number suffixes in a JSON string. |
| 297 | |
| 298 | Based on :func:`human_readable_int` so that the ``k/m/g/t`` (decimal) and |
| 299 | ``K/M/G/T`` (binary) conventions work out the box. |
| 300 | Also works inside JSON config arguments such |
| 301 | as ``--kv-transfer-config '{"cpu_bytes_to_use": 80m}'``. |
| 302 | |
| 303 | Only bare (unquoted) tokens are replaced so that JSON string values |
| 304 | like ``"model_name"`` are never modified. |
| 305 | """ |
| 306 | # Split on quoted strings so we only touch non-string regions. |
| 307 | parts = re.split(r'("(?:[^"\\]|\\.)*")', val) |
| 308 | for i in range(0, len(parts), 2): # even indices = outside strings |
| 309 | parts[i] = re.sub( |
| 310 | r"\b\d+(?:\.\d+)?[kKmMgGtT]\b", |
| 311 | lambda m: str(human_readable_int(m.group())), |
| 312 | parts[i], |
| 313 | ) |
| 314 | return "".join(parts) |
| 315 | |
| 316 | |
| 317 | @functools.lru_cache(maxsize=30) |
| 318 | def _compute_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]: |
| 319 | # Save time only getting attr docs if we're generating help text |
| 320 | cls_docs = get_attr_docs(cls) if NEEDS_HELP else {} |
| 321 | kwargs = {} |
| 322 | for field in fields(cls): |
| 323 | # Get the set of possible types for the field |
| 324 | type_hints: set[TypeHint] = get_type_hints(field.type) |
| 325 | # Subclasses of dict (e.g. MultiModalDummyOptions) are CLI dicts |
| 326 | type_hints = {dict if is_dict_subclass(th) else th for th in type_hints} |
| 327 | |
| 328 | # If the field is a dataclass, we can use the model_validate_json |
| 329 | generator = (th for th in type_hints if is_dataclass(th)) |
| 330 | dataclass_cls = next(generator, None) |
| 331 | |
| 332 | # Get the default value of the field |
| 333 | if field.default is not MISSING: |
| 334 | default = field.default |
| 335 | # Handle pydantic.Field defaults |
| 336 | if isinstance(default, FieldInfo): |
| 337 | if default.default_factory is None: |
| 338 | default = default.default |
| 339 | else: |
| 340 | # VllmConfig's Fields have default_factory set to config classes. |
| 341 | # These could emit logs on init, which would be confusing. |
| 342 | with suppress_logging(): |
| 343 | default = default.default_factory() # type: ignore[call-arg] |
| 344 | elif field.default_factory is not MISSING: |
| 345 | default = field.default_factory() |
| 346 | |
| 347 | # Get the help text for the field |
| 348 | name = field.name |
| 349 | help = cls_docs.get(name, "").strip() |
| 350 | # Escape % for argparse |
| 351 | help = help.replace("%", "%%") |
| 352 | |
| 353 | # Initialise the kwargs dictionary for the field |
| 354 | kwargs[name] = {"default": default, "help": help} |
| 355 | |
| 356 | # Set other kwargs based on the type hints |
| 357 | json_tip = ( |
| 358 | "Should either be a valid JSON string or JSON keys passed individually." |
| 359 | ) |
| 360 | if dataclass_cls is not None: |
| 361 | |
| 362 | def parse_dataclass(val: str, cls=dataclass_cls) -> Any: |
| 363 | try: |
| 364 | val = _expand_json_human_readable_numbers(val) |
| 365 | return TypeAdapter(cls).validate_json(val) |
| 366 | except ValidationError as e: |
| 367 | raise argparse.ArgumentTypeError(repr(e)) from e |
| 368 | |
| 369 | kwargs[name]["type"] = parse_dataclass |
| 370 | kwargs[name]["help"] += _maybe_add_docs_url(dataclass_cls) |
| 371 | kwargs[name]["help"] += f"\n\n{json_tip}" |
| 372 | elif type_hints == {bool, str, type(None)}: |
| 373 | # Optional-valued flag: bare flag -> True, value -> str. |
| 374 | kwargs[name]["type"] = str |
| 375 | kwargs[name]["nargs"] = "?" |
| 376 | kwargs[name]["const"] = True |
| 377 | elif contains_type(type_hints, bool): |
| 378 | # Creates --no-<name> and --<name> flags |
| 379 | kwargs[name]["action"] = argparse.BooleanOptionalAction |
| 380 | elif contains_type(type_hints, Literal): |
| 381 | kwargs[name].update(literal_to_kwargs(type_hints)) |
| 382 | elif contains_type(type_hints, tuple): |
| 383 | kwargs[name].update(collection_to_kwargs(type_hints, tuple)) |
| 384 | elif contains_type(type_hints, list): |
| 385 | kwargs[name].update(collection_to_kwargs(type_hints, list)) |
| 386 | elif contains_type(type_hints, set): |
| 387 | kwargs[name].update(collection_to_kwargs(type_hints, set)) |
| 388 | elif contains_type(type_hints, int): |
| 389 | # Arguments that accept human-readable integer strings (e.g., 1K, 2M, 1G) |
| 390 | human_readable_int_args = { |
| 391 | "max_num_batched_tokens", |
| 392 | "max_num_scheduled_tokens", |
| 393 | "kv_cache_memory_bytes", |
| 394 | "safetensors_prefetch_block_size", |
| 395 | "max_num_queued_tokens", |
| 396 | } |
| 397 | if name == "max_model_len": |
| 398 | kwargs[name]["type"] = human_readable_int_or_auto |
| 399 | kwargs[name]["help"] += f"\n\n{human_readable_int_or_auto.__doc__}" |
| 400 | elif name in human_readable_int_args: |
| 401 | kwargs[name]["type"] = human_readable_int |
| 402 | kwargs[name]["help"] += f"\n\n{human_readable_int.__doc__}" |
| 403 | else: |
| 404 | kwargs[name]["type"] = int |
| 405 | elif contains_type(type_hints, float): |
| 406 | kwargs[name]["type"] = float |
| 407 | elif contains_type(type_hints, dict) and ( |
| 408 | contains_type(type_hints, str) |
| 409 | or any(is_not_builtin(th) for th in type_hints) |
| 410 | ): |
| 411 | kwargs[name]["type"] = union_dict_and_str |
| 412 | elif contains_type(type_hints, dict): |
| 413 | kwargs[name]["type"] = parse_type(json.loads) |
| 414 | kwargs[name]["help"] += f"\n\n{json_tip}" |
| 415 | elif contains_type(type_hints, str) or any( |
| 416 | is_not_builtin(th) for th in type_hints |
| 417 | ): |
| 418 | kwargs[name]["type"] = str |
| 419 | else: |
| 420 | raise ValueError(f"Unsupported type {type_hints} for argument {name}.") |
| 421 | |
| 422 | # If the type hint was a sequence of literals, use the helper function |
| 423 | # to update the type and choices |
| 424 | if get_origin(kwargs[name].get("type")) is Literal: |
| 425 | kwargs[name].update(literal_to_kwargs({kwargs[name]["type"]})) |
| 426 | |
| 427 | # If None is in type_hints, make the argument optional. |
| 428 | # But not if it's a bool, argparse will handle this better. |
| 429 | if type(None) in type_hints and not contains_type(type_hints, bool): |
| 430 | kwargs[name]["type"] = optional_type(kwargs[name]["type"]) |
| 431 | if kwargs[name].get("choices"): |
| 432 | kwargs[name]["choices"].append(None) |
| 433 | return kwargs |
| 434 | |
| 435 | |
| 436 | def get_kwargs(cls: ConfigType) -> dict[str, dict[str, Any]]: |
| 437 | """Return argparse kwargs for the given Config dataclass. |
| 438 | |
| 439 | If `--help` or `mkdocs` are not present in the command line command, the |
| 440 | attribute documentation will not be included in the help output. |
| 441 | |
| 442 | The heavy computation is cached via functools.lru_cache, and a deep copy |
| 443 | is returned so callers can mutate the dictionary without affecting the |
| 444 | cached version. |
| 445 | """ |
| 446 | return copy.deepcopy(_compute_kwargs(cls)) |
| 447 | |
| 448 | |
| 449 | _LOG_CONFIG_FILE_DEPRECATION_MESSAGE = ( |
| 450 | "--log-config-file is deprecated and will be removed in v0.33.0. " |
| 451 | "Use --logging-config.pylogging_config_file instead." |
| 452 | ) |
| 453 | |
| 454 | |
| 455 | class DeprecatedLogConfigFileAction(argparse.Action): |
| 456 | """Warn when the legacy ``--log-config-file`` option is used.""" |
| 457 | |
| 458 | def __call__(self, parser, namespace, values, option_string=None): |
| 459 | warnings.warn(_LOG_CONFIG_FILE_DEPRECATION_MESSAGE, stacklevel=1) |
| 460 | setattr(namespace, self.dest, values) |
| 461 | |
| 462 | |
| 463 | @dataclass |
| 464 | class EngineArgs: |
| 465 | """Arguments for vLLM engine.""" |
| 466 | |
| 467 | model: str = ModelConfig.model |
| 468 | # Public compatibility argument. Canonical runtime state lives in |
| 469 | # AuxOutputConfig. |
| 470 | enable_return_routed_experts: bool = False |
| 471 | return_sampling_mask: bool = ModelConfig.return_sampling_mask |
| 472 | model_weights: str = ModelConfig.model_weights |
| 473 | served_model_name: str | list[str] | None = ModelConfig.served_model_name |
| 474 | tokenizer: str | None = ModelConfig.tokenizer |
| 475 | hf_config_path: str | None = ModelConfig.hf_config_path |
| 476 | runner: RunnerOption = ModelConfig.runner |
| 477 | convert: ConvertOption = ModelConfig.convert |
| 478 | skip_tokenizer_init: bool = ModelConfig.skip_tokenizer_init |
| 479 | enable_prompt_embeds: bool = ModelConfig.enable_prompt_embeds |
| 480 | tokenizer_mode: TokenizerMode | str = ModelConfig.tokenizer_mode |
| 481 | trust_remote_code: bool = ModelConfig.trust_remote_code |
| 482 | allowed_local_media_path: str = ModelConfig.allowed_local_media_path |
| 483 | allowed_media_domains: list[str] | None = ModelConfig.allowed_media_domains |
| 484 | download_dir: str | None = LoadConfig.download_dir |
| 485 | safetensors_load_strategy: SafetensorsLoadStrategy | None = ( |
| 486 | LoadConfig.safetensors_load_strategy |
| 487 | ) |
| 488 | safetensors_prefetch_num_threads: int = LoadConfig.safetensors_prefetch_num_threads |
| 489 | safetensors_prefetch_block_size: int = LoadConfig.safetensors_prefetch_block_size |
| 490 | load_format: str | LoadFormats = LoadConfig.load_format |
| 491 | config_format: str = ModelConfig.config_format |
| 492 | dtype: ModelDType = ModelConfig.dtype |
| 493 | kv_cache_dtype: CacheDType = CacheConfig.cache_dtype |
| 494 | seed: int = ModelConfig.seed |
| 495 | max_model_len: int = ModelConfig.max_model_len |
| 496 | cudagraph_capture_sizes: list[int] | None = ( |
| 497 | CompilationConfig.cudagraph_capture_sizes |
| 498 | ) |
| 499 | max_cudagraph_capture_size: int | None = get_field( |
| 500 | CompilationConfig, "max_cudagraph_capture_size" |
| 501 | ) |
| 502 | ir_op_priority: IrOpPriorityConfig = get_field(KernelConfig, "ir_op_priority") |
| 503 | # Note: Specifying a custom executor backend by passing a class |
| 504 | # is intended for expert use only. The API may change without |
| 505 | # notice. |
| 506 | distributed_executor_backend: ( |
| 507 | str | DistributedExecutorBackend | type[Executor] | None |
| 508 | ) = ParallelConfig.distributed_executor_backend |
| 509 | # number of P/D disaggregation (or other disaggregation) workers |
| 510 | pipeline_parallel_size: int = ParallelConfig.pipeline_parallel_size |
| 511 | master_addr: str = ParallelConfig.master_addr |
| 512 | master_port: int = ParallelConfig.master_port |
| 513 | nnodes: int = ParallelConfig.nnodes |
| 514 | node_rank: int = ParallelConfig.node_rank |
| 515 | distributed_timeout_seconds: int | None = ParallelConfig.distributed_timeout_seconds |
| 516 | cpu_distributed_timeout_seconds: int | None = ( |
| 517 | ParallelConfig.cpu_distributed_timeout_seconds |
| 518 | ) |
| 519 | numa_bind: bool = ParallelConfig.numa_bind |
| 520 | numa_bind_nodes: list[int] | None = ParallelConfig.numa_bind_nodes |
| 521 | numa_bind_cpus: list[str] | None = ParallelConfig.numa_bind_cpus |
| 522 | device_ids: list[int | str] | None = None |
| 523 | tensor_parallel_size: int = ParallelConfig.tensor_parallel_size |
| 524 | prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size |
| 525 | decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size |
| 526 | dcp_comm_backend: DCPCommBackend | None = ParallelConfig.dcp_comm_backend |
| 527 | dcp_q_replicate: bool | None = ParallelConfig.dcp_q_replicate |
| 528 | dcp_kv_cache_interleave_size: int = ParallelConfig.dcp_kv_cache_interleave_size |
| 529 | cp_kv_cache_interleave_size: int | None = None |
| 530 | data_parallel_size: int = ParallelConfig.data_parallel_size |
| 531 | data_parallel_rank: int | None = None |
| 532 | data_parallel_start_rank: int | None = None |
| 533 | data_parallel_size_local: int | None = None |
| 534 | data_parallel_address: str | None = None |
| 535 | data_parallel_rpc_port: int | None = None |
| 536 | data_parallel_hybrid_lb: bool = False |
| 537 | data_parallel_external_lb: bool = False |
| 538 | data_parallel_multi_port_external_lb: bool = False |
| 539 | data_parallel_backend: DataParallelBackend = ParallelConfig.data_parallel_backend |
| 540 | enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel |
| 541 | enable_batch_sharded_sampling: bool | None = ( |
| 542 | ParallelConfig.enable_batch_sharded_sampling |
| 543 | ) |
| 544 | enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter |
| 545 | moe_backend: MoEBackend = KernelConfig.moe_backend |
| 546 | linear_backend: LinearBackend = KernelConfig.linear_backend |
| 547 | sparse_indexer_topk_backend: SparseIndexerTopkBackend = ( |
| 548 | KernelConfig.sparse_indexer_topk_backend |
| 549 | ) |
| 550 | all2all_backend: All2AllBackend = ParallelConfig.all2all_backend |
| 551 | enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep |
| 552 | elastic_ep_max_dp_size: int = ParallelConfig.elastic_ep_max_dp_size |
| 553 | enable_dbo: bool = ParallelConfig.enable_dbo |
| 554 | ubatch_size: int = ParallelConfig.ubatch_size |
| 555 | dbo_decode_token_threshold: int = ParallelConfig.dbo_decode_token_threshold |
| 556 | dp_sync_interval: int = ParallelConfig.dp_sync_interval |
| 557 | dbo_prefill_token_threshold: int = ParallelConfig.dbo_prefill_token_threshold |
| 558 | disable_nccl_for_dp_synchronization: bool | None = ( |
| 559 | ParallelConfig.disable_nccl_for_dp_synchronization |
| 560 | ) |
| 561 | eplb_config: EPLBConfig = get_field(ParallelConfig, "eplb_config") |
| 562 | enable_eplb: bool = ParallelConfig.enable_eplb |
| 563 | expert_placement_strategy: ExpertPlacementStrategy = ( |
| 564 | ParallelConfig.expert_placement_strategy |
| 565 | ) |
| 566 | _api_process_count: int = ParallelConfig._api_process_count |
| 567 | _api_process_rank: int = ParallelConfig._api_process_rank |
| 568 | max_parallel_loading_workers: int | None = ( |
| 569 | ParallelConfig.max_parallel_loading_workers |
| 570 | ) |
| 571 | block_size: int | None = None |
| 572 | enable_prefix_caching: bool | None = None |
| 573 | prefix_caching_hash_algo: PrefixCachingHashAlgo = ( |
| 574 | CacheConfig.prefix_caching_hash_algo |
| 575 | ) |
| 576 | prefix_cache_retention_interval: int | None = get_field( |
| 577 | CacheConfig, "prefix_cache_retention_interval" |
| 578 | ) |
| 579 | disable_sliding_window: bool = ModelConfig.disable_sliding_window |
| 580 | disable_cascade_attn: bool = ModelConfig.disable_cascade_attn |
| 581 | offload_backend: str = OffloadConfig.offload_backend |
| 582 | cpu_offload_gb: float = UVAOffloadConfig.cpu_offload_gb |
| 583 | cpu_offload_params: set[str] = get_field(UVAOffloadConfig, "cpu_offload_params") |
| 584 | offload_group_size: int = PrefetchOffloadConfig.offload_group_size |
| 585 | offload_num_in_group: int = PrefetchOffloadConfig.offload_num_in_group |
| 586 | offload_prefetch_step: int = PrefetchOffloadConfig.offload_prefetch_step |
| 587 | offload_params: set[str] = get_field(PrefetchOffloadConfig, "offload_params") |
| 588 | gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization |
| 589 | kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes |
| 590 | max_num_batched_tokens: int | None = None |
| 591 | max_num_scheduled_tokens: int | None = None |
| 592 | long_prefill_token_threshold: int = SchedulerConfig.long_prefill_token_threshold |
| 593 | long_prefill_token_threshold_adaptive: bool = ( |
| 594 | SchedulerConfig.long_prefill_token_threshold_adaptive |
| 595 | ) |
| 596 | max_num_seqs: int | None = None |
| 597 | max_num_active_seqs: int | None = SchedulerConfig.max_num_active_seqs |
| 598 | max_num_queued_reqs: int | None = None |
| 599 | max_num_queued_tokens: int | None = None |
| 600 | max_logprobs: int = ModelConfig.max_logprobs |
| 601 | logprobs_mode: LogprobsMode = ModelConfig.logprobs_mode |
| 602 | use_fp64_gumbel: bool = ModelConfig.use_fp64_gumbel |
| 603 | enable_trace_replay: bool = ModelConfig.enable_trace_replay |
| 604 | disable_log_stats: bool = False |
| 605 | aggregate_engine_logging: bool = False |
| 606 | revision: str | None = ModelConfig.revision |
| 607 | code_revision: str | None = ModelConfig.code_revision |
| 608 | hf_token: bool | str | None = ModelConfig.hf_token |
| 609 | hf_overrides: HfOverrides = get_field(ModelConfig, "hf_overrides") |
| 610 | model_class_overrides: dict[str, str] = get_field( |
| 611 | ModelConfig, "model_class_overrides" |
| 612 | ) |
| 613 | tokenizer_revision: str | None = ModelConfig.tokenizer_revision |
| 614 | quantization: QuantizationMethods | str | None = ModelConfig.quantization |
| 615 | quantization_config: "dict[str, Any] | QuantizationConfigArgs | None" = None |
| 616 | """User-facing quantization configuration. Carries per-layer-kind |
| 617 | QuantSpecs (linear, moe) and ignore patterns; see |
| 618 | :class:`QuantizationConfigArgs`. Auto-populated from the matching online |
| 619 | shorthand when `quantization` is one of the values in |
| 620 | `ONLINE_QUANT_SHORTHAND_NAMES`.""" |
| 621 | allow_deprecated_quantization: bool = ModelConfig.allow_deprecated_quantization |
| 622 | enforce_eager: bool = ModelConfig.enforce_eager |
| 623 | disable_custom_all_reduce: bool = ParallelConfig.disable_custom_all_reduce |
| 624 | language_model_only: bool = MultiModalConfig.language_model_only |
| 625 | limit_mm_per_prompt: dict[str, int | dict[str, int]] = get_field( |
| 626 | MultiModalConfig, "limit_per_prompt" |
| 627 | ) |
| 628 | enable_mm_embeds: bool = MultiModalConfig.enable_mm_embeds |
| 629 | interleave_mm_strings: bool = MultiModalConfig.interleave_mm_strings |
| 630 | media_io_kwargs: dict[str, dict[str, Any]] = get_field( |
| 631 | MultiModalConfig, "media_io_kwargs" |
| 632 | ) |
| 633 | mm_processor_kwargs: dict[str, Any] | None = MultiModalConfig.mm_processor_kwargs |
| 634 | mm_processor_cache_gb: float = MultiModalConfig.mm_processor_cache_gb |
| 635 | mm_processor_cache_type: MMCacheType | None = ( |
| 636 | MultiModalConfig.mm_processor_cache_type |
| 637 | ) |
| 638 | mm_hasher_algorithm: MMHasherAlgorithm = get_field( |
| 639 | MultiModalConfig, "mm_hasher_algorithm" |
| 640 | ) |
| 641 | mm_shm_cache_max_object_size_mb: int = ( |
| 642 | MultiModalConfig.mm_shm_cache_max_object_size_mb |
| 643 | ) |
| 644 | mm_encoder_only: bool = MultiModalConfig.mm_encoder_only |
| 645 | mm_encoder_tp_mode: MMEncoderTPMode = MultiModalConfig.mm_encoder_tp_mode |
| 646 | mm_encoder_attn_backend: AttentionBackendEnum | str | None = ( |
| 647 | MultiModalConfig.mm_encoder_attn_backend |
| 648 | ) |
| 649 | mm_encoder_attn_dtype: str | None = MultiModalConfig.mm_encoder_attn_dtype |
| 650 | mm_encoder_fp8_scale_path: str | None = MultiModalConfig.mm_encoder_fp8_scale_path |
| 651 | mm_encoder_fp8_scale_save_path: str | None = ( |
| 652 | MultiModalConfig.mm_encoder_fp8_scale_save_path |
| 653 | ) |
| 654 | mm_encoder_fp8_scale_save_margin: float = ( |
| 655 | MultiModalConfig.mm_encoder_fp8_scale_save_margin |
| 656 | ) |
| 657 | io_processor_plugin: str | None = None |
| 658 | renderer_num_workers: int = 1 |
| 659 | skip_mm_profiling: bool = MultiModalConfig.skip_mm_profiling |
| 660 | video_pruning_rate: float | None = MultiModalConfig.video_pruning_rate |
| 661 | video_pruning_method: str = MultiModalConfig.video_pruning_method |
| 662 | mm_tensor_ipc: MMTensorIPC = MultiModalConfig.mm_tensor_ipc |
| 663 | mm_processor_device: MMProcessorDevice = "auto" |
| 664 | mm_ipc_gpu_memory_gb: float = MultiModalConfig.mm_ipc_gpu_memory_gb |
| 665 | mm_device_do_normalize: bool | None = MultiModalConfig.mm_device_do_normalize |
| 666 | # LoRA fields |
| 667 | enable_lora: bool = False |
| 668 | max_loras: int = LoRAConfig.max_loras |
| 669 | max_lora_rank: MaxLoRARanks = LoRAConfig.max_lora_rank |
| 670 | default_mm_loras: dict[str, str] | None = LoRAConfig.default_mm_loras |
| 671 | fully_sharded_loras: bool = LoRAConfig.fully_sharded_loras |
| 672 | max_cpu_loras: int | None = LoRAConfig.max_cpu_loras |
| 673 | lora_dtype: str | torch.dtype | None = LoRAConfig.lora_dtype |
| 674 | lora_target_modules: list[str] | None = LoRAConfig.target_modules |
| 675 | enable_tower_connector_lora: bool = LoRAConfig.enable_tower_connector_lora |
| 676 | specialize_active_lora: bool = LoRAConfig.specialize_active_lora |
| 677 | enable_mixed_moe_lora_format: bool = LoRAConfig.enable_mixed_moe_lora_format |
| 678 | enable_moe_shared_loras: bool = LoRAConfig.enable_moe_shared_loras |
| 679 | max_lora_cls_labels: int | None = LoRAConfig.max_lora_cls_labels |
| 680 | |
| 681 | ray_workers_use_nsight: bool = ParallelConfig.ray_workers_use_nsight |
| 682 | num_gpu_blocks_override: int | None = CacheConfig.num_gpu_blocks_override |
| 683 | model_loader_extra_config: dict = get_field(LoadConfig, "model_loader_extra_config") |
| 684 | ignore_patterns: str | list[str] = get_field(LoadConfig, "ignore_patterns") |
| 685 | |
| 686 | enable_chunked_prefill: bool | None = None |
| 687 | disable_chunked_mm_input: bool = SchedulerConfig.disable_chunked_mm_input |
| 688 | |
| 689 | scheduler_reserve_full_isl: bool = SchedulerConfig.scheduler_reserve_full_isl |
| 690 | prefill_schedule_interval: int = SchedulerConfig.prefill_schedule_interval |
| 691 | |
| 692 | watermark: float = SchedulerConfig.watermark |
| 693 | |
| 694 | disable_hybrid_kv_cache_manager: bool | None = ( |
| 695 | SchedulerConfig.disable_hybrid_kv_cache_manager |
| 696 | ) |
| 697 | |
| 698 | structured_outputs_config: StructuredOutputsConfig = get_field( |
| 699 | VllmConfig, "structured_outputs_config" |
| 700 | ) |
| 701 | aux_output_config: AuxOutputConfig = get_field(VllmConfig, "aux_output_config") |
| 702 | reasoning_parser: str = StructuredOutputsConfig.reasoning_parser |
| 703 | reasoning_parser_plugin: str | None = None |
| 704 | |
| 705 | speculative_config: dict[str, Any] | None = None |
| 706 | spec_method: str | None = None |
| 707 | spec_model: str | None = None |
| 708 | spec_tokens: int | None = None |
| 709 | watermark_config: dict[str, Any] | None = None |
| 710 | diffusion_config: dict[str, Any] | None = None |
| 711 | |
| 712 | show_hidden_metrics_for_version: str | None = ( |
| 713 | ObservabilityConfig.show_hidden_metrics_for_version |
| 714 | ) |
| 715 | otlp_traces_endpoint: str | None = ObservabilityConfig.otlp_traces_endpoint |
| 716 | collect_detailed_traces: list[DetailedTraceModules] | None = ( |
| 717 | ObservabilityConfig.collect_detailed_traces |
| 718 | ) |
| 719 | per_request_spec_decode_metrics: Literal["none", "summary", "detailed"] = ( |
| 720 | ObservabilityConfig.per_request_spec_decode_metrics |
| 721 | ) |
| 722 | kv_cache_metrics: bool = ObservabilityConfig.kv_cache_metrics |
| 723 | kv_cache_metrics_sample: float = get_field( |
| 724 | ObservabilityConfig, "kv_cache_metrics_sample" |
| 725 | ) |
| 726 | custom_histogram_buckets: dict[str, list[float]] | None = ( |
| 727 | ObservabilityConfig.custom_histogram_buckets |
| 728 | ) |
| 729 | cudagraph_metrics: bool = ObservabilityConfig.cudagraph_metrics |
| 730 | enable_layerwise_nvtx_tracing: bool = ( |
| 731 | ObservabilityConfig.enable_layerwise_nvtx_tracing |
| 732 | ) |
| 733 | enable_mfu_metrics: bool = ObservabilityConfig.enable_mfu_metrics |
| 734 | enable_logging_iteration_details: bool = ( |
| 735 | ObservabilityConfig.enable_logging_iteration_details |
| 736 | ) |
| 737 | jit_monitor_mode: Literal["warn", "error"] = ObservabilityConfig.jit_monitor_mode |
| 738 | jit_monitor_verbose: bool = ObservabilityConfig.jit_monitor_verbose |
| 739 | enable_mm_processor_stats: bool = ObservabilityConfig.enable_mm_processor_stats |
| 740 | scheduling_policy: SchedulerPolicy = SchedulerConfig.policy |
| 741 | scheduler_cls: str | type[object] | None = SchedulerConfig.scheduler_cls |
| 742 | |
| 743 | pooler_config: PoolerConfig | None = ModelConfig.pooler_config |
| 744 | compilation_config: CompilationConfig = get_field(VllmConfig, "compilation_config") |
| 745 | attention_config: AttentionConfig = get_field(VllmConfig, "attention_config") |
| 746 | engram_config: EngramConfig | None = VllmConfig.engram_config |
| 747 | mamba_config: MambaConfig = get_field(VllmConfig, "mamba_config") |
| 748 | kernel_config: KernelConfig = get_field(VllmConfig, "kernel_config") |
| 749 | enable_flashinfer_autotune: bool = get_field( |
| 750 | KernelConfig, "enable_flashinfer_autotune" |
| 751 | ) |
| 752 | worker_cls: str = ParallelConfig.worker_cls |
| 753 | worker_extension_cls: str = ParallelConfig.worker_extension_cls |
| 754 | |
| 755 | profiler_config: ProfilerConfig = get_field(VllmConfig, "profiler_config") |
| 756 | |
| 757 | logging_config: LoggingConfig | None = None |
| 758 | log_level: LogLevel | None = None |
| 759 | log_config_file: str | None = None |
| 760 | |
| 761 | kv_transfer_config: KVTransferConfig | None = None |
| 762 | kv_events_config: KVEventsConfig | None = None |
| 763 | |
| 764 | ec_transfer_config: ECTransferConfig | None = None |
| 765 | ec_manager_config: EncoderCacheManagerConfig = get_field( |
| 766 | VllmConfig, "ec_manager_config" |
| 767 | ) |
| 768 | reasoning_config: ReasoningConfig = get_field(VllmConfig, "reasoning_config") |
| 769 | |
| 770 | generation_config: str = ModelConfig.generation_config |
| 771 | enable_sleep_mode: bool = ModelConfig.enable_sleep_mode |
| 772 | sleep_preserve_parameter_names: list[str] = get_field( |
| 773 | ModelConfig, "sleep_preserve_parameter_names" |
| 774 | ) |
| 775 | enable_cumem_allocator: bool = ModelConfig.enable_cumem_allocator |
| 776 | enable_nccl_comm_suspend: bool = ModelConfig.enable_nccl_comm_suspend |
| 777 | override_generation_config: dict[str, Any] = get_field( |
| 778 | ModelConfig, "override_generation_config" |
| 779 | ) |
| 780 | model_impl: str = ModelConfig.model_impl |
| 781 | attention_backend: AttentionBackendEnum | None = AttentionConfig.backend |
| 782 | |
| 783 | kv_cache_dtype_skip_layers: list[str] = get_field( |
| 784 | CacheConfig, "kv_cache_dtype_skip_layers" |
| 785 | ) |
| 786 | mamba_cache_dtype: MambaDType = CacheConfig.mamba_cache_dtype |
| 787 | mamba_ssm_cache_dtype: MambaDType = CacheConfig.mamba_ssm_cache_dtype |
| 788 | mamba_block_size: int | None = get_field(CacheConfig, "mamba_block_size") |
| 789 | prefix_match_unit: int | None = get_field(CacheConfig, "prefix_match_unit") |
| 790 | mamba_cache_mode: MambaCacheMode = CacheConfig.mamba_cache_mode |
| 791 | enable_mamba_shared_prefix_checkpoint: bool = ( |
| 792 | CacheConfig.enable_mamba_shared_prefix_checkpoint |
| 793 | ) |
| 794 | replayssm_buffer_len: int = CacheConfig.replayssm_buffer_len |
| 795 | use_replayssm: bool = CacheConfig.use_replayssm |
| 796 | |
| 797 | mamba_backend: MambaBackendEnum = MambaBackendEnum.TRITON |
| 798 | mamba_ssu_algorithm: MambaSSUAlgorithm | None = None |
| 799 | enable_mamba_cache_stochastic_rounding: bool = ( |
| 800 | MambaConfig.enable_stochastic_rounding |
| 801 | ) |
| 802 | mamba_cache_philox_rounds: int = MambaConfig.stochastic_rounding_philox_rounds |
| 803 | |
| 804 | additional_config: dict[str, Any] = get_field(VllmConfig, "additional_config") |
| 805 | |
| 806 | use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load |
| 807 | pt_load_map_location: str | dict[str, str] = LoadConfig.pt_load_map_location |
| 808 | |
| 809 | logits_processors: list[str | type[LogitsProcessor]] | None = ( |
| 810 | ModelConfig.logits_processors |
| 811 | ) |
| 812 | """Custom logitproc types""" |
| 813 | |
| 814 | async_scheduling: bool | None = SchedulerConfig.async_scheduling |
| 815 | |
| 816 | stream_interval: int = SchedulerConfig.stream_interval |
| 817 | |
| 818 | kv_sharing_fast_prefill: bool = CacheConfig.kv_sharing_fast_prefill |
| 819 | swa_bounded_replay: bool = CacheConfig.swa_bounded_replay |
| 820 | optimization_level: OptimizationLevel = VllmConfig.optimization_level |
| 821 | performance_mode: PerformanceMode = VllmConfig.performance_mode |
| 822 | |
| 823 | fault_tolerance_config: FaultToleranceConfig = get_field( |
| 824 | ParallelConfig, "fault_tolerance_config" |
| 825 | ) |
| 826 | enable_fault_tolerance: bool = ParallelConfig.enable_fault_tolerance |
| 827 | |
| 828 | kv_offloading_size: float | None = CacheConfig.kv_offloading_size |
| 829 | kv_offloading_backend: KVOffloadingBackend = CacheConfig.kv_offloading_backend |
| 830 | tokens_only: bool = False |
| 831 | |
| 832 | shutdown_timeout: int = 0 |
| 833 | |
| 834 | weight_transfer_config: WeightTransferConfig | None = get_field( |
| 835 | VllmConfig, |
| 836 | "weight_transfer_config", |
| 837 | ) |
| 838 | |
| 839 | fail_on_environ_validation: bool = False |
| 840 | gdn_prefill_backend: Literal["flashinfer", "triton", "cutedsl"] | None = None |
| 841 | kda_prefill_backend: ( |
| 842 | Literal["auto", "triton", "flashkda", "flashinfer", "fused"] | None |
| 843 | ) = None |
| 844 | kda_decode_backend: Literal["auto", "native", "flashinfer", "triton"] | None = None |
| 845 | |
| 846 | def __post_init__(self): |
| 847 | # support `EngineArgs(compilation_config={...})` |
| 848 | # without having to manually construct a |
| 849 | # CompilationConfig object |
| 850 | if isinstance(self.compilation_config, dict): |
| 851 | self.compilation_config = CompilationConfig(**self.compilation_config) |
| 852 | if isinstance(self.attention_config, dict): |
| 853 | self.attention_config = AttentionConfig(**self.attention_config) |
| 854 | if isinstance(self.aux_output_config, dict): |
| 855 | self.aux_output_config = AuxOutputConfig(**self.aux_output_config) |
| 856 | if self.enable_return_routed_experts: |
| 857 | self.aux_output_config.enable_return_routed_experts = True |
| 858 | if isinstance(self.engram_config, dict): |
| 859 | self.engram_config = EngramConfig(**self.engram_config) |
| 860 | if isinstance(self.mamba_config, dict): |
| 861 | self.mamba_config = MambaConfig(**self.mamba_config) |
| 862 | if isinstance(self.kernel_config, dict): |
| 863 | self.kernel_config = KernelConfig(**self.kernel_config) |
| 864 | if isinstance(self.ec_manager_config, dict): |
| 865 | self.ec_manager_config = EncoderCacheManagerConfig(**self.ec_manager_config) |
| 866 | if isinstance(self.eplb_config, dict): |
| 867 | self.eplb_config = EPLBConfig(**self.eplb_config) |
| 868 | if isinstance(self.weight_transfer_config, dict): |
| 869 | self.weight_transfer_config = WeightTransferConfig( |
| 870 | **self.weight_transfer_config |
| 871 | ) |
| 872 | if isinstance(self.fault_tolerance_config, dict): |
| 873 | if not self.enable_fault_tolerance: |
| 874 | logger.warning( |
| 875 | "--fault-tolerance-config was passed. Fault tolerance is being " |
| 876 | "automatically enabled." |
| 877 | ) |
| 878 | self.enable_fault_tolerance = True |
| 879 | self.fault_tolerance_config = FaultToleranceConfig( |
| 880 | **self.fault_tolerance_config |
| 881 | ) |
| 882 | if isinstance(self.ir_op_priority, dict): |
| 883 | self.ir_op_priority = IrOpPriorityConfig(**self.ir_op_priority) |
| 884 | |
| 885 | from vllm.config.quantization import resolve_quantization_config |
| 886 | |
| 887 | self.quantization_config = resolve_quantization_config( |
| 888 | self.quantization, self.quantization_config |
| 889 | ) |
| 890 | |
| 891 | # Setup plugins |
| 892 | from vllm.plugins import load_general_plugins |
| 893 | |
| 894 | load_general_plugins() |
| 895 | # when use hf offline,replace model and tokenizer id to local model path |
| 896 | if huggingface_hub.constants.HF_HUB_OFFLINE: |
| 897 | # Skip cloud storage URIs (s3://, gs://, az://) — they are not |
| 898 | # HF repo IDs and will be resolved later by |
| 899 | # ModelConfig.maybe_pull_model_tokenizer_for_runai(). |
| 900 | if not is_cloud_storage(self.model): |
| 901 | model_id = self.model |
| 902 | self.model = get_model_path(self.model, self.revision) |
| 903 | if model_id is not self.model: |
| 904 | logger.info( |
| 905 | "HF_HUB_OFFLINE is True, replace model_id " |
| 906 | "[%s] to model_path [%s]", |
| 907 | model_id, |
| 908 | self.model, |
| 909 | ) |
| 910 | if self.tokenizer is not None and not is_cloud_storage(self.tokenizer): |
| 911 | tokenizer_id = self.tokenizer |
| 912 | self.tokenizer = get_model_path(self.tokenizer, self.tokenizer_revision) |
| 913 | if tokenizer_id is not self.tokenizer: |
| 914 | logger.info( |
| 915 | "HF_HUB_OFFLINE is True, replace tokenizer_id [%s] " |
| 916 | "to tokenizer_path [%s]", |
| 917 | tokenizer_id, |
| 918 | self.tokenizer, |
| 919 | ) |
| 920 | |
| 921 | @staticmethod |
| 922 | def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser: |
| 923 | """Shared CLI arguments for vLLM engine.""" |
| 924 | # Model arguments |
| 925 | model_kwargs = get_kwargs(ModelConfig) |
| 926 | aux_output_kwargs = get_kwargs(AuxOutputConfig) |
| 927 | model_group = parser.add_argument_group( |
| 928 | title="ModelConfig", |
| 929 | description=ModelConfig.__doc__, |
| 930 | ) |
| 931 | if not ("serve" in sys.argv[1:] and "--help" in sys.argv[1:]): |
| 932 | model_group.add_argument("--model", **model_kwargs["model"]) |
| 933 | model_group.add_argument("--runner", **model_kwargs["runner"]) |
| 934 | model_group.add_argument("--convert", **model_kwargs["convert"]) |
| 935 | model_group.add_argument("--tokenizer", **model_kwargs["tokenizer"]) |
| 936 | model_group.add_argument("--tokenizer-mode", **model_kwargs["tokenizer_mode"]) |
| 937 | model_group.add_argument( |
| 938 | "--trust-remote-code", **model_kwargs["trust_remote_code"] |
| 939 | ) |
| 940 | model_group.add_argument("--dtype", **model_kwargs["dtype"]) |
| 941 | model_group.add_argument("--seed", **model_kwargs["seed"]) |
| 942 | model_group.add_argument("--hf-config-path", **model_kwargs["hf_config_path"]) |
| 943 | model_group.add_argument( |
| 944 | "--allowed-local-media-path", **model_kwargs["allowed_local_media_path"] |
| 945 | ) |
| 946 | model_group.add_argument( |
| 947 | "--allowed-media-domains", **model_kwargs["allowed_media_domains"] |
| 948 | ) |
| 949 | model_group.add_argument("--revision", **model_kwargs["revision"]) |
| 950 | model_group.add_argument("--code-revision", **model_kwargs["code_revision"]) |
| 951 | model_group.add_argument( |
| 952 | "--tokenizer-revision", **model_kwargs["tokenizer_revision"] |
| 953 | ) |
| 954 | model_group.add_argument("--max-model-len", **model_kwargs["max_model_len"]) |
| 955 | model_group.add_argument("--quantization", "-q", **model_kwargs["quantization"]) |
| 956 | model_group.add_argument( |
| 957 | "--quantization-config", **model_kwargs["quantization_config"] |
| 958 | ) |
| 959 | model_group.add_argument( |
| 960 | "--allow-deprecated-quantization", |
| 961 | **model_kwargs["allow_deprecated_quantization"], |
| 962 | ) |
| 963 | model_group.add_argument("--enforce-eager", **model_kwargs["enforce_eager"]) |
| 964 | model_group.add_argument( |
| 965 | "--enable-return-routed-experts", |
| 966 | **aux_output_kwargs["enable_return_routed_experts"], |
| 967 | ) |
| 968 | model_group.add_argument( |
| 969 | "--return-sampling-mask", |
| 970 | **model_kwargs["return_sampling_mask"], |
| 971 | ) |
| 972 | model_group.add_argument("--max-logprobs", **model_kwargs["max_logprobs"]) |
| 973 | model_group.add_argument("--logprobs-mode", **model_kwargs["logprobs_mode"]) |
| 974 | model_group.add_argument("--use-fp64-gumbel", **model_kwargs["use_fp64_gumbel"]) |
| 975 | model_group.add_argument( |
| 976 | "--enable-trace-replay", **model_kwargs["enable_trace_replay"] |
| 977 | ) |
| 978 | model_group.add_argument( |
| 979 | "--disable-sliding-window", **model_kwargs["disable_sliding_window"] |
| 980 | ) |
| 981 | model_group.add_argument( |
| 982 | "--disable-cascade-attn", **model_kwargs["disable_cascade_attn"] |
| 983 | ) |
| 984 | model_group.add_argument( |
| 985 | "--skip-tokenizer-init", **model_kwargs["skip_tokenizer_init"] |
| 986 | ) |
| 987 | model_group.add_argument( |
| 988 | "--enable-prompt-embeds", **model_kwargs["enable_prompt_embeds"] |
| 989 | ) |
| 990 | model_group.add_argument( |
| 991 | "--served-model-name", **model_kwargs["served_model_name"] |
| 992 | ) |
| 993 | model_group.add_argument("--config-format", **model_kwargs["config_format"]) |
| 994 | model_group.add_argument("--hf-token", **model_kwargs["hf_token"]) |
| 995 | model_group.add_argument("--hf-overrides", **model_kwargs["hf_overrides"]) |
| 996 | model_group.add_argument( |
| 997 | "--model-class-overrides", **model_kwargs["model_class_overrides"] |
| 998 | ) |
| 999 | model_group.add_argument("--pooler-config", **model_kwargs["pooler_config"]) |
| 1000 | model_group.add_argument( |
| 1001 | "--generation-config", **model_kwargs["generation_config"] |
| 1002 | ) |
| 1003 | model_group.add_argument( |
| 1004 | "--override-generation-config", **model_kwargs["override_generation_config"] |
| 1005 | ) |
| 1006 | model_group.add_argument( |
| 1007 | "--enable-sleep-mode", **model_kwargs["enable_sleep_mode"] |
| 1008 | ) |
| 1009 | model_group.add_argument( |
| 1010 | "--sleep-preserve-parameter-names", |
| 1011 | **model_kwargs["sleep_preserve_parameter_names"], |
| 1012 | ) |
| 1013 | model_group.add_argument( |
| 1014 | "--enable-cumem-allocator", **model_kwargs["enable_cumem_allocator"] |
| 1015 | ) |
| 1016 | model_group.add_argument( |
| 1017 | "--enable-nccl-comm-suspend", |
| 1018 | **model_kwargs["enable_nccl_comm_suspend"], |
| 1019 | ) |
| 1020 | model_group.add_argument("--model-impl", **model_kwargs["model_impl"]) |
| 1021 | model_group.add_argument( |
| 1022 | "--logits-processors", **model_kwargs["logits_processors"] |
| 1023 | ) |
| 1024 | model_group.add_argument( |
| 1025 | "--io-processor-plugin", **model_kwargs["io_processor_plugin"] |
| 1026 | ) |
| 1027 | model_group.add_argument( |
| 1028 | "--renderer-num-workers", |
| 1029 | **model_kwargs["renderer_num_workers"], |
| 1030 | ) |
| 1031 | |
| 1032 | # Model loading arguments |
| 1033 | load_kwargs = get_kwargs(LoadConfig) |
| 1034 | load_group = parser.add_argument_group( |
| 1035 | title="LoadConfig", |
| 1036 | description=LoadConfig.__doc__, |
| 1037 | ) |
| 1038 | load_group.add_argument("--load-format", **load_kwargs["load_format"]) |
| 1039 | load_group.add_argument("--download-dir", **load_kwargs["download_dir"]) |
| 1040 | load_group.add_argument( |
| 1041 | "--safetensors-load-strategy", **load_kwargs["safetensors_load_strategy"] |
| 1042 | ) |
| 1043 | load_group.add_argument( |
| 1044 | "--safetensors-prefetch-num-threads", |
| 1045 | **load_kwargs["safetensors_prefetch_num_threads"], |
| 1046 | ) |
| 1047 | load_group.add_argument( |
| 1048 | "--safetensors-prefetch-block-size", |
| 1049 | **load_kwargs["safetensors_prefetch_block_size"], |
| 1050 | ) |
| 1051 | load_group.add_argument( |
| 1052 | "--model-loader-extra-config", **load_kwargs["model_loader_extra_config"] |
| 1053 | ) |
| 1054 | load_group.add_argument("--ignore-patterns", **load_kwargs["ignore_patterns"]) |
| 1055 | load_group.add_argument("--use-tqdm-on-load", **load_kwargs["use_tqdm_on_load"]) |
| 1056 | load_group.add_argument( |
| 1057 | "--pt-load-map-location", **load_kwargs["pt_load_map_location"] |
| 1058 | ) |
| 1059 | |
| 1060 | # Attention arguments |
| 1061 | attention_kwargs = get_kwargs(AttentionConfig) |
| 1062 | attention_group = parser.add_argument_group( |
| 1063 | title="AttentionConfig", |
| 1064 | description=AttentionConfig.__doc__, |
| 1065 | ) |
| 1066 | attention_group.add_argument( |
| 1067 | "--attention-backend", **attention_kwargs["backend"] |
| 1068 | ) |
| 1069 | |
| 1070 | # Mamba arguments |
| 1071 | mamba_kwargs = get_kwargs(MambaConfig) |
| 1072 | mamba_group = parser.add_argument_group( |
| 1073 | title="MambaConfig", |
| 1074 | description=MambaConfig.__doc__, |
| 1075 | ) |
| 1076 | mamba_group.add_argument("--mamba-backend", **mamba_kwargs["backend"]) |
| 1077 | mamba_group.add_argument( |
| 1078 | "--mamba-ssu-algorithm", **mamba_kwargs["ssu_algorithm"] |
| 1079 | ) |
| 1080 | mamba_group.add_argument( |
| 1081 | "--enable-mamba-cache-stochastic-rounding", |
| 1082 | **mamba_kwargs["enable_stochastic_rounding"], |
| 1083 | ) |
| 1084 | mamba_group.add_argument( |
| 1085 | "--mamba-cache-philox-rounds", |
| 1086 | **mamba_kwargs["stochastic_rounding_philox_rounds"], |
| 1087 | ) |
| 1088 | |
| 1089 | # Structured outputs arguments |
| 1090 | structured_outputs_kwargs = get_kwargs(StructuredOutputsConfig) |
| 1091 | structured_outputs_group = parser.add_argument_group( |
| 1092 | title="StructuredOutputsConfig", |
| 1093 | description=StructuredOutputsConfig.__doc__, |
| 1094 | ) |
| 1095 | structured_outputs_group.add_argument( |
| 1096 | "--reasoning-parser", |
| 1097 | # Choices need to be validated after parsing to include plugins |
| 1098 | **structured_outputs_kwargs["reasoning_parser"], |
| 1099 | ) |
| 1100 | structured_outputs_group.add_argument( |
| 1101 | "--reasoning-parser-plugin", |
| 1102 | **structured_outputs_kwargs["reasoning_parser_plugin"], |
| 1103 | ) |
| 1104 | |
| 1105 | # Parallel arguments |
| 1106 | parallel_kwargs = get_kwargs(ParallelConfig) |
| 1107 | parallel_group = parser.add_argument_group( |
| 1108 | title="ParallelConfig", |
| 1109 | description=ParallelConfig.__doc__, |
| 1110 | ) |
| 1111 | parallel_group.add_argument( |
| 1112 | "--distributed-executor-backend", |
| 1113 | **parallel_kwargs["distributed_executor_backend"], |
| 1114 | ) |
| 1115 | parallel_group.add_argument( |
| 1116 | "--pipeline-parallel-size", |
| 1117 | "-pp", |
| 1118 | **parallel_kwargs["pipeline_parallel_size"], |
| 1119 | ) |
| 1120 | parallel_group.add_argument("--master-addr", **parallel_kwargs["master_addr"]) |
| 1121 | parallel_group.add_argument("--master-port", **parallel_kwargs["master_port"]) |
| 1122 | parallel_group.add_argument("--nnodes", "-n", **parallel_kwargs["nnodes"]) |
| 1123 | parallel_group.add_argument("--node-rank", "-r", **parallel_kwargs["node_rank"]) |
| 1124 | parallel_group.add_argument( |
| 1125 | "--distributed-timeout-seconds", |
| 1126 | **parallel_kwargs["distributed_timeout_seconds"], |
| 1127 | ) |
| 1128 | parallel_group.add_argument( |
| 1129 | "--cpu-distributed-timeout-seconds", |
| 1130 | **parallel_kwargs["cpu_distributed_timeout_seconds"], |
| 1131 | ) |
| 1132 | parallel_group.add_argument("--numa-bind", **parallel_kwargs["numa_bind"]) |
| 1133 | parallel_group.add_argument( |
| 1134 | "--numa-bind-nodes", **parallel_kwargs["numa_bind_nodes"] |
| 1135 | ) |
| 1136 | parallel_group.add_argument( |
| 1137 | "--numa-bind-cpus", **parallel_kwargs["numa_bind_cpus"] |
| 1138 | ) |
| 1139 | parallel_group.add_argument( |
| 1140 | "--device-ids", |
| 1141 | type=lambda s: [ |
| 1142 | int(device_id) if device_id.isdigit() else device_id |
| 1143 | for device_id in (part.strip() for part in s.split(",")) |
| 1144 | ], |
| 1145 | default=None, |
| 1146 | help="Comma-separated physical GPU device IDs or UUIDs to use " |
| 1147 | '(e.g. --device-ids "2,3,5,7"). Avoids setting ' |
| 1148 | "CUDA_VISIBLE_DEVICES, preserving full GPU topology " |
| 1149 | "visibility for GPU-NIC affinity and DeepGEMM. " |
| 1150 | "Note: has no effect with Ray executors; use Ray " |
| 1151 | "placement groups for GPU selection instead.", |
| 1152 | ) |
| 1153 | parallel_group.add_argument( |
| 1154 | "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"] |
| 1155 | ) |
| 1156 | parallel_group.add_argument( |
| 1157 | "--decode-context-parallel-size", |
| 1158 | "-dcp", |
| 1159 | **parallel_kwargs["decode_context_parallel_size"], |
| 1160 | ) |
| 1161 | parallel_group.add_argument( |
| 1162 | "--dcp-comm-backend", |
| 1163 | **parallel_kwargs["dcp_comm_backend"], |
| 1164 | ) |
| 1165 | parallel_group.add_argument( |
| 1166 | "--dcp-q-replicate", |
| 1167 | **parallel_kwargs["dcp_q_replicate"], |
| 1168 | ) |
| 1169 | parallel_group.add_argument( |
| 1170 | "--dcp-kv-cache-interleave-size", |
| 1171 | **parallel_kwargs["dcp_kv_cache_interleave_size"], |
| 1172 | ) |
| 1173 | parallel_group.add_argument( |
| 1174 | "--cp-kv-cache-interleave-size", |
| 1175 | **{ |
| 1176 | **parallel_kwargs["cp_kv_cache_interleave_size"], |
| 1177 | "default": None, |
| 1178 | }, |
| 1179 | ) |
| 1180 | parallel_group.add_argument( |
| 1181 | "--prefill-context-parallel-size", |
| 1182 | "-pcp", |
| 1183 | **parallel_kwargs["prefill_context_parallel_size"], |
| 1184 | ) |
| 1185 | parallel_group.add_argument( |
| 1186 | "--data-parallel-size", "-dp", **parallel_kwargs["data_parallel_size"] |
| 1187 | ) |
| 1188 | parallel_group.add_argument( |
| 1189 | "--data-parallel-rank", |
| 1190 | "-dpn", |
| 1191 | type=int, |
| 1192 | help="Data parallel rank of this instance. " |
| 1193 | "When set, enables external load balancer mode for MoE " |
| 1194 | "data-parallel deployments. Unsupported for non-MoE models; " |
| 1195 | "launch independent vLLM instances instead.", |
| 1196 | ) |
| 1197 | parallel_group.add_argument( |
| 1198 | "--data-parallel-start-rank", |
| 1199 | "-dpr", |
| 1200 | type=int, |
| 1201 | help="Starting data parallel rank for secondary nodes.", |
| 1202 | ) |
| 1203 | parallel_group.add_argument( |
| 1204 | "--data-parallel-size-local", |
| 1205 | "-dpl", |
| 1206 | type=int, |
| 1207 | help="Number of data parallel replicas to run on this node.", |
| 1208 | ) |
| 1209 | parallel_group.add_argument( |
| 1210 | "--data-parallel-address", |
| 1211 | "-dpa", |
| 1212 | type=str, |
| 1213 | help="Address of data parallel cluster head-node.", |
| 1214 | ) |
| 1215 | parallel_group.add_argument( |
| 1216 | "--data-parallel-rpc-port", |
| 1217 | "-dpp", |
| 1218 | type=int, |
| 1219 | help="Fixed port for data parallel RPC communication. All nodes " |
| 1220 | "must use the same port.", |
| 1221 | ) |
| 1222 | parallel_group.add_argument( |
| 1223 | "--data-parallel-backend", |
| 1224 | "-dpb", |
| 1225 | type=str, |
| 1226 | default="mp", |
| 1227 | help='Backend for data parallel, either "mp" or "ray".', |
| 1228 | ) |
| 1229 | parallel_group.add_argument( |
| 1230 | "--data-parallel-hybrid-lb", |
| 1231 | "-dph", |
| 1232 | **parallel_kwargs["data_parallel_hybrid_lb"], |
| 1233 | ) |
| 1234 | parallel_group.add_argument( |
| 1235 | "--data-parallel-external-lb", |
| 1236 | "-dpe", |
| 1237 | **parallel_kwargs["data_parallel_external_lb"], |
| 1238 | ) |
| 1239 | parallel_group.add_argument( |
| 1240 | "--data-parallel-multi-port-external-lb", |
| 1241 | "-dpm", |
| 1242 | action="store_true", |
| 1243 | default=False, |
| 1244 | help="Run a node-local supervisor that launches one external-LB API " |
| 1245 | "server per local data parallel rank and exposes aggregated health on " |
| 1246 | "a supervisor port.", |
| 1247 | ) |
| 1248 | parallel_group.add_argument( |
| 1249 | "--enable-expert-parallel", |
| 1250 | "-ep", |
| 1251 | **parallel_kwargs["enable_expert_parallel"], |
| 1252 | ) |
| 1253 | parallel_group.add_argument( |
| 1254 | "--enable-batch-sharded-sampling", |
| 1255 | **parallel_kwargs["enable_batch_sharded_sampling"], |
| 1256 | ) |
| 1257 | parallel_group.add_argument( |
| 1258 | "--enable-ep-weight-filter", |
| 1259 | **parallel_kwargs["enable_ep_weight_filter"], |
| 1260 | ) |
| 1261 | parallel_group.add_argument( |
| 1262 | "--all2all-backend", **parallel_kwargs["all2all_backend"] |
| 1263 | ) |
| 1264 | parallel_group.add_argument("--enable-dbo", **parallel_kwargs["enable_dbo"]) |
| 1265 | parallel_group.add_argument( |
| 1266 | "--ubatch-size", |
| 1267 | **parallel_kwargs["ubatch_size"], |
| 1268 | ) |
| 1269 | parallel_group.add_argument( |
| 1270 | "--enable-elastic-ep", **parallel_kwargs["enable_elastic_ep"] |
| 1271 | ) |
| 1272 | parallel_group.add_argument( |
| 1273 | "--elastic-ep-max-dp-size", |
| 1274 | **parallel_kwargs["elastic_ep_max_dp_size"], |
| 1275 | ) |
| 1276 | parallel_group.add_argument( |
| 1277 | "--dbo-decode-token-threshold", |
| 1278 | **parallel_kwargs["dbo_decode_token_threshold"], |
| 1279 | ) |
| 1280 | parallel_group.add_argument( |
| 1281 | "--dp-sync-interval", |
| 1282 | **parallel_kwargs["dp_sync_interval"], |
| 1283 | ) |
| 1284 | parallel_group.add_argument( |
| 1285 | "--dbo-prefill-token-threshold", |
| 1286 | **parallel_kwargs["dbo_prefill_token_threshold"], |
| 1287 | ) |
| 1288 | parallel_group.add_argument( |
| 1289 | "--disable-nccl-for-dp-synchronization", |
| 1290 | **parallel_kwargs["disable_nccl_for_dp_synchronization"], |
| 1291 | ) |
| 1292 | parallel_group.add_argument("--enable-eplb", **parallel_kwargs["enable_eplb"]) |
| 1293 | parallel_group.add_argument("--eplb-config", **parallel_kwargs["eplb_config"]) |
| 1294 | parallel_group.add_argument( |
| 1295 | "--expert-placement-strategy", |
| 1296 | **parallel_kwargs["expert_placement_strategy"], |
| 1297 | ) |
| 1298 | |
| 1299 | parallel_group.add_argument( |
| 1300 | "--max-parallel-loading-workers", |
| 1301 | **parallel_kwargs["max_parallel_loading_workers"], |
| 1302 | ) |
| 1303 | parallel_group.add_argument( |
| 1304 | "--ray-workers-use-nsight", **parallel_kwargs["ray_workers_use_nsight"] |
| 1305 | ) |
| 1306 | parallel_group.add_argument( |
| 1307 | "--disable-custom-all-reduce", |
| 1308 | **parallel_kwargs["disable_custom_all_reduce"], |
| 1309 | ) |
| 1310 | parallel_group.add_argument("--worker-cls", **parallel_kwargs["worker_cls"]) |
| 1311 | parallel_group.add_argument( |
| 1312 | "--worker-extension-cls", **parallel_kwargs["worker_extension_cls"] |
| 1313 | ) |
| 1314 | parallel_group.add_argument( |
| 1315 | "--enable-fault-tolerance", **parallel_kwargs["enable_fault_tolerance"] |
| 1316 | ) |
| 1317 | parallel_group.add_argument( |
| 1318 | "--fault-tolerance-config", **parallel_kwargs["fault_tolerance_config"] |
| 1319 | ) |
| 1320 | |
| 1321 | # KV cache arguments |
| 1322 | cache_kwargs = get_kwargs(CacheConfig) |
| 1323 | cache_group = parser.add_argument_group( |
| 1324 | title="CacheConfig", |
| 1325 | description=CacheConfig.__doc__, |
| 1326 | ) |
| 1327 | cache_group.add_argument("--block-size", **cache_kwargs["block_size"]) |
| 1328 | cache_group.add_argument( |
| 1329 | "--gpu-memory-utilization", |
| 1330 | "--device-memory-utilization", |
| 1331 | dest="gpu_memory_utilization", |
| 1332 | **cache_kwargs["gpu_memory_utilization"], |
| 1333 | ) |
| 1334 | cache_group.add_argument( |
| 1335 | "--kv-cache-memory-bytes", **cache_kwargs["kv_cache_memory_bytes"] |
| 1336 | ) |
| 1337 | cache_group.add_argument("--kv-cache-dtype", **cache_kwargs["cache_dtype"]) |
| 1338 | cache_group.add_argument( |
| 1339 | "--num-gpu-blocks-override", **cache_kwargs["num_gpu_blocks_override"] |
| 1340 | ) |
| 1341 | cache_group.add_argument( |
| 1342 | "--enable-prefix-caching", |
| 1343 | **{ |
| 1344 | **cache_kwargs["enable_prefix_caching"], |
| 1345 | "default": None, |
| 1346 | }, |
| 1347 | ) |
| 1348 | cache_group.add_argument( |
| 1349 | "--prefix-caching-hash-algo", **cache_kwargs["prefix_caching_hash_algo"] |
| 1350 | ) |
| 1351 | cache_group.add_argument( |
| 1352 | "--prefix-cache-retention-interval", |
| 1353 | **cache_kwargs["prefix_cache_retention_interval"], |
| 1354 | ) |
| 1355 | cache_group.add_argument( |
| 1356 | "--kv-cache-dtype-skip-layers", **cache_kwargs["kv_cache_dtype_skip_layers"] |
| 1357 | ) |
| 1358 | cache_group.add_argument( |
| 1359 | "--kv-sharing-fast-prefill", **cache_kwargs["kv_sharing_fast_prefill"] |
| 1360 | ) |
| 1361 | cache_group.add_argument( |
| 1362 | "--swa-bounded-replay", **cache_kwargs["swa_bounded_replay"] |
| 1363 | ) |
| 1364 | cache_group.add_argument( |
| 1365 | "--mamba-cache-dtype", **cache_kwargs["mamba_cache_dtype"] |
| 1366 | ) |
| 1367 | cache_group.add_argument( |
| 1368 | "--mamba-ssm-cache-dtype", **cache_kwargs["mamba_ssm_cache_dtype"] |
| 1369 | ) |
| 1370 | cache_group.add_argument( |
| 1371 | "--mamba-block-size", **cache_kwargs["mamba_block_size"] |
| 1372 | ) |
| 1373 | cache_group.add_argument( |
| 1374 | "--prefix-match-unit", **cache_kwargs["prefix_match_unit"] |
| 1375 | ) |
| 1376 | cache_group.add_argument( |
| 1377 | "--mamba-cache-mode", **cache_kwargs["mamba_cache_mode"] |
| 1378 | ) |
| 1379 | cache_group.add_argument( |
| 1380 | "--enable-mamba-shared-prefix-checkpoint", |
| 1381 | **cache_kwargs["enable_mamba_shared_prefix_checkpoint"], |
| 1382 | ) |
| 1383 | cache_group.add_argument( |
| 1384 | "--replayssm-buffer-len", **cache_kwargs["replayssm_buffer_len"] |
| 1385 | ) |
| 1386 | cache_group.add_argument("--use-replayssm", **cache_kwargs["use_replayssm"]) |
| 1387 | cache_group.add_argument( |
| 1388 | "--kv-offloading-size", **cache_kwargs["kv_offloading_size"] |
| 1389 | ) |
| 1390 | cache_group.add_argument( |
| 1391 | "--kv-offloading-backend", **cache_kwargs["kv_offloading_backend"] |
| 1392 | ) |
| 1393 | |
| 1394 | # Model weight offload related configs |
| 1395 | offload_kwargs = get_kwargs(OffloadConfig) |
| 1396 | uva_kwargs = get_kwargs(UVAOffloadConfig) |
| 1397 | prefetch_kwargs = get_kwargs(PrefetchOffloadConfig) |
| 1398 | offload_group = parser.add_argument_group( |
| 1399 | title="OffloadConfig", |
| 1400 | description=OffloadConfig.__doc__, |
| 1401 | ) |
| 1402 | offload_group.add_argument( |
| 1403 | "--offload-backend", **offload_kwargs["offload_backend"] |
| 1404 | ) |
| 1405 | offload_group.add_argument("--cpu-offload-gb", **uva_kwargs["cpu_offload_gb"]) |
| 1406 | offload_group.add_argument( |
| 1407 | "--cpu-offload-params", **uva_kwargs["cpu_offload_params"] |
| 1408 | ) |
| 1409 | offload_group.add_argument( |
| 1410 | "--offload-group-size", |
| 1411 | **prefetch_kwargs["offload_group_size"], |
| 1412 | ) |
| 1413 | offload_group.add_argument( |
| 1414 | "--offload-num-in-group", |
| 1415 | **prefetch_kwargs["offload_num_in_group"], |
| 1416 | ) |
| 1417 | offload_group.add_argument( |
| 1418 | "--offload-prefetch-step", |
| 1419 | **prefetch_kwargs["offload_prefetch_step"], |
| 1420 | ) |
| 1421 | offload_group.add_argument( |
| 1422 | "--offload-params", **prefetch_kwargs["offload_params"] |
| 1423 | ) |
| 1424 | |
| 1425 | # Multimodal related configs |
| 1426 | multimodal_kwargs = get_kwargs(MultiModalConfig) |
| 1427 | multimodal_group = parser.add_argument_group( |
| 1428 | title="MultiModalConfig", |
| 1429 | description=MultiModalConfig.__doc__, |
| 1430 | ) |
| 1431 | multimodal_group.add_argument( |
| 1432 | "--language-model-only", **multimodal_kwargs["language_model_only"] |
| 1433 | ) |
| 1434 | multimodal_group.add_argument( |
| 1435 | "--limit-mm-per-prompt", **multimodal_kwargs["limit_per_prompt"] |
| 1436 | ) |
| 1437 | multimodal_group.add_argument( |
| 1438 | "--enable-mm-embeds", **multimodal_kwargs["enable_mm_embeds"] |
| 1439 | ) |
| 1440 | multimodal_group.add_argument( |
| 1441 | "--media-io-kwargs", **multimodal_kwargs["media_io_kwargs"] |
| 1442 | ) |
| 1443 | multimodal_group.add_argument( |
| 1444 | "--mm-processor-kwargs", **multimodal_kwargs["mm_processor_kwargs"] |
| 1445 | ) |
| 1446 | multimodal_group.add_argument( |
| 1447 | "--mm-processor-cache-gb", **multimodal_kwargs["mm_processor_cache_gb"] |
| 1448 | ) |
| 1449 | multimodal_group.add_argument( |
| 1450 | "--mm-processor-cache-type", **multimodal_kwargs["mm_processor_cache_type"] |
| 1451 | ) |
| 1452 | multimodal_group.add_argument( |
| 1453 | "--mm-hasher-algorithm", **multimodal_kwargs["mm_hasher_algorithm"] |
| 1454 | ) |
| 1455 | multimodal_group.add_argument( |
| 1456 | "--mm-shm-cache-max-object-size-mb", |
| 1457 | **multimodal_kwargs["mm_shm_cache_max_object_size_mb"], |
| 1458 | ) |
| 1459 | multimodal_group.add_argument( |
| 1460 | "--mm-encoder-only", **multimodal_kwargs["mm_encoder_only"] |
| 1461 | ) |
| 1462 | multimodal_group.add_argument( |
| 1463 | "--mm-encoder-tp-mode", **multimodal_kwargs["mm_encoder_tp_mode"] |
| 1464 | ) |
| 1465 | multimodal_group.add_argument( |
| 1466 | "--mm-encoder-attn-backend", |
| 1467 | **multimodal_kwargs["mm_encoder_attn_backend"], |
| 1468 | ) |
| 1469 | multimodal_group.add_argument( |
| 1470 | "--mm-encoder-attn-dtype", |
| 1471 | **multimodal_kwargs["mm_encoder_attn_dtype"], |
| 1472 | ) |
| 1473 | multimodal_group.add_argument( |
| 1474 | "--mm-encoder-fp8-scale-path", |
| 1475 | **multimodal_kwargs["mm_encoder_fp8_scale_path"], |
| 1476 | ) |
| 1477 | multimodal_group.add_argument( |
| 1478 | "--mm-encoder-fp8-scale-save-path", |
| 1479 | **multimodal_kwargs["mm_encoder_fp8_scale_save_path"], |
| 1480 | ) |
| 1481 | multimodal_group.add_argument( |
| 1482 | "--mm-encoder-fp8-scale-save-margin", |
| 1483 | **multimodal_kwargs["mm_encoder_fp8_scale_save_margin"], |
| 1484 | ) |
| 1485 | multimodal_group.add_argument( |
| 1486 | "--interleave-mm-strings", **multimodal_kwargs["interleave_mm_strings"] |
| 1487 | ) |
| 1488 | multimodal_group.add_argument( |
| 1489 | "--skip-mm-profiling", **multimodal_kwargs["skip_mm_profiling"] |
| 1490 | ) |
| 1491 | |
| 1492 | multimodal_group.add_argument( |
| 1493 | "--video-pruning-rate", **multimodal_kwargs["video_pruning_rate"] |
| 1494 | ) |
| 1495 | multimodal_group.add_argument( |
| 1496 | "--video-pruning-method", |
| 1497 | **multimodal_kwargs["video_pruning_method"], |
| 1498 | ) |
| 1499 | multimodal_group.add_argument( |
| 1500 | "--mm-tensor-ipc", **multimodal_kwargs["mm_tensor_ipc"] |
| 1501 | ) |
| 1502 | multimodal_group.add_argument( |
| 1503 | "--mm-processor-device", |
| 1504 | choices=["auto", "cpu"] |
| 1505 | + ( |
| 1506 | [current_platform.device_type] |
| 1507 | if current_platform.device_type not in ("", "cpu") |
| 1508 | else [] |
| 1509 | ), |
| 1510 | default="auto", |
| 1511 | help="Device the HF multi-modal processor runs the image/video " |
| 1512 | "transform on. Convenience for `--mm-processor-kwargs " |
| 1513 | "'{\"device\": ...}'`: the value is resolved here and stored there, " |
| 1514 | "it is not kept as separate state. Only takes effect for HF " |
| 1515 | '"fast" (torchvision-backed) processors, which accept a `device` ' |
| 1516 | "argument; the others ignore it and stay on CPU.\n\n" |
| 1517 | '"auto" uses the accelerator on encoder instances of an ' |
| 1518 | "encode/prefill/decode deployment -- an EC producer that is not " |
| 1519 | "also a consumer allocates no KV cache, so its accelerator is not " |
| 1520 | "contended by the language model -- and then only when " |
| 1521 | "`--mm-tensor-ipc=torch_shm` can carry device tensors, since every " |
| 1522 | "other transport would copy the result back to the host and that " |
| 1523 | 'copy costs more than it saves. "auto" resolves to "cpu" ' |
| 1524 | "everywhere else.", |
| 1525 | ) |
| 1526 | multimodal_group.add_argument( |
| 1527 | "--mm-ipc-gpu-memory-gb", |
| 1528 | **multimodal_kwargs["mm_ipc_gpu_memory_gb"], |
| 1529 | ) |
| 1530 | multimodal_group.add_argument( |
| 1531 | "--mm-device-do-normalize", |
| 1532 | **{ |
| 1533 | **multimodal_kwargs["mm_device_do_normalize"], |
| 1534 | "default": None, |
| 1535 | }, |
| 1536 | ) |
| 1537 | |
| 1538 | # LoRA related configs |
| 1539 | lora_kwargs = get_kwargs(LoRAConfig) |
| 1540 | lora_group = parser.add_argument_group( |
| 1541 | title="LoRAConfig", |
| 1542 | description=LoRAConfig.__doc__, |
| 1543 | ) |
| 1544 | lora_group.add_argument( |
| 1545 | "--enable-lora", |
| 1546 | action=argparse.BooleanOptionalAction, |
| 1547 | help="If True, enable handling of LoRA adapters.", |
| 1548 | ) |
| 1549 | lora_group.add_argument("--max-loras", **lora_kwargs["max_loras"]) |
| 1550 | lora_group.add_argument("--max-lora-rank", **lora_kwargs["max_lora_rank"]) |
| 1551 | lora_group.add_argument( |
| 1552 | "--lora-dtype", |
| 1553 | **lora_kwargs["lora_dtype"], |
| 1554 | ) |
| 1555 | lora_group.add_argument( |
| 1556 | "--enable-tower-connector-lora", |
| 1557 | **lora_kwargs["enable_tower_connector_lora"], |
| 1558 | ) |
| 1559 | lora_group.add_argument("--max-cpu-loras", **lora_kwargs["max_cpu_loras"]) |
| 1560 | lora_group.add_argument( |
| 1561 | "--fully-sharded-loras", **lora_kwargs["fully_sharded_loras"] |
| 1562 | ) |
| 1563 | lora_group.add_argument( |
| 1564 | "--lora-target-modules", **lora_kwargs["target_modules"] |
| 1565 | ) |
| 1566 | lora_group.add_argument("--default-mm-loras", **lora_kwargs["default_mm_loras"]) |
| 1567 | lora_group.add_argument( |
| 1568 | "--specialize-active-lora", **lora_kwargs["specialize_active_lora"] |
| 1569 | ) |
| 1570 | lora_group.add_argument( |
| 1571 | "--enable-mixed-moe-lora-format", |
| 1572 | **lora_kwargs["enable_mixed_moe_lora_format"], |
| 1573 | ) |
| 1574 | lora_group.add_argument( |
| 1575 | "--enable-moe-shared-loras", |
| 1576 | **lora_kwargs["enable_moe_shared_loras"], |
| 1577 | ) |
| 1578 | lora_group.add_argument( |
| 1579 | "--max-lora-cls-labels", |
| 1580 | **lora_kwargs["max_lora_cls_labels"], |
| 1581 | ) |
| 1582 | |
| 1583 | # Logging arguments |
| 1584 | logging_group = parser.add_argument_group( |
| 1585 | title="LoggingConfig", |
| 1586 | description=LoggingConfig.__doc__, |
| 1587 | ) |
| 1588 | logging_config_kwargs = get_kwargs(VllmConfig)["logging_config"] |
| 1589 | logging_config_kwargs["default"] = argparse.SUPPRESS |
| 1590 | logging_group.add_argument("--logging-config", **logging_config_kwargs) |
| 1591 | logging_group.add_argument( |
| 1592 | "--log-level", |
| 1593 | choices=get_args(LogLevel), |
| 1594 | default=argparse.SUPPRESS, |
| 1595 | help=( |
| 1596 | "Shortcut for --logging-config.log_level. " |
| 1597 | "Overrides that field if both are specified." |
| 1598 | ), |
| 1599 | ) |
| 1600 | logging_group.add_argument( |
| 1601 | "--log-config-file", |
| 1602 | dest="log_config_file", |
| 1603 | default=argparse.SUPPRESS, |
| 1604 | metavar="PYLOGGING_CONFIG_FILE", |
| 1605 | action=DeprecatedLogConfigFileAction, |
| 1606 | help=_LOG_CONFIG_FILE_DEPRECATION_MESSAGE, |
| 1607 | ) |
| 1608 | # Retain the warning in v0.31.0 and v0.32.0. Remove this option and its |
| 1609 | # compatibility mapping in create_logging_config() in v0.33.0. |
| 1610 | |
| 1611 | # Observability arguments |
| 1612 | observability_kwargs = get_kwargs(ObservabilityConfig) |
| 1613 | observability_group = parser.add_argument_group( |
| 1614 | title="ObservabilityConfig", |
| 1615 | description=ObservabilityConfig.__doc__, |
| 1616 | ) |
| 1617 | observability_group.add_argument( |
| 1618 | "--show-hidden-metrics-for-version", |
| 1619 | **observability_kwargs["show_hidden_metrics_for_version"], |
| 1620 | ) |
| 1621 | observability_group.add_argument( |
| 1622 | "--otlp-traces-endpoint", **observability_kwargs["otlp_traces_endpoint"] |
| 1623 | ) |
| 1624 | observability_group.add_argument( |
| 1625 | "--collect-detailed-traces", |
| 1626 | **observability_kwargs["collect_detailed_traces"], |
| 1627 | ) |
| 1628 | observability_group.add_argument( |
| 1629 | "--per-request-spec-decode-metrics", |
| 1630 | **observability_kwargs["per_request_spec_decode_metrics"], |
| 1631 | ) |
| 1632 | observability_group.add_argument( |
| 1633 | "--kv-cache-metrics", **observability_kwargs["kv_cache_metrics"] |
| 1634 | ) |
| 1635 | observability_group.add_argument( |
| 1636 | "--kv-cache-metrics-sample", |
| 1637 | **observability_kwargs["kv_cache_metrics_sample"], |
| 1638 | ) |
| 1639 | observability_group.add_argument( |
| 1640 | "--custom-histogram-buckets", |
| 1641 | **observability_kwargs["custom_histogram_buckets"], |
| 1642 | ) |
| 1643 | observability_group.add_argument( |
| 1644 | "--cudagraph-metrics", |
| 1645 | **observability_kwargs["cudagraph_metrics"], |
| 1646 | ) |
| 1647 | observability_group.add_argument( |
| 1648 | "--enable-layerwise-nvtx-tracing", |
| 1649 | **observability_kwargs["enable_layerwise_nvtx_tracing"], |
| 1650 | ) |
| 1651 | observability_group.add_argument( |
| 1652 | "--enable-mfu-metrics", |
| 1653 | **observability_kwargs["enable_mfu_metrics"], |
| 1654 | ) |
| 1655 | observability_group.add_argument( |
| 1656 | "--enable-logging-iteration-details", |
| 1657 | **observability_kwargs["enable_logging_iteration_details"], |
| 1658 | ) |
| 1659 | observability_group.add_argument( |
| 1660 | "--jit-monitor-mode", |
| 1661 | **observability_kwargs["jit_monitor_mode"], |
| 1662 | ) |
| 1663 | observability_group.add_argument( |
| 1664 | "--jit-monitor-verbose", |
| 1665 | **observability_kwargs["jit_monitor_verbose"], |
| 1666 | ) |
| 1667 | |
| 1668 | # Scheduler arguments |
| 1669 | scheduler_kwargs = get_kwargs(SchedulerConfig) |
| 1670 | scheduler_group = parser.add_argument_group( |
| 1671 | title="SchedulerConfig", |
| 1672 | description=SchedulerConfig.__doc__, |
| 1673 | ) |
| 1674 | scheduler_group.add_argument( |
| 1675 | "--max-num-batched-tokens", |
| 1676 | **{ |
| 1677 | **scheduler_kwargs["max_num_batched_tokens"], |
| 1678 | "default": None, |
| 1679 | }, |
| 1680 | ) |
| 1681 | scheduler_group.add_argument( |
| 1682 | "--max-num-scheduled-tokens", |
| 1683 | **{ |
| 1684 | **scheduler_kwargs["max_num_scheduled_tokens"], |
| 1685 | "default": None, |
| 1686 | }, |
| 1687 | ) |
| 1688 | scheduler_group.add_argument( |
| 1689 | "--max-num-seqs", |
| 1690 | **{ |
| 1691 | **scheduler_kwargs["max_num_seqs"], |
| 1692 | "default": None, |
| 1693 | }, |
| 1694 | ) |
| 1695 | scheduler_group.add_argument( |
| 1696 | "--max-num-active-seqs", |
| 1697 | **scheduler_kwargs["max_num_active_seqs"], |
| 1698 | ) |
| 1699 | scheduler_group.add_argument( |
| 1700 | "--max-num-queued-reqs", **scheduler_kwargs["max_num_queued_reqs"] |
| 1701 | ) |
| 1702 | scheduler_group.add_argument( |
| 1703 | "--max-num-queued-tokens", |
| 1704 | **scheduler_kwargs["max_num_queued_tokens"], |
| 1705 | ) |
| 1706 | scheduler_group.add_argument( |
| 1707 | "--long-prefill-token-threshold", |
| 1708 | **scheduler_kwargs["long_prefill_token_threshold"], |
| 1709 | ) |
| 1710 | scheduler_group.add_argument( |
| 1711 | "--long-prefill-token-threshold-adaptive", |
| 1712 | **scheduler_kwargs["long_prefill_token_threshold_adaptive"], |
| 1713 | ) |
| 1714 | # multi-step scheduling has been removed; corresponding arguments |
| 1715 | # are no longer supported. |
| 1716 | scheduler_group.add_argument( |
| 1717 | "--scheduling-policy", **scheduler_kwargs["policy"] |
| 1718 | ) |
| 1719 | scheduler_group.add_argument( |
| 1720 | "--enable-chunked-prefill", |
| 1721 | **{ |
| 1722 | **scheduler_kwargs["enable_chunked_prefill"], |
| 1723 | "default": None, |
| 1724 | }, |
| 1725 | ) |
| 1726 | scheduler_group.add_argument( |
| 1727 | "--disable-chunked-mm-input", **scheduler_kwargs["disable_chunked_mm_input"] |
| 1728 | ) |
| 1729 | scheduler_group.add_argument( |
| 1730 | "--scheduler-cls", **scheduler_kwargs["scheduler_cls"] |
| 1731 | ) |
| 1732 | scheduler_group.add_argument( |
| 1733 | "--scheduler-reserve-full-isl", |
| 1734 | **scheduler_kwargs["scheduler_reserve_full_isl"], |
| 1735 | ) |
| 1736 | scheduler_group.add_argument("--watermark", **scheduler_kwargs["watermark"]) |
| 1737 | scheduler_group.add_argument( |
| 1738 | "--prefill-schedule-interval", |
| 1739 | **scheduler_kwargs["prefill_schedule_interval"], |
| 1740 | ) |
| 1741 | scheduler_group.add_argument( |
| 1742 | "--disable-hybrid-kv-cache-manager", |
| 1743 | **scheduler_kwargs["disable_hybrid_kv_cache_manager"], |
| 1744 | ) |
| 1745 | scheduler_group.add_argument( |
| 1746 | "--async-scheduling", **scheduler_kwargs["async_scheduling"] |
| 1747 | ) |
| 1748 | scheduler_group.add_argument( |
| 1749 | "--stream-interval", **scheduler_kwargs["stream_interval"] |
| 1750 | ) |
| 1751 | |
| 1752 | # Compilation arguments |
| 1753 | compilation_kwargs = get_kwargs(CompilationConfig) |
| 1754 | compilation_group = parser.add_argument_group( |
| 1755 | title="CompilationConfig", |
| 1756 | description=CompilationConfig.__doc__, |
| 1757 | ) |
| 1758 | compilation_group.add_argument( |
| 1759 | "--cudagraph-capture-sizes", **compilation_kwargs["cudagraph_capture_sizes"] |
| 1760 | ) |
| 1761 | compilation_group.add_argument( |
| 1762 | "--max-cudagraph-capture-size", |
| 1763 | **compilation_kwargs["max_cudagraph_capture_size"], |
| 1764 | ) |
| 1765 | |
| 1766 | # Kernel arguments |
| 1767 | kernel_kwargs = get_kwargs(KernelConfig) |
| 1768 | kernel_group = parser.add_argument_group( |
| 1769 | title="KernelConfig", |
| 1770 | description=KernelConfig.__doc__, |
| 1771 | ) |
| 1772 | kernel_group.add_argument("--ir-op-priority", **kernel_kwargs["ir_op_priority"]) |
| 1773 | kernel_group.add_argument( |
| 1774 | "--enable-flashinfer-autotune", |
| 1775 | **kernel_kwargs["enable_flashinfer_autotune"], |
| 1776 | ) |
| 1777 | moe_backend_kwargs = kernel_kwargs["moe_backend"] |
| 1778 | moe_backend_kwargs["type"] = lambda s: s.lower().replace("-", "_") |
| 1779 | kernel_group.add_argument("--moe-backend", **moe_backend_kwargs) |
| 1780 | linear_backend_kwargs = kernel_kwargs["linear_backend"] |
| 1781 | linear_backend_kwargs["type"] = lambda s: s.lower().replace("-", "_") |
| 1782 | kernel_group.add_argument("--linear-backend", **linear_backend_kwargs) |
| 1783 | sparse_indexer_topk_kwargs = kernel_kwargs["sparse_indexer_topk_backend"] |
| 1784 | sparse_indexer_topk_kwargs["type"] = lambda s: s.lower().replace("-", "_") |
| 1785 | kernel_group.add_argument( |
| 1786 | "--sparse-indexer-topk-backend", **sparse_indexer_topk_kwargs |
| 1787 | ) |
| 1788 | |
| 1789 | # vLLM arguments |
| 1790 | vllm_kwargs = get_kwargs(VllmConfig) |
| 1791 | vllm_group = parser.add_argument_group( |
| 1792 | title="VllmConfig", |
| 1793 | description=VllmConfig.__doc__, |
| 1794 | ) |
| 1795 | # We construct SpeculativeConfig using fields from other configs in |
| 1796 | # create_engine_config. So we set the type to a JSON string here to |
| 1797 | # delay the Pydantic validation that comes with SpeculativeConfig. |
| 1798 | vllm_kwargs["speculative_config"]["type"] = optional_type(json.loads) |
| 1799 | vllm_group.add_argument( |
| 1800 | "--speculative-config", "-sc", **vllm_kwargs["speculative_config"] |
| 1801 | ) |
| 1802 | speculative_kwargs = get_kwargs(SpeculativeConfig) |
| 1803 | vllm_group.add_argument("--spec-method", **speculative_kwargs["method"]) |
| 1804 | vllm_group.add_argument("--spec-model", **speculative_kwargs["model"]) |
| 1805 | vllm_group.add_argument( |
| 1806 | "--spec-tokens", **speculative_kwargs["num_speculative_tokens"] |
| 1807 | ) |
| 1808 | vllm_kwargs["watermark_config"]["type"] = optional_type(json.loads) |
| 1809 | vllm_group.add_argument("--watermark-config", **vllm_kwargs["watermark_config"]) |
| 1810 | vllm_kwargs["diffusion_config"]["type"] = optional_type(json.loads) |
| 1811 | vllm_group.add_argument( |
| 1812 | "--diffusion-config", "-dc", **vllm_kwargs["diffusion_config"] |
| 1813 | ) |
| 1814 | vllm_group.add_argument( |
| 1815 | "--kv-transfer-config", **vllm_kwargs["kv_transfer_config"] |
| 1816 | ) |
| 1817 | vllm_group.add_argument("--kv-events-config", **vllm_kwargs["kv_events_config"]) |
| 1818 | vllm_group.add_argument( |
| 1819 | "--ec-transfer-config", **vllm_kwargs["ec_transfer_config"] |
| 1820 | ) |
| 1821 | vllm_group.add_argument( |
| 1822 | "--ec-manager-config", **vllm_kwargs["ec_manager_config"] |
| 1823 | ) |
| 1824 | vllm_group.add_argument( |
| 1825 | "--compilation-config", "-cc", **vllm_kwargs["compilation_config"] |
| 1826 | ) |
| 1827 | vllm_group.add_argument( |
| 1828 | "--attention-config", "-ac", **vllm_kwargs["attention_config"] |
| 1829 | ) |
| 1830 | vllm_group.add_argument("--engram-config", **vllm_kwargs["engram_config"]) |
| 1831 | vllm_group.add_argument("--reasoning-config", **vllm_kwargs["reasoning_config"]) |
| 1832 | vllm_group.add_argument("--kernel-config", **vllm_kwargs["kernel_config"]) |
| 1833 | vllm_group.add_argument( |
| 1834 | "--additional-config", **vllm_kwargs["additional_config"] |
| 1835 | ) |
| 1836 | vllm_group.add_argument( |
| 1837 | "--structured-outputs-config", **vllm_kwargs["structured_outputs_config"] |
| 1838 | ) |
| 1839 | vllm_group.add_argument( |
| 1840 | "--aux-output-config", **vllm_kwargs["aux_output_config"] |
| 1841 | ) |
| 1842 | vllm_group.add_argument("--profiler-config", **vllm_kwargs["profiler_config"]) |
| 1843 | vllm_group.add_argument( |
| 1844 | "--optimization-level", **vllm_kwargs["optimization_level"] |
| 1845 | ) |
| 1846 | vllm_group.add_argument("--performance-mode", **vllm_kwargs["performance_mode"]) |
| 1847 | vllm_group.add_argument( |
| 1848 | "--weight-transfer-config", **vllm_kwargs["weight_transfer_config"] |
| 1849 | ) |
| 1850 | |
| 1851 | # Other arguments |
| 1852 | parser.add_argument( |
| 1853 | "--disable-log-stats", |
| 1854 | action="store_true", |
| 1855 | help="Disable logging statistics.", |
| 1856 | ) |
| 1857 | |
| 1858 | parser.add_argument( |
| 1859 | "--aggregate-engine-logging", |
| 1860 | action="store_true", |
| 1861 | help="Log aggregate rather than per-engine statistics " |
| 1862 | "when using data parallelism.", |
| 1863 | ) |
| 1864 | |
| 1865 | parser.add_argument( |
| 1866 | "--fail-on-environ-validation", |
| 1867 | help="If set, the engine will raise an error if " |
| 1868 | "environment validation fails.", |
| 1869 | default=False, |
| 1870 | action=argparse.BooleanOptionalAction, |
| 1871 | ) |
| 1872 | |
| 1873 | parser.add_argument( |
| 1874 | "--shutdown-timeout", |
| 1875 | type=int, |
| 1876 | default=0, |
| 1877 | help="Shutdown timeout in seconds. 0 = abort, >0 = wait.", |
| 1878 | ) |
| 1879 | |
| 1880 | parser.add_argument( |
| 1881 | "--gdn-prefill-backend", |
| 1882 | dest="gdn_prefill_backend", |
| 1883 | choices=["flashinfer", "triton", "cutedsl"], |
| 1884 | default=None, |
| 1885 | help="Select GDN prefill backend.", |
| 1886 | ) |
| 1887 | parser.add_argument( |
| 1888 | "--kda-prefill-backend", |
| 1889 | dest="kda_prefill_backend", |
| 1890 | choices=["auto", "triton", "flashkda", "flashinfer", "fused"], |
| 1891 | default=None, |
| 1892 | help="Select KDA prefill backend. 'flashkda' is CUDA-only and " |
| 1893 | "'fused' is ROCm-only; 'auto' picks a supported backend.", |
| 1894 | ) |
| 1895 | parser.add_argument( |
| 1896 | "--kda-decode-backend", |
| 1897 | dest="kda_decode_backend", |
| 1898 | choices=["auto", "native", "flashinfer", "triton"], |
| 1899 | default=None, |
| 1900 | help="Select KDA decode backend.", |
| 1901 | ) |
| 1902 | return parser |
| 1903 | |
| 1904 | @classmethod |
| 1905 | def from_cli_args(cls, args: argparse.Namespace): |
| 1906 | # Get the list of attributes of this dataclass. |
| 1907 | attrs = [attr.name for attr in dataclasses.fields(cls)] |
| 1908 | |
| 1909 | # Set the attributes from the parsed arguments. |
| 1910 | engine_args = cls( |
| 1911 | **{attr: getattr(args, attr) for attr in attrs if hasattr(args, attr)} |
| 1912 | ) |
| 1913 | return engine_args |
| 1914 | |
| 1915 | def create_model_config(self) -> ModelConfig: |
| 1916 | if not envs.VLLM_ENABLE_V1_MULTIPROCESSING: |
| 1917 | logger.warning( |
| 1918 | "The global random seed is set to %d. Since " |
| 1919 | "VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may " |
| 1920 | "affect the random state of the Python process that " |
| 1921 | "launched vLLM.", |
| 1922 | self.seed, |
| 1923 | ) |
| 1924 | |
| 1925 | return ModelConfig( |
| 1926 | model=self.model, |
| 1927 | model_weights=self.model_weights, |
| 1928 | hf_config_path=self.hf_config_path, |
| 1929 | runner=self.runner, |
| 1930 | convert=self.convert, |
| 1931 | tokenizer=self.tokenizer, # type: ignore[arg-type] |
| 1932 | tokenizer_mode=self.tokenizer_mode, |
| 1933 | trust_remote_code=self.trust_remote_code, |
| 1934 | allowed_local_media_path=self.allowed_local_media_path, |
| 1935 | allowed_media_domains=self.allowed_media_domains, |
| 1936 | dtype=self.dtype, |
| 1937 | seed=self.seed, |
| 1938 | revision=self.revision, |
| 1939 | code_revision=self.code_revision, |
| 1940 | hf_token=self.hf_token, |
| 1941 | hf_overrides=self.hf_overrides, |
| 1942 | model_class_overrides=self.model_class_overrides, |
| 1943 | tokenizer_revision=self.tokenizer_revision, |
| 1944 | max_model_len=self.max_model_len, |
| 1945 | quantization=self.quantization, |
| 1946 | quantization_config=self.quantization_config, |
| 1947 | allow_deprecated_quantization=self.allow_deprecated_quantization, |
| 1948 | enforce_eager=self.enforce_eager, |
| 1949 | return_sampling_mask=self.return_sampling_mask, |
| 1950 | max_logprobs=self.max_logprobs, |
| 1951 | logprobs_mode=self.logprobs_mode, |
| 1952 | use_fp64_gumbel=self.use_fp64_gumbel, |
| 1953 | enable_trace_replay=self.enable_trace_replay, |
| 1954 | disable_sliding_window=self.disable_sliding_window, |
| 1955 | disable_cascade_attn=self.disable_cascade_attn, |
| 1956 | skip_tokenizer_init=self.skip_tokenizer_init, |
| 1957 | enable_prompt_embeds=self.enable_prompt_embeds, |
| 1958 | served_model_name=self.served_model_name, |
| 1959 | language_model_only=self.language_model_only, |
| 1960 | limit_mm_per_prompt=self.limit_mm_per_prompt, |
| 1961 | enable_mm_embeds=self.enable_mm_embeds, |
| 1962 | interleave_mm_strings=self.interleave_mm_strings, |
| 1963 | media_io_kwargs=self.media_io_kwargs, |
| 1964 | skip_mm_profiling=self.skip_mm_profiling, |
| 1965 | config_format=self.config_format, |
| 1966 | mm_processor_kwargs=self.mm_processor_kwargs, |
| 1967 | mm_processor_cache_gb=self.mm_processor_cache_gb, |
| 1968 | mm_processor_cache_type=self.mm_processor_cache_type, |
| 1969 | mm_shm_cache_max_object_size_mb=self.mm_shm_cache_max_object_size_mb, |
| 1970 | mm_hasher_algorithm=self.mm_hasher_algorithm, |
| 1971 | mm_encoder_only=self.mm_encoder_only, |
| 1972 | mm_encoder_tp_mode=self.mm_encoder_tp_mode, |
| 1973 | mm_encoder_attn_backend=self.mm_encoder_attn_backend, |
| 1974 | mm_encoder_attn_dtype=self.mm_encoder_attn_dtype, |
| 1975 | mm_encoder_fp8_scale_path=self.mm_encoder_fp8_scale_path, |
| 1976 | mm_encoder_fp8_scale_save_path=self.mm_encoder_fp8_scale_save_path, |
| 1977 | mm_encoder_fp8_scale_save_margin=self.mm_encoder_fp8_scale_save_margin, |
| 1978 | pooler_config=self.pooler_config, |
| 1979 | generation_config=self.generation_config, |
| 1980 | override_generation_config=self.override_generation_config, |
| 1981 | enable_sleep_mode=self.enable_sleep_mode, |
| 1982 | sleep_preserve_parameter_names=self.sleep_preserve_parameter_names, |
| 1983 | enable_cumem_allocator=self.enable_cumem_allocator, |
| 1984 | enable_nccl_comm_suspend=self.enable_nccl_comm_suspend, |
| 1985 | model_impl=self.model_impl, |
| 1986 | logits_processors=self.logits_processors, |
| 1987 | video_pruning_rate=self.video_pruning_rate, |
| 1988 | video_pruning_method=self.video_pruning_method, |
| 1989 | mm_tensor_ipc=self.mm_tensor_ipc, |
| 1990 | mm_ipc_gpu_memory_gb=self.mm_ipc_gpu_memory_gb, |
| 1991 | mm_device_do_normalize=self.mm_device_do_normalize, |
| 1992 | mm_processor_device=self.mm_processor_device, |
| 1993 | io_processor_plugin=self.io_processor_plugin, |
| 1994 | renderer_num_workers=self.renderer_num_workers, |
| 1995 | ) |
| 1996 | |
| 1997 | def validate_tensorizer_args(self): |
| 1998 | from vllm.model_executor.model_loader.tensorizer import TensorizerConfig |
| 1999 | |
| 2000 | for key in self.model_loader_extra_config: |
| 2001 | if key in TensorizerConfig._fields: |
| 2002 | self.model_loader_extra_config["tensorizer_config"][key] = ( |
| 2003 | self.model_loader_extra_config[key] |
| 2004 | ) |
| 2005 | |
| 2006 | def create_load_config(self) -> LoadConfig: |
| 2007 | if self.load_format == "tensorizer": |
| 2008 | if hasattr(self.model_loader_extra_config, "to_serializable"): |
| 2009 | self.model_loader_extra_config = ( |
| 2010 | self.model_loader_extra_config.to_serializable() |
| 2011 | ) |
| 2012 | self.model_loader_extra_config["tensorizer_config"] = {} |
| 2013 | self.model_loader_extra_config["tensorizer_config"]["tensorizer_dir"] = ( |
| 2014 | self.model |
| 2015 | ) |
| 2016 | self.validate_tensorizer_args() |
| 2017 | |
| 2018 | return LoadConfig( |
| 2019 | load_format=self.load_format, |
| 2020 | download_dir=self.download_dir, |
| 2021 | safetensors_load_strategy=self.safetensors_load_strategy, |
| 2022 | safetensors_prefetch_num_threads=self.safetensors_prefetch_num_threads, |
| 2023 | safetensors_prefetch_block_size=self.safetensors_prefetch_block_size, |
| 2024 | model_loader_extra_config=self.model_loader_extra_config, |
| 2025 | ignore_patterns=self.ignore_patterns, |
| 2026 | use_tqdm_on_load=self.use_tqdm_on_load, |
| 2027 | pt_load_map_location=self.pt_load_map_location, |
| 2028 | ) |
| 2029 | |
| 2030 | def create_speculative_config( |
| 2031 | self, |
| 2032 | target_model_config: ModelConfig, |
| 2033 | target_parallel_config: ParallelConfig, |
| 2034 | ) -> SpeculativeConfig | None: |
| 2035 | """Initializes and returns a SpeculativeConfig object based on |
| 2036 | `speculative_config`. |
| 2037 | """ |
| 2038 | for flag, key, value in ( |
| 2039 | ("--spec-method", "method", self.spec_method), |
| 2040 | ("--spec-model", "model", self.spec_model), |
| 2041 | ("--spec-tokens", "num_speculative_tokens", self.spec_tokens), |
| 2042 | ): |
| 2043 | if value is None: |
| 2044 | continue |
| 2045 | if self.speculative_config is None: |
| 2046 | self.speculative_config = {} |
| 2047 | if key in self.speculative_config: |
| 2048 | raise ValueError( |
| 2049 | f"{flag} and --speculative-config['{key}'] are mutually exclusive" |
| 2050 | ) |
| 2051 | self.speculative_config[key] = value |
| 2052 | |
| 2053 | if self.speculative_config is None: |
| 2054 | return None |
| 2055 | |
| 2056 | self.speculative_config = { |
| 2057 | k.replace("-", "_"): v for k, v in self.speculative_config.items() |
| 2058 | } |
| 2059 | |
| 2060 | # Note(Shangming): These parameters are not obtained from the cli arg |
| 2061 | # '--speculative-config' and must be passed in when creating the engine |
| 2062 | # config. |
| 2063 | self.speculative_config.update( |
| 2064 | { |
| 2065 | "target_model_config": target_model_config, |
| 2066 | "target_parallel_config": target_parallel_config, |
| 2067 | } |
| 2068 | ) |
| 2069 | return SpeculativeConfig(**self.speculative_config) |
| 2070 | |
| 2071 | def _resolve_device_ids(self) -> list[int] | None: |
| 2072 | if not self.device_ids: |
| 2073 | return None |
| 2074 | if self.distributed_executor_backend == "ray": |
| 2075 | logger.warning( |
| 2076 | "--device-ids has no effect when using the Ray executor. " |
| 2077 | "Use Ray placement groups for GPU selection instead." |
| 2078 | ) |
| 2079 | ids = self.device_ids |
| 2080 | if len(set(ids)) != len(ids): |
| 2081 | raise ValueError(f"--device-ids must not contain duplicates: {ids}") |
| 2082 | if all(isinstance(i, str) for i in ids): |
| 2083 | return [ |
| 2084 | current_platform.device_control_id_to_physical_device_id(i) |
| 2085 | for i in cast(list[str], ids) |
| 2086 | ] |
| 2087 | if any(isinstance(i, str) for i in ids): |
| 2088 | raise ValueError("--device-ids must not mix integer IDs and UUIDs") |
| 2089 | int_ids = cast(list[int], ids) |
| 2090 | # Compose with CUDA_VISIBLE_DEVICES: if CVD is set, treat |
| 2091 | # --device-ids values as indices into the CVD-visible set. |
| 2092 | cvd = getattr( |
| 2093 | envs, |
| 2094 | current_platform.device_control_env_var, |
| 2095 | os.environ.get(current_platform.device_control_env_var), |
| 2096 | ) |
| 2097 | if cvd: |
| 2098 | cvd_ids = [ |
| 2099 | current_platform.device_control_id_to_physical_device_id(x) |
| 2100 | for x in cvd.split(",") |
| 2101 | ] |
| 2102 | for i in int_ids: |
| 2103 | if i >= len(cvd_ids): |
| 2104 | raise ValueError( |
| 2105 | f"--device-ids index {i} is out of range for " |
| 2106 | f"{current_platform.device_control_env_var}" |
| 2107 | f"={cvd} ({len(cvd_ids)} devices visible)" |
| 2108 | ) |
| 2109 | return [cvd_ids[i] for i in int_ids] |
| 2110 | return int_ids |
| 2111 | |
| 2112 | def create_diffusion_config(self) -> DiffusionConfig | None: |
| 2113 | if self.diffusion_config is None: |
| 2114 | return None |
| 2115 | cfg = self.diffusion_config |
| 2116 | if isinstance(cfg, str): |
| 2117 | cfg = json.loads(cfg) |
| 2118 | return DiffusionConfig(**cfg) |
| 2119 | |
| 2120 | def create_watermark_config(self) -> WatermarkConfig | None: |
| 2121 | if self.watermark_config is None: |
| 2122 | return None |
| 2123 | cfg = self.watermark_config |
| 2124 | if isinstance(cfg, str): |
| 2125 | cfg = json.loads(cfg) |
| 2126 | return WatermarkConfig(**cfg) |
| 2127 | |
| 2128 | def create_observability_config(self) -> ObservabilityConfig: |
| 2129 | return ObservabilityConfig( |
| 2130 | show_hidden_metrics_for_version=self.show_hidden_metrics_for_version, |
| 2131 | otlp_traces_endpoint=self.otlp_traces_endpoint, |
| 2132 | collect_detailed_traces=self.collect_detailed_traces, |
| 2133 | per_request_spec_decode_metrics=self.per_request_spec_decode_metrics, |
| 2134 | kv_cache_metrics=self.kv_cache_metrics, |
| 2135 | kv_cache_metrics_sample=self.kv_cache_metrics_sample, |
| 2136 | custom_histogram_buckets=self.custom_histogram_buckets, |
| 2137 | cudagraph_metrics=self.cudagraph_metrics, |
| 2138 | enable_layerwise_nvtx_tracing=self.enable_layerwise_nvtx_tracing, |
| 2139 | enable_mfu_metrics=self.enable_mfu_metrics, |
| 2140 | enable_mm_processor_stats=self.enable_mm_processor_stats, |
| 2141 | enable_logging_iteration_details=self.enable_logging_iteration_details, |
| 2142 | jit_monitor_mode=self.jit_monitor_mode, |
| 2143 | jit_monitor_verbose=self.jit_monitor_verbose, |
| 2144 | ) |
| 2145 | |
| 2146 | def create_logging_config(self) -> LoggingConfig: |
| 2147 | config = self.logging_config or LoggingConfig() |
| 2148 | if self.log_level is not None: |
| 2149 | config = dataclasses.replace(config, log_level=self.log_level) |
| 2150 | if self.log_config_file is not None: |
| 2151 | config = dataclasses.replace( |
| 2152 | config, pylogging_config_file=self.log_config_file |
| 2153 | ) |
| 2154 | return config |
| 2155 | |
| 2156 | def create_engine_config( |
| 2157 | self, |
| 2158 | usage_context: UsageContext | None = None, |
| 2159 | headless: bool = False, |
| 2160 | ) -> VllmConfig: |
| 2161 | """Create the VllmConfig. |
| 2162 | |
| 2163 | NOTE: If VllmConfig is incompatible, we raise an error. |
| 2164 | """ |
| 2165 | logging_config = self.create_logging_config() |
| 2166 | configure_logging_if_needed(logging_config) |
| 2167 | |
| 2168 | current_platform.pre_register_and_update() |
| 2169 | |
| 2170 | device_config = DeviceConfig(device=cast(Device, current_platform.device_type)) |
| 2171 | |
| 2172 | current_platform.validate_environ(self.fail_on_environ_validation) |
| 2173 | |
| 2174 | # Check if the model is a speculator and override model/tokenizer/config |
| 2175 | # BEFORE creating ModelConfig, so the config is created with the target model |
| 2176 | # Skip speculator detection for cloud storage models (eg: S3, GCS) since |
| 2177 | # HuggingFace cannot load configs directly from S3 URLs. S3 models can still |
| 2178 | # use speculators with explicit --speculative-config. |
| 2179 | if not is_cloud_storage(self.model): |
| 2180 | (self.model, self.tokenizer, self.speculative_config) = ( |
| 2181 | maybe_override_with_speculators( |
| 2182 | model=self.model, |
| 2183 | tokenizer=self.tokenizer, |
| 2184 | revision=self.revision, |
| 2185 | trust_remote_code=self.trust_remote_code, |
| 2186 | vllm_speculative_config=self.speculative_config, |
| 2187 | hf_token=self.hf_token, |
| 2188 | ) |
| 2189 | ) |
| 2190 | |
| 2191 | model_config = self.create_model_config() |
| 2192 | self.model = model_config.model |
| 2193 | self.model_weights = model_config.model_weights |
| 2194 | self.tokenizer = model_config.tokenizer |
| 2195 | |
| 2196 | self._check_feature_supported() |
| 2197 | self._set_default_chunked_prefill_and_prefix_caching_args(model_config) |
| 2198 | self._set_default_reasoning_config_args() |
| 2199 | sliding_window: int | None = None |
| 2200 | layer_types = getattr(model_config.hf_text_config, "layer_types", None) |
| 2201 | if layer_types is None or all(lt == "sliding_attention" for lt in layer_types): |
| 2202 | # Only set CacheConfig.sliding_window if the model is all sliding |
| 2203 | # window. Otherwise CacheConfig.sliding_window will override the |
| 2204 | # global layers in interleaved sliding window models. |
| 2205 | sliding_window = model_config.get_sliding_window() |
| 2206 | |
| 2207 | # Resolve "auto" kv_cache_dtype to actual value from model config |
| 2208 | resolved_cache_dtype = resolve_kv_cache_dtype_string( |
| 2209 | self.kv_cache_dtype, model_config |
| 2210 | ) |
| 2211 | |
| 2212 | assert self.enable_prefix_caching is not None, ( |
| 2213 | "enable_prefix_caching must be set by this point" |
| 2214 | ) |
| 2215 | |
| 2216 | cache_config = CacheConfig( |
| 2217 | block_size=self.block_size, # type: ignore[arg-type] |
| 2218 | gpu_memory_utilization=self.gpu_memory_utilization, |
| 2219 | kv_cache_memory_bytes=self.kv_cache_memory_bytes, |
| 2220 | cache_dtype=resolved_cache_dtype, # type: ignore[arg-type] |
| 2221 | is_attention_free=model_config.is_attention_free, |
| 2222 | num_gpu_blocks_override=self.num_gpu_blocks_override, |
| 2223 | sliding_window=sliding_window, |
| 2224 | enable_prefix_caching=self.enable_prefix_caching, |
| 2225 | prefix_caching_hash_algo=self.prefix_caching_hash_algo, |
| 2226 | prefix_cache_retention_interval=self.prefix_cache_retention_interval, |
| 2227 | kv_cache_dtype_skip_layers=self.kv_cache_dtype_skip_layers, |
| 2228 | kv_sharing_fast_prefill=self.kv_sharing_fast_prefill, |
| 2229 | swa_bounded_replay=self.swa_bounded_replay, |
| 2230 | mamba_cache_dtype=self.mamba_cache_dtype, |
| 2231 | mamba_ssm_cache_dtype=self.mamba_ssm_cache_dtype, |
| 2232 | mamba_block_size=self.mamba_block_size, |
| 2233 | prefix_match_unit=self.prefix_match_unit, |
| 2234 | mamba_cache_mode=self.mamba_cache_mode, |
| 2235 | enable_mamba_shared_prefix_checkpoint=( |
| 2236 | self.enable_mamba_shared_prefix_checkpoint |
| 2237 | ), |
| 2238 | replayssm_buffer_len=self.replayssm_buffer_len, |
| 2239 | use_replayssm=self.use_replayssm, |
| 2240 | kv_offloading_size=self.kv_offloading_size, |
| 2241 | kv_offloading_backend=self.kv_offloading_backend, |
| 2242 | ) |
| 2243 | |
| 2244 | if resolved_cache_dtype.startswith("turboquant_"): |
| 2245 | from vllm.model_executor.layers.quantization.turboquant.config import ( |
| 2246 | TurboQuantConfig, |
| 2247 | ) |
| 2248 | |
| 2249 | boundary = TurboQuantConfig.get_boundary_skip_layers(model_config) |
| 2250 | existing = set(cache_config.kv_cache_dtype_skip_layers) |
| 2251 | cache_config.kv_cache_dtype_skip_layers = sorted( |
| 2252 | existing | set(boundary), key=int |
| 2253 | ) |
| 2254 | |
| 2255 | ray_runtime_env = None |
| 2256 | if is_ray_initialized(): |
| 2257 | # Ray Serve LLM calls `create_engine_config` in the context |
| 2258 | # of a Ray task, therefore we check is_ray_initialized() |
| 2259 | # as opposed to is_in_ray_actor(). |
| 2260 | import ray |
| 2261 | |
| 2262 | ray_runtime_env = ray.get_runtime_context().runtime_env |
| 2263 | # Avoid logging sensitive environment variables |
| 2264 | sanitized_env = ray_runtime_env.to_dict() if ray_runtime_env else {} |
| 2265 | if "env_vars" in sanitized_env: |
| 2266 | sanitized_env["env_vars"] = { |
| 2267 | k: "***" for k in sanitized_env["env_vars"] |
| 2268 | } |
| 2269 | logger.info("Using ray runtime env (env vars redacted): %s", sanitized_env) |
| 2270 | |
| 2271 | # Get the current placement group if Ray is initialized and |
| 2272 | # we are in a Ray actor. If so, then the placement group will be |
| 2273 | # passed to spawned processes. |
| 2274 | placement_group = None |
| 2275 | if is_in_ray_actor(): |
| 2276 | import ray |
| 2277 | |
| 2278 | # This call initializes Ray automatically if it is not initialized, |
| 2279 | # but we should not do this here. |
| 2280 | placement_group = ray.util.get_current_placement_group() |
| 2281 | |
| 2282 | assert not headless or not self.data_parallel_hybrid_lb, ( |
| 2283 | "data_parallel_hybrid_lb is not applicable in headless mode" |
| 2284 | ) |
| 2285 | if self.data_parallel_hybrid_lb and self.data_parallel_external_lb: |
| 2286 | raise ValueError( |
| 2287 | "Invalid data-parallel launch options: " |
| 2288 | "`--data-parallel-hybrid-lb` and " |
| 2289 | "`--data-parallel-external-lb` cannot be enabled together. " |
| 2290 | "Enable only one load-balancing mode." |
| 2291 | ) |
| 2292 | if self.nnodes > 1 and self.data_parallel_backend != "mp": |
| 2293 | raise ValueError( |
| 2294 | "Invalid data-parallel launch options: " |
| 2295 | f"`--nnodes {self.nnodes}` requires " |
| 2296 | "`--data-parallel-backend mp`; got " |
| 2297 | f"`--data-parallel-backend {self.data_parallel_backend}`. " |
| 2298 | "Use the MP backend or set `--nnodes 1`." |
| 2299 | ) |
| 2300 | inferred_data_parallel_rank = 0 |
| 2301 | if self.nnodes > 1: |
| 2302 | world_size_within_dp = ( |
| 2303 | self.pipeline_parallel_size |
| 2304 | * self.tensor_parallel_size |
| 2305 | * self.prefill_context_parallel_size |
| 2306 | ) |
| 2307 | world_size = self.data_parallel_size * world_size_within_dp |
| 2308 | if world_size % self.nnodes != 0: |
| 2309 | raise ValueError( |
| 2310 | "Invalid data-parallel launch options: " |
| 2311 | f"`--nnodes {self.nnodes}` must evenly divide the total " |
| 2312 | f"world size ({world_size}). Adjust `--nnodes`, " |
| 2313 | "`--data-parallel-size`, `--pipeline-parallel-size`, " |
| 2314 | "`--tensor-parallel-size`, or `--prefill-context-parallel-size`." |
| 2315 | ) |
| 2316 | if not 0 <= self.node_rank < self.nnodes: |
| 2317 | raise ValueError( |
| 2318 | "Invalid data-parallel launch options: `--node-rank` must " |
| 2319 | f"be between 0 and {self.nnodes - 1}; got " |
| 2320 | f"`--node-rank {self.node_rank}`. Set it to this node's " |
| 2321 | "zero-based index." |
| 2322 | ) |
| 2323 | local_world_size = world_size // self.nnodes |
| 2324 | inferred_data_parallel_rank = ( |
| 2325 | self.node_rank * local_world_size |
| 2326 | ) // world_size_within_dp |
| 2327 | if self.data_parallel_size > 1 and self.data_parallel_external_lb: |
| 2328 | if self.data_parallel_rank is None: |
| 2329 | if self.nnodes % self.data_parallel_size != 0: |
| 2330 | raise ValueError( |
| 2331 | "Invalid data-parallel launch options: " |
| 2332 | "`--node-rank` cannot unambiguously identify external " |
| 2333 | "data-parallel ranks when `--nnodes` is not divisible " |
| 2334 | "by `--data-parallel-size`. Set a unique " |
| 2335 | "`--data-parallel-rank` for each external-LB process." |
| 2336 | ) |
| 2337 | self.data_parallel_rank = inferred_data_parallel_rank |
| 2338 | logger.info( |
| 2339 | "Inferred data_parallel_rank %d from node_rank %d " |
| 2340 | "for external lb", |
| 2341 | self.data_parallel_rank, |
| 2342 | self.node_rank, |
| 2343 | ) |
| 2344 | elif self.data_parallel_size_local is None: |
| 2345 | # Infer data parallel size local for internal dplb: |
| 2346 | self.data_parallel_size_local = max( |
| 2347 | local_world_size // world_size_within_dp, 1 |
| 2348 | ) |
| 2349 | data_parallel_external_lb = ( |
| 2350 | self.data_parallel_external_lb or self.data_parallel_rank is not None |
| 2351 | ) |
| 2352 | if self.enable_fault_tolerance and not data_parallel_external_lb: |
| 2353 | raise ValueError( |
| 2354 | "Fault tolerance requires external load balancer mode " |
| 2355 | "(--data-parallel-external-lb or --data-parallel-rank). " |
| 2356 | "Internal LB mode is not supported." |
| 2357 | ) |
| 2358 | if ( |
| 2359 | self.data_parallel_size > 1 |
| 2360 | and data_parallel_external_lb |
| 2361 | and not model_config.is_moe |
| 2362 | ): |
| 2363 | raise ValueError( |
| 2364 | "Non-MoE models do not support external data parallel mode. " |
| 2365 | "For external load balancing, launch independent vLLM " |
| 2366 | "instances without --data-parallel-* arguments." |
| 2367 | ) |
| 2368 | # Local DP rank = 1, use pure-external LB. |
| 2369 | if data_parallel_external_lb: |
| 2370 | if self.data_parallel_rank is None: |
| 2371 | raise ValueError( |
| 2372 | "Invalid data-parallel launch options: " |
| 2373 | "`--data-parallel-external-lb` requires a data-parallel " |
| 2374 | "rank. Set `--data-parallel-rank`, or set " |
| 2375 | "`--data-parallel-size` greater than 1 and use `--nnodes` " |
| 2376 | "with `--node-rank` so the rank can be inferred." |
| 2377 | ) |
| 2378 | if self.data_parallel_size_local not in (1, None): |
| 2379 | raise ValueError( |
| 2380 | "Invalid data-parallel launch options: an external " |
| 2381 | "data-parallel rank requires `--data-parallel-size-local " |
| 2382 | f"1`; got {self.data_parallel_size_local}. Set it to 1 or " |
| 2383 | "omit it." |
| 2384 | ) |
| 2385 | data_parallel_size_local = 1 |
| 2386 | # Use full external lb if we have local_size of 1. |
| 2387 | self.data_parallel_hybrid_lb = False |
| 2388 | elif self.data_parallel_size_local is not None: |
| 2389 | data_parallel_size_local = self.data_parallel_size_local |
| 2390 | |
| 2391 | if self.data_parallel_start_rank is not None and not headless: |
| 2392 | # Infer hybrid LB mode. |
| 2393 | self.data_parallel_hybrid_lb = True |
| 2394 | |
| 2395 | if self.data_parallel_hybrid_lb and data_parallel_size_local == 1: |
| 2396 | # Use full external lb if we have local_size of 1. |
| 2397 | logger.warning( |
| 2398 | "data_parallel_hybrid_lb is not eligible when " |
| 2399 | "data_parallel_size_local = 1, autoswitch to " |
| 2400 | "data_parallel_external_lb." |
| 2401 | ) |
| 2402 | data_parallel_external_lb = True |
| 2403 | self.data_parallel_hybrid_lb = False |
| 2404 | |
| 2405 | if data_parallel_size_local == self.data_parallel_size: |
| 2406 | # Disable hybrid LB mode if set for a single node |
| 2407 | self.data_parallel_hybrid_lb = False |
| 2408 | |
| 2409 | self.data_parallel_rank = ( |
| 2410 | self.data_parallel_start_rank |
| 2411 | if self.data_parallel_start_rank is not None |
| 2412 | else inferred_data_parallel_rank |
| 2413 | ) |
| 2414 | if self.nnodes > 1: |
| 2415 | logger.info( |
| 2416 | "Inferred data_parallel_rank %d from node_rank %d", |
| 2417 | self.data_parallel_rank, |
| 2418 | self.node_rank, |
| 2419 | ) |
| 2420 | else: |
| 2421 | if self.data_parallel_hybrid_lb: |
| 2422 | raise ValueError( |
| 2423 | "Invalid data-parallel launch options: " |
| 2424 | "`--data-parallel-hybrid-lb` requires " |
| 2425 | "`--data-parallel-size-local`. Set it to the number of " |
| 2426 | "data-parallel ranks on this node." |
| 2427 | ) |
| 2428 | |
| 2429 | if self.data_parallel_backend == "ray" and ( |
| 2430 | envs.VLLM_RAY_DP_PACK_STRATEGY == "span" |
| 2431 | ): |
| 2432 | # Data parallel size defaults to 1 if DP ranks are spanning |
| 2433 | # multiple nodes |
| 2434 | data_parallel_size_local = 1 |
| 2435 | else: |
| 2436 | # Otherwise local DP size defaults to global DP size if not set |
| 2437 | data_parallel_size_local = self.data_parallel_size |
| 2438 | |
| 2439 | # DP address, used in multi-node case for torch distributed group |
| 2440 | # and ZMQ sockets. |
| 2441 | if self.data_parallel_address is None: |
| 2442 | if self.data_parallel_backend == "ray": |
| 2443 | host_ip = get_ip() |
| 2444 | logger.info( |
| 2445 | "Using host IP %s as ray-based data parallel address", host_ip |
| 2446 | ) |
| 2447 | data_parallel_address = host_ip |
| 2448 | else: |
| 2449 | assert self.data_parallel_backend == "mp", ( |
| 2450 | "data_parallel_backend can only be ray or mp, got %s", |
| 2451 | self.data_parallel_backend, |
| 2452 | ) |
| 2453 | data_parallel_address = ( |
| 2454 | self.master_addr or ParallelConfig.data_parallel_master_ip |
| 2455 | ) |
| 2456 | else: |
| 2457 | data_parallel_address = self.data_parallel_address |
| 2458 | |
| 2459 | # This port is only used when there are remote data parallel engines, |
| 2460 | # otherwise the local IPC transport is used. |
| 2461 | data_parallel_rpc_port = ( |
| 2462 | self.data_parallel_rpc_port |
| 2463 | if (self.data_parallel_rpc_port is not None) |
| 2464 | else ParallelConfig.data_parallel_rpc_port |
| 2465 | ) |
| 2466 | |
| 2467 | if self.tokens_only and not model_config.skip_tokenizer_init: |
| 2468 | model_config.skip_tokenizer_init = True |
| 2469 | logger.info("Skipping tokenizer initialization for tokens-only mode.") |
| 2470 | |
| 2471 | parallel_config = ParallelConfig( |
| 2472 | pipeline_parallel_size=self.pipeline_parallel_size, |
| 2473 | tensor_parallel_size=self.tensor_parallel_size, |
| 2474 | prefill_context_parallel_size=self.prefill_context_parallel_size, |
| 2475 | data_parallel_size=self.data_parallel_size, |
| 2476 | data_parallel_rank=self.data_parallel_rank or 0, |
| 2477 | data_parallel_external_lb=data_parallel_external_lb, |
| 2478 | data_parallel_size_local=data_parallel_size_local, |
| 2479 | master_addr=self.master_addr, |
| 2480 | master_port=self.master_port, |
| 2481 | nnodes=self.nnodes, |
| 2482 | node_rank=self.node_rank, |
| 2483 | distributed_timeout_seconds=self.distributed_timeout_seconds, |
| 2484 | cpu_distributed_timeout_seconds=self.cpu_distributed_timeout_seconds, |
| 2485 | data_parallel_master_ip=data_parallel_address, |
| 2486 | data_parallel_rpc_port=data_parallel_rpc_port, |
| 2487 | data_parallel_backend=self.data_parallel_backend, |
| 2488 | data_parallel_hybrid_lb=self.data_parallel_hybrid_lb, |
| 2489 | is_moe_model=model_config.is_moe, |
| 2490 | enable_expert_parallel=self.enable_expert_parallel, |
| 2491 | enable_batch_sharded_sampling=self.enable_batch_sharded_sampling, |
| 2492 | enable_ep_weight_filter=self.enable_ep_weight_filter, |
| 2493 | all2all_backend=self.all2all_backend, |
| 2494 | enable_elastic_ep=self.enable_elastic_ep, |
| 2495 | elastic_ep_max_dp_size=self.elastic_ep_max_dp_size, |
| 2496 | enable_dbo=self.enable_dbo, |
| 2497 | ubatch_size=self.ubatch_size, |
| 2498 | dbo_decode_token_threshold=self.dbo_decode_token_threshold, |
| 2499 | dp_sync_interval=self.dp_sync_interval, |
| 2500 | dbo_prefill_token_threshold=self.dbo_prefill_token_threshold, |
| 2501 | disable_nccl_for_dp_synchronization=self.disable_nccl_for_dp_synchronization, |
| 2502 | enable_eplb=self.enable_eplb, |
| 2503 | eplb_config=self.eplb_config, |
| 2504 | expert_placement_strategy=self.expert_placement_strategy, |
| 2505 | max_parallel_loading_workers=self.max_parallel_loading_workers, |
| 2506 | disable_custom_all_reduce=self.disable_custom_all_reduce, |
| 2507 | ray_workers_use_nsight=self.ray_workers_use_nsight, |
| 2508 | ray_runtime_env=ray_runtime_env, |
| 2509 | placement_group=placement_group, |
| 2510 | distributed_executor_backend=self.distributed_executor_backend, |
| 2511 | worker_cls=self.worker_cls, |
| 2512 | worker_extension_cls=self.worker_extension_cls, |
| 2513 | decode_context_parallel_size=self.decode_context_parallel_size, |
| 2514 | dcp_comm_backend=self.dcp_comm_backend, |
| 2515 | dcp_q_replicate=self.dcp_q_replicate, |
| 2516 | dcp_kv_cache_interleave_size=self.dcp_kv_cache_interleave_size, |
| 2517 | cp_kv_cache_interleave_size=( |
| 2518 | self.cp_kv_cache_interleave_size |
| 2519 | if self.cp_kv_cache_interleave_size is not None |
| 2520 | else ParallelConfig.cp_kv_cache_interleave_size |
| 2521 | ), |
| 2522 | _allow_auto_resolve_cp_interleave_size=( |
| 2523 | self.cp_kv_cache_interleave_size is None |
| 2524 | ), |
| 2525 | _api_process_count=self._api_process_count, |
| 2526 | _api_process_rank=self._api_process_rank, |
| 2527 | assigned_physical_gpu_ids=self._resolve_device_ids(), |
| 2528 | enable_fault_tolerance=self.enable_fault_tolerance, |
| 2529 | fault_tolerance_config=self.fault_tolerance_config, |
| 2530 | numa_bind=self.numa_bind, |
| 2531 | numa_bind_nodes=self.numa_bind_nodes, |
| 2532 | numa_bind_cpus=self.numa_bind_cpus, |
| 2533 | ) |
| 2534 | |
| 2535 | speculative_config = self.create_speculative_config( |
| 2536 | target_model_config=model_config, |
| 2537 | target_parallel_config=parallel_config, |
| 2538 | ) |
| 2539 | diffusion_config = self.create_diffusion_config() |
| 2540 | watermark_config = self.create_watermark_config() |
| 2541 | |
| 2542 | self._set_default_max_num_seqs_and_batched_tokens_args( |
| 2543 | usage_context, |
| 2544 | model_config, |
| 2545 | parallel_config, |
| 2546 | ) |
| 2547 | |
| 2548 | assert self.max_num_batched_tokens is not None, ( |
| 2549 | "max_num_batched_tokens must be set by this point" |
| 2550 | ) |
| 2551 | assert self.max_num_seqs is not None, "max_num_seqs must be set by this point" |
| 2552 | assert self.enable_chunked_prefill is not None, ( |
| 2553 | "enable_chunked_prefill must be set by this point" |
| 2554 | ) |
| 2555 | assert model_config.max_model_len is not None, ( |
| 2556 | "max_model_len must be set by this point" |
| 2557 | ) |
| 2558 | scheduler_config = SchedulerConfig( |
| 2559 | runner_type=model_config.runner_type, |
| 2560 | max_num_batched_tokens=self.max_num_batched_tokens, |
| 2561 | max_num_scheduled_tokens=self.max_num_scheduled_tokens, |
| 2562 | max_num_seqs=self.max_num_seqs, |
| 2563 | max_num_active_seqs=self.max_num_active_seqs, |
| 2564 | max_num_queued_reqs=self.max_num_queued_reqs, |
| 2565 | max_num_queued_tokens=self.max_num_queued_tokens, |
| 2566 | max_model_len=model_config.max_model_len, |
| 2567 | enable_chunked_prefill=self.enable_chunked_prefill, |
| 2568 | disable_chunked_mm_input=self.disable_chunked_mm_input, |
| 2569 | is_multimodal_model=model_config.is_multimodal_model, |
| 2570 | is_encoder_decoder=model_config.is_encoder_decoder, |
| 2571 | policy=self.scheduling_policy, |
| 2572 | scheduler_cls=self.scheduler_cls, |
| 2573 | long_prefill_token_threshold=self.long_prefill_token_threshold, |
| 2574 | long_prefill_token_threshold_adaptive=( |
| 2575 | self.long_prefill_token_threshold_adaptive |
| 2576 | ), |
| 2577 | scheduler_reserve_full_isl=self.scheduler_reserve_full_isl, |
| 2578 | watermark=self.watermark, |
| 2579 | prefill_schedule_interval=self.prefill_schedule_interval, |
| 2580 | disable_hybrid_kv_cache_manager=self.disable_hybrid_kv_cache_manager, |
| 2581 | async_scheduling=self.async_scheduling, |
| 2582 | stream_interval=self.stream_interval, |
| 2583 | ) |
| 2584 | |
| 2585 | if not model_config.is_multimodal_model and self.default_mm_loras: |
| 2586 | raise ValueError( |
| 2587 | "Default modality-specific LoRA(s) were provided for a " |
| 2588 | "non multimodal model" |
| 2589 | ) |
| 2590 | |
| 2591 | lora_config = ( |
| 2592 | LoRAConfig( |
| 2593 | max_lora_rank=self.max_lora_rank, |
| 2594 | max_loras=self.max_loras, |
| 2595 | default_mm_loras=self.default_mm_loras, |
| 2596 | fully_sharded_loras=self.fully_sharded_loras, |
| 2597 | lora_dtype=self.lora_dtype, |
| 2598 | target_modules=self.lora_target_modules, |
| 2599 | enable_tower_connector_lora=self.enable_tower_connector_lora, |
| 2600 | specialize_active_lora=self.specialize_active_lora, |
| 2601 | enable_mixed_moe_lora_format=self.enable_mixed_moe_lora_format, |
| 2602 | enable_moe_shared_loras=self.enable_moe_shared_loras, |
| 2603 | max_lora_cls_labels=self.max_lora_cls_labels, |
| 2604 | max_cpu_loras=self.max_cpu_loras |
| 2605 | if self.max_cpu_loras and self.max_cpu_loras > 0 |
| 2606 | else None, |
| 2607 | ) |
| 2608 | if self.enable_lora |
| 2609 | else None |
| 2610 | ) |
| 2611 | |
| 2612 | if ( |
| 2613 | lora_config is not None |
| 2614 | and speculative_config is not None |
| 2615 | and scheduler_config.max_num_batched_tokens |
| 2616 | < ( |
| 2617 | scheduler_config.max_num_seqs |
| 2618 | * (speculative_config.num_speculative_tokens + 1) |
| 2619 | ) |
| 2620 | ): |
| 2621 | raise ValueError( |
| 2622 | "Consider increasing max_num_batched_tokens or " |
| 2623 | "decreasing num_speculative_tokens" |
| 2624 | ) |
| 2625 | |
| 2626 | # Attention config overrides |
| 2627 | attention_config = copy.deepcopy(self.attention_config) |
| 2628 | if self.attention_backend is not None: |
| 2629 | if attention_config.backend is not None: |
| 2630 | raise ValueError( |
| 2631 | "attention_backend and attention_config.backend " |
| 2632 | "are mutually exclusive" |
| 2633 | ) |
| 2634 | # Reuse the validator to handle "auto" and string-to-enum conversion |
| 2635 | attention_config.backend = AttentionConfig.validate_backend_before( |
| 2636 | self.attention_backend |
| 2637 | ) |
| 2638 | |
| 2639 | # Batch-invariant mode requires deterministic attention behavior. |
| 2640 | # If no backend is explicitly requested, prefer Triton Attention. |
| 2641 | if ( |
| 2642 | envs.VLLM_BATCH_INVARIANT |
| 2643 | and attention_config.backend is None |
| 2644 | and current_platform.is_xpu() |
| 2645 | ): |
| 2646 | attention_config.backend = AttentionBackendEnum.TRITON_ATTN |
| 2647 | logger.info( |
| 2648 | "VLLM_BATCH_INVARIANT is enabled and no attention backend was " |
| 2649 | "specified; defaulting to TRITON_ATTN." |
| 2650 | ) |
| 2651 | |
| 2652 | # TurboQuant requires FlashAttention 2 — FA3 boundary layers assert |
| 2653 | # FlashAttentionImpl which fails with TurboQuantAttentionImpl. |
| 2654 | if resolved_cache_dtype.startswith("turboquant_") and ( |
| 2655 | attention_config.flash_attn_version is None |
| 2656 | or attention_config.flash_attn_version >= 3 |
| 2657 | ): |
| 2658 | logger.warning( |
| 2659 | "TurboQuant is not yet compatible with FlashAttention >= 3. " |
| 2660 | "Overriding flash_attn_version to 2. To silence this " |
| 2661 | "warning, pass --attention-config.flash_attn_version=2" |
| 2662 | ) |
| 2663 | attention_config.flash_attn_version = 2 |
| 2664 | |
| 2665 | # Mamba config overrides |
| 2666 | mamba_config = copy.deepcopy(self.mamba_config) |
| 2667 | # Convert string to enum if needed (CLI parsing returns a string) |
| 2668 | if isinstance(self.mamba_backend, str): |
| 2669 | mamba_config.backend = MambaBackendEnum[self.mamba_backend.upper()] |
| 2670 | else: |
| 2671 | mamba_config.backend = self.mamba_backend |
| 2672 | if self.mamba_ssu_algorithm is not None: |
| 2673 | mamba_config.ssu_algorithm = self.mamba_ssu_algorithm |
| 2674 | if self.enable_mamba_cache_stochastic_rounding: |
| 2675 | mamba_config.enable_stochastic_rounding = ( |
| 2676 | self.enable_mamba_cache_stochastic_rounding |
| 2677 | ) |
| 2678 | if self.mamba_cache_philox_rounds: |
| 2679 | mamba_config.stochastic_rounding_philox_rounds = ( |
| 2680 | self.mamba_cache_philox_rounds |
| 2681 | ) |
| 2682 | mamba_config.validate_ssu_algorithm() |
| 2683 | |
| 2684 | # Kernel config overrides |
| 2685 | kernel_config = copy.deepcopy(self.kernel_config) |
| 2686 | if self.enable_flashinfer_autotune is not None: |
| 2687 | if kernel_config.enable_flashinfer_autotune is not None: |
| 2688 | raise ValueError( |
| 2689 | "enable_flashinfer_autotune and " |
| 2690 | "kernel_config.enable_flashinfer_autotune " |
| 2691 | "are mutually exclusive" |
| 2692 | ) |
| 2693 | kernel_config.enable_flashinfer_autotune = self.enable_flashinfer_autotune |
| 2694 | if self.moe_backend != "auto": |
| 2695 | kernel_config.moe_backend = self.moe_backend |
| 2696 | if self.linear_backend != "auto": |
| 2697 | kernel_config.linear_backend = self.linear_backend |
| 2698 | if self.sparse_indexer_topk_backend != "auto": |
| 2699 | kernel_config.sparse_indexer_topk_backend = self.sparse_indexer_topk_backend |
| 2700 | |
| 2701 | # Transfer top-level ir_op_priority into KernelConfig.ir_op_priority |
| 2702 | for op_name, op_priority in asdict(self.ir_op_priority).items(): |
| 2703 | # Empty means unset |
| 2704 | if not op_priority: |
| 2705 | continue |
| 2706 | |
| 2707 | # Priority cannot be set 2x for the same op |
| 2708 | if getattr(kernel_config.ir_op_priority, op_name): |
| 2709 | raise ValueError( |
| 2710 | f"Op priority for {op_name} specified via both ir_op_priority " |
| 2711 | f"and KernelConfig.ir_op_priority, only one allowed at a time." |
| 2712 | ) |
| 2713 | |
| 2714 | # Set the attribute |
| 2715 | setattr(kernel_config.ir_op_priority, op_name, op_priority) |
| 2716 | |
| 2717 | load_config = self.create_load_config() |
| 2718 | |
| 2719 | # Pass reasoning_parser into StructuredOutputsConfig |
| 2720 | if self.reasoning_parser: |
| 2721 | self.structured_outputs_config.reasoning_parser = self.reasoning_parser |
| 2722 | |
| 2723 | if self.reasoning_parser_plugin: |
| 2724 | self.structured_outputs_config.reasoning_parser_plugin = ( |
| 2725 | self.reasoning_parser_plugin |
| 2726 | ) |
| 2727 | |
| 2728 | observability_config = self.create_observability_config() |
| 2729 | |
| 2730 | # Compilation config overrides |
| 2731 | compilation_config = copy.deepcopy(self.compilation_config) |
| 2732 | if self.cudagraph_capture_sizes is not None: |
| 2733 | if compilation_config.cudagraph_capture_sizes is not None: |
| 2734 | raise ValueError( |
| 2735 | "cudagraph_capture_sizes and compilation_config." |
| 2736 | "cudagraph_capture_sizes are mutually exclusive" |
| 2737 | ) |
| 2738 | compilation_config.cudagraph_capture_sizes = self.cudagraph_capture_sizes |
| 2739 | if self.max_cudagraph_capture_size is not None: |
| 2740 | if compilation_config.max_cudagraph_capture_size is not None: |
| 2741 | raise ValueError( |
| 2742 | "max_cudagraph_capture_size and compilation_config." |
| 2743 | "max_cudagraph_capture_size are mutually exclusive" |
| 2744 | ) |
| 2745 | compilation_config.max_cudagraph_capture_size = ( |
| 2746 | self.max_cudagraph_capture_size |
| 2747 | ) |
| 2748 | |
| 2749 | offload_config = OffloadConfig( |
| 2750 | offload_backend=self.offload_backend, |
| 2751 | uva=UVAOffloadConfig( |
| 2752 | cpu_offload_gb=self.cpu_offload_gb, |
| 2753 | cpu_offload_params=self.cpu_offload_params, |
| 2754 | ), |
| 2755 | prefetch=PrefetchOffloadConfig( |
| 2756 | offload_group_size=self.offload_group_size, |
| 2757 | offload_num_in_group=self.offload_num_in_group, |
| 2758 | offload_prefetch_step=self.offload_prefetch_step, |
| 2759 | offload_params=self.offload_params, |
| 2760 | ), |
| 2761 | ) |
| 2762 | |
| 2763 | if self.gdn_prefill_backend is not None: |
| 2764 | self.additional_config["gdn_prefill_backend"] = self.gdn_prefill_backend |
| 2765 | if self.kda_prefill_backend is not None: |
| 2766 | if ( |
| 2767 | self.kda_prefill_backend == "flashkda" |
| 2768 | and not current_platform.is_cuda() |
| 2769 | ): |
| 2770 | raise ValueError( |
| 2771 | "--kda-prefill-backend=flashkda is only available on CUDA." |
| 2772 | ) |
| 2773 | if self.kda_prefill_backend == "fused" and not current_platform.is_rocm(): |
| 2774 | raise ValueError( |
| 2775 | "--kda-prefill-backend=fused is only available on ROCm." |
| 2776 | ) |
| 2777 | self.additional_config["kda_prefill_backend"] = self.kda_prefill_backend |
| 2778 | if self.kda_decode_backend is not None: |
| 2779 | self.additional_config["kda_decode_backend"] = self.kda_decode_backend |
| 2780 | |
| 2781 | config = VllmConfig( |
| 2782 | model_config=model_config, |
| 2783 | cache_config=cache_config, |
| 2784 | parallel_config=parallel_config, |
| 2785 | scheduler_config=scheduler_config, |
| 2786 | device_config=device_config, |
| 2787 | load_config=load_config, |
| 2788 | offload_config=offload_config, |
| 2789 | attention_config=attention_config, |
| 2790 | aux_output_config=self.aux_output_config, |
| 2791 | engram_config=self.engram_config, |
| 2792 | mamba_config=mamba_config, |
| 2793 | kernel_config=kernel_config, |
| 2794 | lora_config=lora_config, |
| 2795 | speculative_config=speculative_config, |
| 2796 | watermark_config=watermark_config, |
| 2797 | diffusion_config=diffusion_config, |
| 2798 | structured_outputs_config=self.structured_outputs_config, |
| 2799 | observability_config=observability_config, |
| 2800 | logging_config=logging_config, |
| 2801 | compilation_config=compilation_config, |
| 2802 | kv_transfer_config=self.kv_transfer_config, |
| 2803 | kv_events_config=self.kv_events_config, |
| 2804 | ec_transfer_config=self.ec_transfer_config, |
| 2805 | ec_manager_config=self.ec_manager_config, |
| 2806 | reasoning_config=self.reasoning_config, |
| 2807 | profiler_config=self.profiler_config, |
| 2808 | additional_config=self.additional_config, |
| 2809 | optimization_level=self.optimization_level, |
| 2810 | performance_mode=self.performance_mode, |
| 2811 | weight_transfer_config=self.weight_transfer_config, |
| 2812 | shutdown_timeout=self.shutdown_timeout, |
| 2813 | ) |
| 2814 | |
| 2815 | return config |
| 2816 | |
| 2817 | def _check_feature_supported(self): |
| 2818 | """Raise an error if the feature is not supported.""" |
| 2819 | if self.pipeline_parallel_size > 1: |
| 2820 | supports_pp = getattr( |
| 2821 | self.distributed_executor_backend, "supports_pp", False |
| 2822 | ) |
| 2823 | if not supports_pp and self.distributed_executor_backend not in ( |
| 2824 | ParallelConfig.distributed_executor_backend, |
| 2825 | "ray", |
| 2826 | "mp", |
| 2827 | "external_launcher", |
| 2828 | ): |
| 2829 | name = ( |
| 2830 | "Pipeline Parallelism without Ray distributed " |
| 2831 | "executor or multiprocessing executor or external " |
| 2832 | "launcher" |
| 2833 | ) |
| 2834 | _raise_unsupported_error(feature_name=name) |
| 2835 | |
| 2836 | @classmethod |
| 2837 | def get_batch_defaults( |
| 2838 | cls, |
| 2839 | world_size: int, |
| 2840 | ) -> tuple[dict[UsageContext | None, int], dict[UsageContext | None, int]]: |
| 2841 | from vllm.usage.usage_lib import UsageContext |
| 2842 | |
| 2843 | default_max_num_batched_tokens: dict[UsageContext | None, int] |
| 2844 | default_max_num_seqs: dict[UsageContext | None, int] |
| 2845 | |
| 2846 | # When no user override, set the default values based on the usage |
| 2847 | # context. |
| 2848 | # Use different default values for different hardware. |
| 2849 | |
| 2850 | # Try to query the device name on the current platform. If it fails, |
| 2851 | # it may be because the platform that imports vLLM is not the same |
| 2852 | # as the platform that vLLM is running on (e.g. the case of scaling |
| 2853 | # vLLM with Ray) and has no GPUs. In this case we use the default |
| 2854 | # values for non-H100/H200 GPUs. |
| 2855 | try: |
| 2856 | device_memory = current_platform.get_device_total_memory() |
| 2857 | device_name = current_platform.get_device_name().lower() |
| 2858 | except Exception: |
| 2859 | # This is only used to set default_max_num_batched_tokens |
| 2860 | device_memory = 0 |
| 2861 | device_name = "" |
| 2862 | |
| 2863 | # NOTE(Kuntai): Setting large `max_num_batched_tokens` for A100 reduces |
| 2864 | # throughput, see PR #17885 for more details. |
| 2865 | # So here we do an extra device name check to prevent such regression. |
| 2866 | if device_memory >= 160 * GiB_bytes: |
| 2867 | # for GPUs like B200/B300 with >= 160GB memory, use the largest defaults |
| 2868 | default_max_num_batched_tokens = { |
| 2869 | UsageContext.LLM_CLASS: 16384, |
| 2870 | UsageContext.OPENAI_API_SERVER: 16384, |
| 2871 | } |
| 2872 | default_max_num_seqs = { |
| 2873 | UsageContext.LLM_CLASS: 1024, |
| 2874 | UsageContext.OPENAI_API_SERVER: 1024, |
| 2875 | } |
| 2876 | elif device_memory >= 70 * GiB_bytes and "a100" not in device_name: |
| 2877 | # For GPUs like H100 and H200, use larger offline defaults. |
| 2878 | default_max_num_batched_tokens = { |
| 2879 | UsageContext.LLM_CLASS: 16384, |
| 2880 | UsageContext.OPENAI_API_SERVER: 8192, |
| 2881 | } |
| 2882 | default_max_num_seqs = { |
| 2883 | UsageContext.LLM_CLASS: 1024, |
| 2884 | UsageContext.OPENAI_API_SERVER: 1024, |
| 2885 | } |
| 2886 | else: |
| 2887 | # TODO(woosuk): Tune the default values for other hardware. |
| 2888 | default_max_num_batched_tokens = { |
| 2889 | UsageContext.LLM_CLASS: 8192, |
| 2890 | UsageContext.OPENAI_API_SERVER: 2048, |
| 2891 | } |
| 2892 | default_max_num_seqs = { |
| 2893 | UsageContext.LLM_CLASS: 256, |
| 2894 | UsageContext.OPENAI_API_SERVER: 256, |
| 2895 | } |
| 2896 | |
| 2897 | # tpu specific default values. |
| 2898 | if current_platform.is_tpu(): |
| 2899 | chip_name = current_platform.get_device_name() |
| 2900 | |
| 2901 | if chip_name == "V6E": |
| 2902 | default_max_num_batched_tokens = { |
| 2903 | UsageContext.LLM_CLASS: 2048, |
| 2904 | UsageContext.OPENAI_API_SERVER: 1024, |
| 2905 | } |
| 2906 | elif chip_name == "V5E": |
| 2907 | default_max_num_batched_tokens = { |
| 2908 | UsageContext.LLM_CLASS: 1024, |
| 2909 | UsageContext.OPENAI_API_SERVER: 512, |
| 2910 | } |
| 2911 | elif chip_name == "V5P": |
| 2912 | default_max_num_batched_tokens = { |
| 2913 | UsageContext.LLM_CLASS: 512, |
| 2914 | UsageContext.OPENAI_API_SERVER: 256, |
| 2915 | } |
| 2916 | |
| 2917 | # cpu specific default values. |
| 2918 | if current_platform.is_cpu(): |
| 2919 | default_max_num_batched_tokens = { |
| 2920 | UsageContext.LLM_CLASS: 4096 * world_size, |
| 2921 | UsageContext.OPENAI_API_SERVER: 2048 * world_size, |
| 2922 | } |
| 2923 | default_max_num_seqs = { |
| 2924 | UsageContext.LLM_CLASS: 256 * world_size, |
| 2925 | UsageContext.OPENAI_API_SERVER: 128 * world_size, |
| 2926 | } |
| 2927 | |
| 2928 | return default_max_num_batched_tokens, default_max_num_seqs |
| 2929 | |
| 2930 | def _set_default_chunked_prefill_and_prefix_caching_args( |
| 2931 | self, model_config: ModelConfig |
| 2932 | ) -> None: |
| 2933 | default_chunked_prefill = model_config.is_chunked_prefill_supported |
| 2934 | default_prefix_caching = model_config.is_prefix_caching_supported |
| 2935 | |
| 2936 | if self.enable_chunked_prefill is None: |
| 2937 | self.enable_chunked_prefill = default_chunked_prefill |
| 2938 | |
| 2939 | logger.debug( |
| 2940 | "%s chunked prefill by default", |
| 2941 | "Enabling" if default_chunked_prefill else "Disabling", |
| 2942 | ) |
| 2943 | elif ( |
| 2944 | model_config.runner_type == "generate" |
| 2945 | and not self.enable_chunked_prefill |
| 2946 | and default_chunked_prefill |
| 2947 | ): |
| 2948 | logger.warning_once( |
| 2949 | "This model does not officially support disabling chunked prefill. " |
| 2950 | "Disabling this manually may cause the engine to crash " |
| 2951 | "or produce incorrect outputs.", |
| 2952 | ) |
| 2953 | elif ( |
| 2954 | model_config.runner_type == "pooling" |
| 2955 | and self.enable_chunked_prefill |
| 2956 | and not default_chunked_prefill |
| 2957 | ): |
| 2958 | logger.warning_once( |
| 2959 | "This model does not officially support chunked prefill. " |
| 2960 | "Enabling this manually may cause the engine to crash " |
| 2961 | "or produce incorrect outputs.", |
| 2962 | ) |
| 2963 | |
| 2964 | if self.enable_prefix_caching is None: |
| 2965 | self.enable_prefix_caching = default_prefix_caching |
| 2966 | |
| 2967 | logger.debug( |
| 2968 | "%s prefix caching by default", |
| 2969 | "Enabling" if default_prefix_caching else "Disabling", |
| 2970 | ) |
| 2971 | elif ( |
| 2972 | model_config.runner_type == "pooling" |
| 2973 | and self.enable_prefix_caching |
| 2974 | and not default_prefix_caching |
| 2975 | ): |
| 2976 | logger.warning_once( |
| 2977 | "This model does not officially support prefix caching. " |
| 2978 | "Enabling this manually may cause the engine to crash " |
| 2979 | "or produce incorrect outputs.", |
| 2980 | ) |
| 2981 | |
| 2982 | # Disable chunked prefill and prefix caching for: |
| 2983 | # RISCV CPUs in V1 |
| 2984 | if current_platform.is_cpu() and current_platform.get_cpu_architecture() in ( |
| 2985 | CpuArchEnum.RISCV, |
| 2986 | ): |
| 2987 | logger.info( |
| 2988 | "Chunked prefill is not supported for" |
| 2989 | "RISC-V CPUs; " |
| 2990 | "disabling it for V1 backend." |
| 2991 | ) |
| 2992 | self.enable_chunked_prefill = False |
| 2993 | logger.info( |
| 2994 | "Prefix caching is not supported for " |
| 2995 | "RISC-V CPUs; " |
| 2996 | "disabling it for V1 backend." |
| 2997 | ) |
| 2998 | self.enable_prefix_caching = False |
| 2999 | |
| 3000 | def _set_default_reasoning_config_args(self): |
| 3001 | if not self.reasoning_parser: |
| 3002 | return |
| 3003 | if self.reasoning_config is None: |
| 3004 | self.reasoning_config = ReasoningConfig() |
| 3005 | self.reasoning_config.reasoning_parser = self.reasoning_parser |
| 3006 | |
| 3007 | @staticmethod |
| 3008 | def _get_min_mm_batched_tokens( |
| 3009 | model_config: ModelConfig, |
| 3010 | ) -> tuple[int, str] | None: |
| 3011 | """Get the minimum max_num_batched_tokens needed for a multimodal |
| 3012 | prefix-LM model to process at least one item of any supported modality. |
| 3013 | |
| 3014 | Returns (token_count, modality_name) for the most expensive modality, |
| 3015 | or None if the value cannot be determined at this stage. |
| 3016 | """ |
| 3017 | try: |
| 3018 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 3019 | |
| 3020 | # get_processing_info returns the model's multimodal processing |
| 3021 | # metadata (supported modalities, token limits) without loading |
| 3022 | # model weights or generating dummy data. |
| 3023 | info = MULTIMODAL_REGISTRY.get_processing_info(model_config) |
| 3024 | mm_counts = {modality: 1 for modality in info.supported_mm_limits} |
| 3025 | # get_mm_max_tokens_per_item returns pre-computed per-item token |
| 3026 | # ceilings for models that override it (e.g., Gemma4), or None |
| 3027 | # for models that rely on dummy-input profiling. When None is |
| 3028 | # returned we bail out — no dummy generation is triggered here. |
| 3029 | max_tokens = info.get_mm_max_tokens_per_item( |
| 3030 | seq_len=model_config.max_model_len, |
| 3031 | mm_counts=mm_counts, |
| 3032 | ) |
| 3033 | if max_tokens is not None: |
| 3034 | modality = max(max_tokens, key=max_tokens.__getitem__) |
| 3035 | return (max_tokens[modality], modality) |
| 3036 | except Exception as e: |
| 3037 | logger.warning("Failed to determine min multimodal batched tokens: %s", e) |
| 3038 | return None |
| 3039 | |
| 3040 | def _set_default_max_num_seqs_and_batched_tokens_args( |
| 3041 | self, |
| 3042 | usage_context: UsageContext | None, |
| 3043 | model_config: ModelConfig, |
| 3044 | parallel_config: ParallelConfig, |
| 3045 | ): |
| 3046 | world_size = self.pipeline_parallel_size * self.tensor_parallel_size |
| 3047 | ( |
| 3048 | default_max_num_batched_tokens, |
| 3049 | default_max_num_seqs, |
| 3050 | ) = self.get_batch_defaults(world_size) |
| 3051 | |
| 3052 | orig_max_num_batched_tokens = self.max_num_batched_tokens |
| 3053 | orig_max_num_seqs = self.max_num_seqs |
| 3054 | |
| 3055 | if self.max_num_batched_tokens is None: |
| 3056 | if parallel_config.use_batched_dp_moe: |
| 3057 | self.max_num_batched_tokens = ( |
| 3058 | SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS_FOR_BATCHED_DP |
| 3059 | ) |
| 3060 | else: |
| 3061 | self.max_num_batched_tokens = default_max_num_batched_tokens.get( |
| 3062 | usage_context, |
| 3063 | SchedulerConfig.DEFAULT_MAX_NUM_BATCHED_TOKENS, |
| 3064 | ) |
| 3065 | |
| 3066 | if self.max_num_seqs is None: |
| 3067 | self.max_num_seqs = default_max_num_seqs.get( |
| 3068 | usage_context, |
| 3069 | SchedulerConfig.DEFAULT_MAX_NUM_SEQS, |
| 3070 | ) |
| 3071 | |
| 3072 | # If throughput mode is set, double max_num_batched_tokens and max_num_seqs. |
| 3073 | if self.performance_mode == "throughput": |
| 3074 | if orig_max_num_batched_tokens is None: |
| 3075 | self.max_num_batched_tokens *= 2 |
| 3076 | if orig_max_num_seqs is None: |
| 3077 | self.max_num_seqs *= 2 |
| 3078 | |
| 3079 | if orig_max_num_batched_tokens is None: |
| 3080 | assert model_config.max_model_len is not None, ( |
| 3081 | "max_model_len must be set by this point" |
| 3082 | ) |
| 3083 | if not self.enable_chunked_prefill: |
| 3084 | # If max_model_len is too short, use the default for higher throughput. |
| 3085 | self.max_num_batched_tokens = max( |
| 3086 | model_config.max_model_len, |
| 3087 | self.max_num_batched_tokens, |
| 3088 | ) |
| 3089 | |
| 3090 | # For multimodal prefix-LM models (e.g., Gemma 4) that disable |
| 3091 | # chunked MM input, a single multimodal item must fit in one batch. |
| 3092 | # Raise the floor to accommodate the largest per-item token count. |
| 3093 | if model_config.is_multimodal_model and model_config.is_mm_prefix_lm: |
| 3094 | result = self._get_min_mm_batched_tokens(model_config) |
| 3095 | if result is not None and result[0] > self.max_num_batched_tokens: |
| 3096 | mm_min, modality = result |
| 3097 | logger.info( |
| 3098 | "Raising max_num_batched_tokens from %d to %d to " |
| 3099 | "accommodate '%s' input for prefix-LM model %s.", |
| 3100 | self.max_num_batched_tokens, |
| 3101 | mm_min, |
| 3102 | modality, |
| 3103 | model_config.model, |
| 3104 | ) |
| 3105 | self.max_num_batched_tokens = mm_min |
| 3106 | |
| 3107 | # When using default settings, |
| 3108 | # Ensure max_num_batched_tokens does not exceed model limit. |
| 3109 | # Some models (e.g., Whisper) have embeddings tied to max length. |
| 3110 | self.max_num_batched_tokens = min( |
| 3111 | self.max_num_seqs * model_config.max_model_len, |
| 3112 | self.max_num_batched_tokens, |
| 3113 | ) |
| 3114 | |
| 3115 | logger.debug( |
| 3116 | "Defaulting max_num_batched_tokens to %d for %s usage context.", |
| 3117 | self.max_num_batched_tokens, |
| 3118 | usage_context.value if usage_context else None, |
| 3119 | ) |
| 3120 | |
| 3121 | if orig_max_num_seqs is None: |
| 3122 | assert self.max_num_batched_tokens is not None # For type checking |
| 3123 | self.max_num_seqs = min(self.max_num_seqs, self.max_num_batched_tokens) |
| 3124 | |
| 3125 | logger.debug( |
| 3126 | "Defaulting max_num_seqs to %d for %s usage context.", |
| 3127 | self.max_num_seqs, |
| 3128 | usage_context.value if usage_context else None, |
| 3129 | ) |
| 3130 | |
| 3131 | |
| 3132 | @dataclass |
| 3133 | class AsyncEngineArgs(EngineArgs): |
| 3134 | """Arguments for asynchronous vLLM engine.""" |
| 3135 | |
| 3136 | enable_log_requests: bool = False |
| 3137 | |
| 3138 | @staticmethod |
| 3139 | def add_cli_args( |
| 3140 | parser: FlexibleArgumentParser, async_args_only: bool = False |
| 3141 | ) -> FlexibleArgumentParser: |
| 3142 | # Initialize plugin to update the parser, for example, The plugin may |
| 3143 | # add a new kind of quantization method to --quantization argument or |
| 3144 | # a new device to --device argument. |
| 3145 | load_general_plugins() |
| 3146 | if not async_args_only: |
| 3147 | parser = EngineArgs.add_cli_args(parser) |
| 3148 | parser.add_argument( |
| 3149 | "--enable-log-requests", |
| 3150 | action=argparse.BooleanOptionalAction, |
| 3151 | default=AsyncEngineArgs.enable_log_requests, |
| 3152 | help="Enable logging request information, dependent on log level:\n" |
| 3153 | "- INFO: Request ID, parameters and LoRA request.\n" |
| 3154 | "- DEBUG: Prompt inputs (e.g: text, token IDs).\n" |
| 3155 | "You can set the minimum log level via `VLLM_LOGGING_LEVEL`.", |
| 3156 | ) |
| 3157 | current_platform.pre_register_and_update(parser) |
| 3158 | return parser |
| 3159 | |
| 3160 | |
| 3161 | def _raise_unsupported_error(feature_name: str): |
| 3162 | msg = ( |
| 3163 | f"{feature_name} is not supported. We recommend to " |
| 3164 | f"remove {feature_name} from your config." |
| 3165 | ) |
| 3166 | raise NotImplementedError(msg) |