| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import itertools |
| 5 | from collections.abc import Callable, Iterable, Mapping |
| 6 | from contextlib import contextmanager |
| 7 | from dataclasses import dataclass, field, replace |
| 8 | from typing import TYPE_CHECKING, Any, Literal, Protocol, TypeAlias, overload |
| 9 | |
| 10 | import regex as re |
| 11 | import torch |
| 12 | import torch.nn as nn |
| 13 | from torch.nn.modules.module import register_module_module_registration_hook |
| 14 | |
| 15 | from vllm.config import VllmConfig |
| 16 | from vllm.distributed import ( |
| 17 | get_tensor_model_parallel_rank, |
| 18 | get_tensor_model_parallel_world_size, |
| 19 | ) |
| 20 | from vllm.logger import init_logger |
| 21 | from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding |
| 22 | from vllm.model_executor.model_loader.reload import ( |
| 23 | support_quantized_model_reload_from_hp_weights, |
| 24 | ) |
| 25 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 26 | from vllm.model_executor.models.interfaces import supports_any_eagle |
| 27 | from vllm.multimodal import NestedTensors |
| 28 | from vllm.sequence import IntermediateTensors |
| 29 | from vllm.utils.math_utils import cdiv |
| 30 | from vllm.utils.torch_utils import ( |
| 31 | direct_register_custom_op, |
| 32 | ) |
| 33 | |
| 34 | if TYPE_CHECKING: |
| 35 | from transformers import PreTrainedConfig |
| 36 | from transformers.conversion_mapping import WeightRenaming |
| 37 | |
| 38 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 39 | |
| 40 | logger = init_logger(__name__) |
| 41 | |
| 42 | ShardId: TypeAlias = str | int | tuple[int, ...] |
| 43 | """One shard of a stacked parameter. A tuple is a single contiguous span.""" |
| 44 | ShardIds: TypeAlias = ShardId | list[ShardId] |
| 45 | """One shard, or a list of shards the same weight is loaded into in turn.""" |
| 46 | |
| 47 | |
| 48 | @dataclass |
| 49 | class WeightsMapper: |
| 50 | """Maps the name of each weight if they match the following patterns. |
| 51 | |
| 52 | If a key maps to a value of `None`, the corresponding weight is ignored.""" |
| 53 | |
| 54 | orig_to_new_renaming: list["WeightRenaming"] = field(default_factory=list) |
| 55 | orig_to_new_regex: Mapping[re.Pattern, str | None] = field(default_factory=dict) |
| 56 | orig_to_new_substr: Mapping[str, str | None] = field(default_factory=dict) |
| 57 | orig_to_new_stacked: Mapping[str, tuple[str, ShardIds]] = field( |
| 58 | default_factory=dict |
| 59 | ) |
| 60 | orig_to_new_prefix: Mapping[str, str | None] = field(default_factory=dict) |
| 61 | orig_to_new_suffix: Mapping[str, str | None] = field(default_factory=dict) |
| 62 | |
| 63 | def __or__(self, other: "WeightsMapper") -> "WeightsMapper": |
| 64 | """Combine two `WeightsMapper`s by merging their mappings.""" |
| 65 | return WeightsMapper( |
| 66 | orig_to_new_renaming=[ |
| 67 | *self.orig_to_new_renaming, |
| 68 | *other.orig_to_new_renaming, |
| 69 | ], |
| 70 | orig_to_new_regex={**self.orig_to_new_regex, **other.orig_to_new_regex}, |
| 71 | orig_to_new_substr={**self.orig_to_new_substr, **other.orig_to_new_substr}, |
| 72 | orig_to_new_stacked={ |
| 73 | **self.orig_to_new_stacked, |
| 74 | **other.orig_to_new_stacked, |
| 75 | }, |
| 76 | orig_to_new_prefix={**self.orig_to_new_prefix, **other.orig_to_new_prefix}, |
| 77 | orig_to_new_suffix={**self.orig_to_new_suffix, **other.orig_to_new_suffix}, |
| 78 | ) |
| 79 | |
| 80 | def map_name(self, key: str) -> str | None: |
| 81 | """Map a weight name; returns ``None`` if the weight should be ignored.""" |
| 82 | result = self._map_name_with_shard(key) |
| 83 | return result[0] if result is not None else None |
| 84 | |
| 85 | def _map_name(self, key: str) -> str | None: |
| 86 | """Backward-compatible alias for :meth:`map_name`.""" |
| 87 | return self.map_name(key) |
| 88 | |
| 89 | def _map_name_with_shard(self, key: str) -> tuple[str, ShardIds | None] | None: |
| 90 | """Map a weight name and extract any shard_id metadata. |
| 91 | |
| 92 | Returns: |
| 93 | (mapped_name, shard_id) if the name should be kept. A list of shard |
| 94 | ids means the weight is loaded into each of those shards in turn. |
| 95 | None if the name should be dropped. |
| 96 | |
| 97 | """ |
| 98 | # Deprecation warnings |
| 99 | if key.endswith(".kv_scale"): |
| 100 | logger.warning_once( |
| 101 | "DEPRECATED. Found kv_scale in the checkpoint. " |
| 102 | "This format is deprecated in favor of separate k_scale and " |
| 103 | "v_scale tensors and will be removed in a future release. " |
| 104 | "Functionally, we will remap kv_scale to k_scale and duplicate " |
| 105 | "k_scale to v_scale" |
| 106 | ) |
| 107 | |
| 108 | for renaming in self.orig_to_new_renaming: |
| 109 | key, _ = renaming.rename_source_key(key) |
| 110 | |
| 111 | for pattern, new_key in self.orig_to_new_regex.items(): |
| 112 | if pattern.search(key): |
| 113 | if new_key is None: |
| 114 | return None |
| 115 | |
| 116 | key = pattern.sub(new_key, key) |
| 117 | |
| 118 | for substr, new_key in self.orig_to_new_substr.items(): |
| 119 | if substr in key: |
| 120 | if new_key is None: |
| 121 | return None |
| 122 | |
| 123 | key = key.replace(substr, new_key, 1) |
| 124 | |
| 125 | shard_id: ShardIds | None = None |
| 126 | for substr, (new_key, new_shard_id) in self.orig_to_new_stacked.items(): |
| 127 | if substr in key: |
| 128 | key = key.replace(substr, new_key, 1) |
| 129 | shard_id = new_shard_id |
| 130 | |
| 131 | for prefix, new_key in self.orig_to_new_prefix.items(): |
| 132 | if key.startswith(prefix): |
| 133 | if new_key is None: |
| 134 | return None |
| 135 | |
| 136 | key = key.replace(prefix, new_key, 1) |
| 137 | |
| 138 | for suffix, new_key in self.orig_to_new_suffix.items(): |
| 139 | if key.endswith(suffix): |
| 140 | if new_key is None: |
| 141 | return None |
| 142 | |
| 143 | key = new_key.join(key.rsplit(suffix, 1)) |
| 144 | |
| 145 | return key, shard_id |
| 146 | |
| 147 | def apply( |
| 148 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 149 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 150 | for name, data in weights: |
| 151 | result = self._map_name_with_shard(name) |
| 152 | if result is None: |
| 153 | continue |
| 154 | out_name, shard_id = result |
| 155 | if shard_id is None: |
| 156 | yield out_name, data |
| 157 | continue |
| 158 | shard_ids = shard_id if isinstance(shard_id, list) else [shard_id] |
| 159 | for i, one_shard_id in enumerate(shard_ids): |
| 160 | # Each shard carries its own id, so extra shards need their own |
| 161 | # tensor object; `detach` aliases the storage instead of copying. |
| 162 | shard_data = data if i == 0 else data.detach() |
| 163 | shard_data.shard_id = one_shard_id |
| 164 | yield out_name, shard_data |
| 165 | |
| 166 | def apply_list(self, values: list[str]) -> list[str]: |
| 167 | return [ |
| 168 | out_name |
| 169 | for name in values |
| 170 | if (out_name := self._map_name(name)) is not None |
| 171 | ] |
| 172 | |
| 173 | def apply_dict(self, values: dict[str, Any]) -> dict[str, Any]: |
| 174 | return { |
| 175 | out_name: value |
| 176 | for name, value in values.items() |
| 177 | if (out_name := self._map_name(name)) is not None |
| 178 | } |
| 179 | |
| 180 | def get_rename_mapper(self) -> "WeightsMapper": |
| 181 | """Mapper variant keeping only the renames. |
| 182 | |
| 183 | This is what consumers that *name* modules rather than load them need: |
| 184 | LoRA name parsing and the quantization config's layer lists. |
| 185 | |
| 186 | Stacked maps are dropped so that constituent names (e.g. `q_proj`) survive |
| 187 | rather than being rewritten to the stacked vLLM name (`qkv_proj`). Mappings to |
| 188 | `None` are dropped because "do not load this weight" is meaningless to such a |
| 189 | consumer, and applying it would silently shrink a quantization config's ignore |
| 190 | list or make LoRA name parsing fail.""" |
| 191 | remove_none = lambda d: {k: v for k, v in d.items() if v is not None} |
| 192 | return replace( |
| 193 | self, |
| 194 | orig_to_new_regex=remove_none(self.orig_to_new_regex), |
| 195 | orig_to_new_substr=remove_none(self.orig_to_new_substr), |
| 196 | orig_to_new_stacked={}, |
| 197 | orig_to_new_prefix=remove_none(self.orig_to_new_prefix), |
| 198 | orig_to_new_suffix=remove_none(self.orig_to_new_suffix), |
| 199 | ) |
| 200 | |
| 201 | |
| 202 | def _get_tied_embedding_params(module: nn.Module) -> dict[str, str]: |
| 203 | """Map each tied word embedding qualname to the first name it aliases.""" |
| 204 | canonical = dict[int, str]() |
| 205 | aliased = dict[str, str]() |
| 206 | for prefix, submodule in module.named_modules(remove_duplicate=False): |
| 207 | if not isinstance(submodule, VocabParallelEmbedding): |
| 208 | continue |
| 209 | for name, param in submodule.named_parameters(remove_duplicate=False): |
| 210 | qualname = f"{prefix}.{name}" if prefix else name |
| 211 | if (first_name := canonical.setdefault(id(param), qualname)) != qualname: |
| 212 | aliased[qualname] = first_name |
| 213 | return aliased |
| 214 | |
| 215 | |
| 216 | class AutoWeightsLoader: |
| 217 | """Helper class to load weights into a [`torch.nn.Module`][]. It is able |
| 218 | to automatically detect child modules and parameters while iterating over |
| 219 | the weights only once. |
| 220 | |
| 221 | The weight loading logic for individual modules can be overridden |
| 222 | by defining a `load_weights` method. |
| 223 | |
| 224 | Similarly, the weight loading logic for individual parameters can be |
| 225 | overridden by defining a `weight_loader` method. |
| 226 | |
| 227 | Detailed weight loading information can be viewed by setting the |
| 228 | environment variable `VLLM_LOGGING_LEVEL=DEBUG`. |
| 229 | """ |
| 230 | |
| 231 | # Models trained using early version ColossalAI or quantized by |
| 232 | # GPTQModel may include these tensors in checkpoint. Drop them. |
| 233 | REMOVE_UNUSED_ROTARY_EMBEDS_MAPPER = WeightsMapper( |
| 234 | orig_to_new_substr={ |
| 235 | "rotary_pos_emb.inv_freq": None, |
| 236 | "rotary_emb.inv_freq": None, |
| 237 | "rotary_emb.cos_cached": None, |
| 238 | "rotary_emb.sin_cached": None, |
| 239 | } |
| 240 | ) |
| 241 | |
| 242 | def __init__( |
| 243 | self, |
| 244 | module: nn.Module, |
| 245 | *, |
| 246 | ignore_unexpected_prefixes: list[str] | None = None, |
| 247 | ignore_unexpected_suffixes: list[str] | None = None, |
| 248 | ) -> None: |
| 249 | super().__init__() |
| 250 | |
| 251 | self.module = module |
| 252 | self.ignore_unexpected_prefixes = ignore_unexpected_prefixes or [] |
| 253 | self.ignore_unexpected_suffixes = ignore_unexpected_suffixes or [] |
| 254 | |
| 255 | # Weight tying makes two qualnames point at the same `nn.Parameter` |
| 256 | # (e.g. `lm_head.weight` and `model.embed_tokens.weight`). Loading both |
| 257 | # would write the same storage twice, so only the first name reached by |
| 258 | # module traversal is loaded and the rest are skipped. |
| 259 | self.aliased_params = _get_tied_embedding_params(module) |
| 260 | self._skipped_aliases = dict[str, str]() |
| 261 | self._loaded_params_are_complete = True |
| 262 | |
| 263 | def _groupby_prefix( |
| 264 | self, |
| 265 | weights: Iterable[tuple[str, torch.Tensor]], |
| 266 | ) -> Iterable[tuple[str, Iterable[tuple[str, torch.Tensor]]]]: |
| 267 | weights_by_parts = ( |
| 268 | (weight_name.split(".", 1), weight_data) |
| 269 | for weight_name, weight_data in weights |
| 270 | ) |
| 271 | |
| 272 | for prefix, group in itertools.groupby(weights_by_parts, key=lambda x: x[0][0]): |
| 273 | yield ( |
| 274 | prefix, |
| 275 | # Because maxsplit=1 in weight_name.split(...), |
| 276 | # the length of `parts` must either be 1 or 2 |
| 277 | ( |
| 278 | ("" if len(parts) == 1 else parts[1], weights_data) |
| 279 | for parts, weights_data in group |
| 280 | ), |
| 281 | ) |
| 282 | |
| 283 | def _get_qualname(self, prefix: str, rest: str) -> str: |
| 284 | if prefix == "": |
| 285 | return rest |
| 286 | if rest == "": |
| 287 | return prefix |
| 288 | |
| 289 | return ".".join((prefix, rest)) |
| 290 | |
| 291 | def _can_skip(self, qualname: str) -> bool: |
| 292 | return qualname in self.aliased_params |
| 293 | |
| 294 | def _can_ignore_unexpected(self, qualname: str) -> bool: |
| 295 | iup = (qualname.startswith(p) for p in self.ignore_unexpected_prefixes) |
| 296 | ius = (qualname.endswith(s) for s in self.ignore_unexpected_suffixes) |
| 297 | return any(iup) or any(ius) |
| 298 | |
| 299 | def _load_param( |
| 300 | self, |
| 301 | base_prefix: str, |
| 302 | param: nn.Parameter, |
| 303 | weights: Iterable[tuple[str, torch.Tensor]], |
| 304 | ) -> Iterable[str]: |
| 305 | for weight_name, weight_data in weights: |
| 306 | weight_qualname = self._get_qualname(base_prefix, weight_name) |
| 307 | |
| 308 | if self._can_skip(weight_qualname): |
| 309 | logger.debug("Skipping weight %s", weight_qualname) |
| 310 | |
| 311 | continue |
| 312 | |
| 313 | if weight_name != "": |
| 314 | if self._can_ignore_unexpected(weight_qualname): |
| 315 | logger.debug("Ignoring weight %s", weight_qualname) |
| 316 | |
| 317 | continue |
| 318 | |
| 319 | raise ValueError( |
| 320 | f"Attempted to load nested weight {weight_qualname!r} " |
| 321 | f"into a single parameter {base_prefix!r}" |
| 322 | ) |
| 323 | |
| 324 | if param.dtype != weight_data.dtype: |
| 325 | logger.warning( |
| 326 | "Attempted to load weight %s with dtype %s into " |
| 327 | "parameter with dtype %s", |
| 328 | weight_qualname, |
| 329 | weight_data.dtype, |
| 330 | param.dtype, |
| 331 | ) |
| 332 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 333 | weight_loader(param, weight_data) |
| 334 | |
| 335 | logger.debug("Loaded weight %s with shape %s", weight_qualname, param.shape) |
| 336 | |
| 337 | yield weight_qualname |
| 338 | |
| 339 | def _add_loadable_non_param_tensors( |
| 340 | self, module: nn.Module, child_params: dict[str, torch.Tensor] |
| 341 | ): |
| 342 | """Add tensor names that are not in the model params that may be in the |
| 343 | safetensors, e.g., batch normalization stats and registered buffers. |
| 344 | """ |
| 345 | # Add persistent registered buffers. |
| 346 | # Non-persistent buffers are excluded, matching PyTorch state_dict(). |
| 347 | non_persistent: set[str] = getattr(module, "_non_persistent_buffers_set", set()) |
| 348 | for buf_name, buf in module.named_buffers(recurse=False): |
| 349 | if buf_name not in child_params and buf_name not in non_persistent: |
| 350 | child_params[buf_name] = buf |
| 351 | |
| 352 | if isinstance( |
| 353 | module, |
| 354 | ( |
| 355 | nn.BatchNorm1d, |
| 356 | nn.BatchNorm2d, |
| 357 | nn.BatchNorm3d, |
| 358 | nn.LazyBatchNorm1d, |
| 359 | nn.LazyBatchNorm2d, |
| 360 | nn.LazyBatchNorm3d, |
| 361 | nn.SyncBatchNorm, |
| 362 | ), |
| 363 | ): |
| 364 | module_state_dict = module.state_dict() |
| 365 | for stat_name in ("running_mean", "running_var", "num_batches_tracked"): |
| 366 | child_params[stat_name] = module_state_dict[stat_name] |
| 367 | |
| 368 | def _load_module( |
| 369 | self, |
| 370 | base_prefix: str, |
| 371 | module: nn.Module, |
| 372 | weights: Iterable[tuple[str, torch.Tensor]], |
| 373 | ) -> Iterable[str]: |
| 374 | if isinstance(module, (StageMissingLayer, PPMissingLayer)): |
| 375 | return |
| 376 | |
| 377 | # Avoid infinite recursion since this function is typically |
| 378 | # called inside load_weights of the module itself |
| 379 | if module != self.module: |
| 380 | module_load_weights = getattr(module, "load_weights", None) |
| 381 | if callable(module_load_weights): |
| 382 | loaded_params = module_load_weights(weights) |
| 383 | if loaded_params is None: |
| 384 | logger.warning( |
| 385 | "Unable to collect loaded parameters for module %s", module |
| 386 | ) |
| 387 | self._loaded_params_are_complete = False |
| 388 | else: |
| 389 | yield from map( |
| 390 | lambda x: self._get_qualname(base_prefix, x), |
| 391 | loaded_params, |
| 392 | ) |
| 393 | |
| 394 | child_modules = dict(module.named_children()) |
| 395 | child_params = dict(module.named_parameters(recurse=False)) |
| 396 | |
| 397 | # Add missing tensors the weight loader needs to be able to load |
| 398 | # that aren't registered as params, e.g., batchnorm statistics. |
| 399 | self._add_loadable_non_param_tensors(module, child_params) |
| 400 | |
| 401 | for child_prefix, child_weights in self._groupby_prefix(weights): |
| 402 | prefix = self._get_qualname(base_prefix, child_prefix) |
| 403 | |
| 404 | if child_prefix in child_modules: |
| 405 | yield from self._load_module( |
| 406 | prefix, child_modules[child_prefix], child_weights |
| 407 | ) |
| 408 | elif child_prefix in child_params: |
| 409 | if self._can_skip(prefix): |
| 410 | logger.debug("Skipping param %s", prefix) |
| 411 | |
| 412 | continue |
| 413 | |
| 414 | yield from self._load_param( |
| 415 | prefix, child_params[child_prefix], child_weights |
| 416 | ) |
| 417 | else: |
| 418 | if self._can_skip(prefix): |
| 419 | logger.debug("Skipping missing %s", prefix) |
| 420 | |
| 421 | continue |
| 422 | |
| 423 | can_ignore_module = self._can_ignore_unexpected(prefix + ".") |
| 424 | can_ignore_param = self._can_ignore_unexpected(prefix) |
| 425 | if can_ignore_module or can_ignore_param: |
| 426 | logger.debug("Ignoring missing %s", prefix) |
| 427 | |
| 428 | continue |
| 429 | |
| 430 | named_parameters = module.named_parameters(recurse=True) |
| 431 | desc_param_keys = { |
| 432 | maybe_prefix(base_prefix, k) for k, _ in named_parameters |
| 433 | } |
| 434 | msg = ( |
| 435 | f"There is no module or parameter named {prefix!r} " |
| 436 | f"in {self.module._get_name()}. " |
| 437 | f"The available parameters belonging to {base_prefix} " |
| 438 | f"({module._get_name()}) are: {desc_param_keys}" |
| 439 | ) |
| 440 | raise ValueError(msg) |
| 441 | |
| 442 | @support_quantized_model_reload_from_hp_weights |
| 443 | def load_weights( |
| 444 | self, |
| 445 | weights: Iterable[tuple[str, torch.Tensor]], |
| 446 | *, |
| 447 | mapper: WeightsMapper | None = None, |
| 448 | ) -> set[str]: |
| 449 | # Ignore unexpected biases (typically from GPTQ models) |
| 450 | self.ignore_unexpected_suffixes.append(".bias") |
| 451 | |
| 452 | mapper = mapper or WeightsMapper() |
| 453 | # Many models store quant_config in the base model instead of the causal model. |
| 454 | # We look at the causal model's direct children for this reason. |
| 455 | modules = (self.module, *self.module.children()) |
| 456 | iterator = (m.quant_config for m in modules if hasattr(m, "quant_config")) |
| 457 | if quant_config := next(iterator, None): |
| 458 | # Apply mappings for quantization-specific checkpoint tensors. |
| 459 | mapper |= quant_config.get_cache_scale_mapper() |
| 460 | mapper |= quant_config.get_checkpoint_weight_mapper() |
| 461 | ignore_unexpected_suffixes = quant_config._ignore_unexpected_suffixes |
| 462 | self.ignore_unexpected_suffixes.extend(ignore_unexpected_suffixes) |
| 463 | mapper |= self.REMOVE_UNUSED_ROTARY_EMBEDS_MAPPER |
| 464 | |
| 465 | weights = mapper.apply(weights) |
| 466 | weights = self._filter_skipped(weights) |
| 467 | |
| 468 | autoloaded_weights = set(self._load_module("", self.module, weights)) |
| 469 | self._check_skipped_aliases(autoloaded_weights) |
| 470 | return autoloaded_weights |
| 471 | |
| 472 | def _filter_skipped( |
| 473 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 474 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 475 | for name, weight in weights: |
| 476 | if (canonical := self.aliased_params.get(name)) is not None: |
| 477 | self._skipped_aliases[name] = canonical |
| 478 | if self._can_skip(name): |
| 479 | continue |
| 480 | |
| 481 | yield name, weight |
| 482 | |
| 483 | def _check_skipped_aliases(self, autoloaded_weights: set[str]) -> None: |
| 484 | """Guard against skipping an alias whose canonical name never loads.""" |
| 485 | if not self._loaded_params_are_complete: |
| 486 | return |
| 487 | for alias, canonical in self._skipped_aliases.items(): |
| 488 | if canonical not in autoloaded_weights: |
| 489 | raise ValueError( |
| 490 | f"{alias!r} was skipped because it is tied to {canonical!r} " |
| 491 | f"in {self.module._get_name()}, but {canonical!r} was not " |
| 492 | "found in the checkpoint, so the tied weight is " |
| 493 | "uninitialized." |
| 494 | ) |
| 495 | |
| 496 | |
| 497 | def maybe_fuse_shared_experts( |
| 498 | weights: Iterable[tuple[str, torch.Tensor]], |
| 499 | *, |
| 500 | n_routed_experts: int, |
| 501 | n_shared_experts: int, |
| 502 | ckpt_prefix: str = "mlp.shared_experts", |
| 503 | enabled: bool | None = None, |
| 504 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 505 | """Route AITER fused-shared-expert checkpoint weights into fused slots. |
| 506 | |
| 507 | When AITER fused-shared-experts is active, shared experts are packed into |
| 508 | the routed expert tensor. The checkpoint stores them under `ckpt_prefix` |
| 509 | as a single (possibly widened) tensor; this splits it into |
| 510 | `n_shared_experts` chunks named `mlp.experts.{n_routed_experts + j}` so |
| 511 | the `RoutedExperts` loader treats them as extra experts. Yields the input |
| 512 | unchanged when the fusion is inactive, so callers can wrap unconditionally. |
| 513 | |
| 514 | Args: |
| 515 | weights: Iterable of `(name, tensor)` checkpoint pairs. |
| 516 | n_routed_experts: Number of routed experts; offsets the fused slots. |
| 517 | n_shared_experts: Number of shared experts packed into the tensor. |
| 518 | ckpt_prefix: Checkpoint module name of the shared experts. |
| 519 | enabled: Whether AITER fused-shared-experts is active. Defaults to |
| 520 | `rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()`; pass an |
| 521 | explicit value only when the model gates on something more (e.g. |
| 522 | quant-spec compatibility) and it must match its construction-time |
| 523 | decision. |
| 524 | |
| 525 | Yields: |
| 526 | `(name, tensor)` pairs with shared experts routed to fused slots. |
| 527 | |
| 528 | """ |
| 529 | if enabled is None: |
| 530 | from vllm._aiter_ops import rocm_aiter_ops |
| 531 | |
| 532 | enabled = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() |
| 533 | if not enabled: |
| 534 | yield from weights |
| 535 | return |
| 536 | |
| 537 | # Match on the dotted boundary so e.g. "mlp.shared_expert." does not also |
| 538 | # catch a sibling "mlp.shared_expert_gate". |
| 539 | prefix = f"{ckpt_prefix}." |
| 540 | for name, loaded_weight in weights: |
| 541 | if prefix not in name: |
| 542 | yield name, loaded_weight |
| 543 | continue |
| 544 | # gate/up split on the output dim; down_proj on the input dim. |
| 545 | split_dim = 1 if ("down_proj.weight" in name and loaded_weight.ndim > 1) else 0 |
| 546 | total = loaded_weight.shape[split_dim] |
| 547 | if total % n_shared_experts != 0: |
| 548 | raise ValueError( |
| 549 | f"FSE shared-expert weight {name!r} has size {total} along axis " |
| 550 | f"{split_dim}, not divisible by n_shared_experts={n_shared_experts}." |
| 551 | ) |
| 552 | chunk = total // n_shared_experts |
| 553 | for j in range(n_shared_experts): |
| 554 | sl = slice(j * chunk, (j + 1) * chunk) |
| 555 | if loaded_weight.ndim == 1: |
| 556 | chunk_weight = loaded_weight[sl] |
| 557 | elif split_dim == 0: |
| 558 | chunk_weight = loaded_weight[sl, :] |
| 559 | else: |
| 560 | chunk_weight = loaded_weight[:, sl] |
| 561 | yield ( |
| 562 | name.replace(prefix, f"mlp.experts.{n_routed_experts + j}."), |
| 563 | chunk_weight, |
| 564 | ) |
| 565 | |
| 566 | |
| 567 | def get_spec_layer_idx_from_weight_name( |
| 568 | config: "PreTrainedConfig", weight_name: str |
| 569 | ) -> int | None: |
| 570 | """Return the MTP layer index a weight belongs to, or None. |
| 571 | |
| 572 | Args: |
| 573 | config: The model config; must expose `num_hidden_layers` and, for MTP |
| 574 | checkpoints, `num_nextn_predict_layers`. |
| 575 | weight_name: Checkpoint weight name to classify. |
| 576 | |
| 577 | Returns: |
| 578 | The absolute layer index for an MTP-layer weight, else None. |
| 579 | |
| 580 | """ |
| 581 | if not (n := getattr(config, "num_nextn_predict_layers", 0)): |
| 582 | return None |
| 583 | base = config.num_hidden_layers |
| 584 | for i in range(n): |
| 585 | if weight_name.startswith( |
| 586 | ( |
| 587 | f"model.layers.{base + i}.", |
| 588 | f"layers.{base + i}.", |
| 589 | f"model.language_model.layers.{base + i}.", |
| 590 | ) |
| 591 | ): |
| 592 | return base + i |
| 593 | return None |
| 594 | |
| 595 | |
| 596 | def skip_spec_layers( |
| 597 | weights: Iterable[tuple[str, torch.Tensor]], config: "PreTrainedConfig" |
| 598 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 599 | """Drop MTP spec-layer weights (loaded by the MTP head, not the base model). |
| 600 | |
| 601 | Args: |
| 602 | weights: Iterable of `(name, tensor)` checkpoint pairs. |
| 603 | config: The model config, passed to `get_spec_layer_idx_from_weight_name`. |
| 604 | |
| 605 | Yields: |
| 606 | `(name, tensor)` pairs whose weight is not an MTP-layer weight. |
| 607 | |
| 608 | """ |
| 609 | return ( |
| 610 | (name, w) |
| 611 | for name, w in weights |
| 612 | if get_spec_layer_idx_from_weight_name(config, name) is None |
| 613 | ) |
| 614 | |
| 615 | |
| 616 | def init_vllm_registered_model( |
| 617 | vllm_config: VllmConfig, |
| 618 | *, |
| 619 | prefix: str = "", |
| 620 | hf_config: "PreTrainedConfig | None" = None, |
| 621 | architectures: list[str] | None = None, |
| 622 | ) -> nn.Module: |
| 623 | """Helper function to initialize an inner model registered to vLLM, |
| 624 | based on the arguments passed to the outer vLLM model. |
| 625 | """ |
| 626 | from vllm.model_executor.model_loader.utils import initialize_model |
| 627 | |
| 628 | if hf_config is None and architectures is not None: |
| 629 | # So that the architectures field is overridden |
| 630 | hf_config = vllm_config.model_config.hf_config |
| 631 | |
| 632 | if hf_config is not None: |
| 633 | vllm_config = vllm_config.with_hf_config(hf_config, architectures=architectures) |
| 634 | |
| 635 | return initialize_model(vllm_config=vllm_config, prefix=prefix) |
| 636 | |
| 637 | |
| 638 | @overload |
| 639 | def flatten_bn(x: list[torch.Tensor]) -> list[torch.Tensor]: ... |
| 640 | |
| 641 | |
| 642 | @overload |
| 643 | def flatten_bn(x: torch.Tensor) -> torch.Tensor: ... |
| 644 | |
| 645 | |
| 646 | @overload |
| 647 | def flatten_bn( |
| 648 | x: list[torch.Tensor] | torch.Tensor, |
| 649 | *, |
| 650 | concat: Literal[True], |
| 651 | ) -> torch.Tensor: ... |
| 652 | |
| 653 | |
| 654 | @overload |
| 655 | def flatten_bn( |
| 656 | x: list[torch.Tensor] | torch.Tensor, |
| 657 | *, |
| 658 | concat: bool = False, |
| 659 | ) -> list[torch.Tensor] | torch.Tensor: ... |
| 660 | |
| 661 | |
| 662 | def flatten_bn( |
| 663 | x: list[torch.Tensor] | torch.Tensor, |
| 664 | *, |
| 665 | concat: bool = False, |
| 666 | ) -> list[torch.Tensor] | torch.Tensor: |
| 667 | """Flatten the `B` and `N` dimensions of batched multimodal inputs. |
| 668 | |
| 669 | The input tensor should have shape `(B, N, ...)`. |
| 670 | """ |
| 671 | if isinstance(x, torch.Tensor): |
| 672 | return x.flatten(0, 1) |
| 673 | |
| 674 | if concat: |
| 675 | return torch.cat(x) |
| 676 | |
| 677 | return [x_n for x_b in x for x_n in x_b] |
| 678 | |
| 679 | |
| 680 | def _flatten_embeddings(embeddings: NestedTensors) -> torch.Tensor: |
| 681 | """Recursively flattens and concatenates NestedTensors on all but the last |
| 682 | dimension. |
| 683 | """ |
| 684 | if isinstance(embeddings, torch.Tensor): |
| 685 | # Flatten all but the last dimension. |
| 686 | return embeddings.flatten(0, -2) |
| 687 | |
| 688 | return torch.cat(tuple(_flatten_embeddings(t) for t in embeddings)) |
| 689 | |
| 690 | |
| 691 | def _embedding_count_expression(embeddings: NestedTensors) -> str: |
| 692 | """Constructs a debugging representation of the number of embeddings in the |
| 693 | NestedTensors. |
| 694 | """ |
| 695 | if isinstance(embeddings, torch.Tensor): |
| 696 | return " x ".join([str(dim) for dim in embeddings.shape[:-1]]) |
| 697 | |
| 698 | return " + ".join(_embedding_count_expression(inner) for inner in embeddings) |
| 699 | |
| 700 | |
| 701 | def split_list_into_ranges(lst: torch.Tensor, interval: int) -> list[list[int]]: |
| 702 | ranges: list[list[int]] = [[] for _ in range((max(lst) // interval) + 1)] |
| 703 | for num in lst: |
| 704 | index = num // interval |
| 705 | ranges[index].append(num) |
| 706 | return ranges |
| 707 | |
| 708 | |
| 709 | def _merge_multimodal_embeddings( |
| 710 | inputs_embeds: torch.Tensor, |
| 711 | multimodal_embeddings: NestedTensors, |
| 712 | is_multimodal: torch.Tensor, |
| 713 | ) -> torch.Tensor: |
| 714 | """Merge `multimodal_embeddings` into `inputs_embeds` by overwriting the |
| 715 | positions in `inputs_embeds` corresponding to placeholder tokens in |
| 716 | `input_ids`. |
| 717 | |
| 718 | Note: |
| 719 | This updates `inputs_embeds` in place. |
| 720 | |
| 721 | """ |
| 722 | if len(multimodal_embeddings) == 0: |
| 723 | return inputs_embeds |
| 724 | |
| 725 | mm_embeds_flat = _flatten_embeddings(multimodal_embeddings) |
| 726 | input_dtype = inputs_embeds.dtype |
| 727 | |
| 728 | try: |
| 729 | # If is_multimodal is on CPU this avoids a D2H sync |
| 730 | inputs_embeds[is_multimodal] = mm_embeds_flat.to(dtype=input_dtype) |
| 731 | except RuntimeError as e: |
| 732 | num_actual_tokens = len(mm_embeds_flat) |
| 733 | num_expected_tokens = is_multimodal.sum().item() |
| 734 | |
| 735 | if num_actual_tokens != num_expected_tokens: |
| 736 | expr = _embedding_count_expression(multimodal_embeddings) |
| 737 | |
| 738 | raise ValueError( |
| 739 | f"Attempted to assign {expr} = {num_actual_tokens} " |
| 740 | f"multimodal tokens to {num_expected_tokens} placeholders" |
| 741 | ) from e |
| 742 | |
| 743 | raise ValueError("Error during index put operation") from e |
| 744 | |
| 745 | return inputs_embeds |
| 746 | |
| 747 | |
| 748 | class StageMissingLayer(nn.Module): |
| 749 | def __init__(self, stage_name: str, module: nn.Module | None = None) -> None: |
| 750 | super().__init__() |
| 751 | |
| 752 | self.stage_name = stage_name |
| 753 | |
| 754 | # Don't register this as a child module in order to |
| 755 | # avoid missing keys when loading weights |
| 756 | self.__dict__["module"] = module |
| 757 | |
| 758 | def __getattr__(self, name: str): |
| 759 | return getattr(self.__dict__["module"], name) |
| 760 | |
| 761 | def __call__(self, *args, **kwargs): |
| 762 | raise RuntimeError(f"{self} should not be called") |
| 763 | |
| 764 | def extra_repr(self) -> str: |
| 765 | return f"stage_name={self.stage_name!r}" |
| 766 | |
| 767 | |
| 768 | @contextmanager |
| 769 | def collect_children( |
| 770 | module: nn.Module, |
| 771 | *, |
| 772 | targets: type[nn.Module] | tuple[type[nn.Module], ...] | None = None, |
| 773 | ): |
| 774 | """Within this context, collect all direct child assignments to `module`, |
| 775 | returning a list of children names that is internally updated until the |
| 776 | context is exited. |
| 777 | |
| 778 | If `targets` is set, instead collect descendents of `module` |
| 779 | that are an instance of `targets`, even if they aren't direct children. |
| 780 | """ |
| 781 | children_names = list[str]() |
| 782 | |
| 783 | if targets is None: |
| 784 | |
| 785 | def hook(module_: nn.Module, name: str, submodule: nn.Module): |
| 786 | if module_ is module: |
| 787 | children_names.append(name) |
| 788 | |
| 789 | with register_module_module_registration_hook(hook): |
| 790 | yield children_names |
| 791 | else: |
| 792 | yield children_names |
| 793 | |
| 794 | for name, module_ in module.named_modules(): |
| 795 | if isinstance(module_, targets): |
| 796 | children_names.append(name) |
| 797 | |
| 798 | |
| 799 | @contextmanager |
| 800 | def no_init_weights( |
| 801 | module: nn.Module, |
| 802 | placeholder: Callable[[nn.Module], nn.Module], |
| 803 | *, |
| 804 | targets: type[nn.Module] | tuple[type[nn.Module], ...] | None = None, |
| 805 | ): |
| 806 | """Within this context, prevent weight initialization from using device memory and |
| 807 | replace direct child assignments to `module` with the result of `placeholder()`. |
| 808 | |
| 809 | If `targets` is set, instead prevent weight initialization and |
| 810 | replace assignments where the child is an instance of `targets`, |
| 811 | even if they aren't direct children of `module`. |
| 812 | """ |
| 813 | if targets is None: |
| 814 | |
| 815 | def hook(module_: nn.Module, name: str, submodule: nn.Module): |
| 816 | if module_ is module: |
| 817 | return placeholder(submodule) |
| 818 | |
| 819 | return submodule |
| 820 | |
| 821 | with register_module_module_registration_hook(hook), torch.device("meta"): |
| 822 | yield |
| 823 | else: |
| 824 | |
| 825 | def hook(module_: nn.Module, name: str, submodule: nn.Module): |
| 826 | if isinstance(module_, targets): |
| 827 | submodule.to("meta") # Free memory |
| 828 | if isinstance(submodule, targets): |
| 829 | submodule.to("meta") # Free memory |
| 830 | return placeholder(submodule) |
| 831 | |
| 832 | return submodule |
| 833 | |
| 834 | # Not all descendents are targeted, so we can't use a blanket |
| 835 | # `torch.device("meta")` context |
| 836 | with register_module_module_registration_hook(hook): |
| 837 | yield |
| 838 | |
| 839 | |
| 840 | class LayerFn(Protocol): |
| 841 | def __call__(self, prefix: str) -> torch.nn.Module: ... |
| 842 | |
| 843 | |
| 844 | class PPMissingLayer(torch.nn.Identity): |
| 845 | """A placeholder layer for missing layers in a pipeline parallel model.""" |
| 846 | |
| 847 | def __init__(self, *args, **kwargs): |
| 848 | super().__init__() |
| 849 | |
| 850 | def forward(self, *args, **kwargs): |
| 851 | """Return the first arg from args or the first value from kwargs.""" |
| 852 | return args[0] if args else next(iter(kwargs.values())) |
| 853 | |
| 854 | |
| 855 | def spec_decode_needs_target_embed( |
| 856 | vllm_config: VllmConfig, *, include_mtp: bool = False |
| 857 | ) -> bool: |
| 858 | """Whether the last PP rank needs the target input embedding. |
| 859 | |
| 860 | MTP targets opt in when their drafter supports sharing this embedding. |
| 861 | """ |
| 862 | from vllm.distributed.parallel_state import get_pp_group |
| 863 | |
| 864 | speculative_config = vllm_config.speculative_config |
| 865 | if speculative_config is None: |
| 866 | return False |
| 867 | if speculative_config.method not in ( |
| 868 | "eagle", |
| 869 | "eagle3", |
| 870 | "dflash", |
| 871 | "dspark", |
| 872 | ) and not (include_mtp and speculative_config.method == "mtp"): |
| 873 | return False |
| 874 | pp = get_pp_group() |
| 875 | return pp.world_size > 1 and pp.is_last_rank |
| 876 | |
| 877 | |
| 878 | def make_layers( |
| 879 | num_hidden_layers: int, |
| 880 | layer_fn: LayerFn, |
| 881 | prefix: str, |
| 882 | ) -> tuple[int, int, torch.nn.ModuleList]: |
| 883 | """Make a list of layers with the given layer function, taking |
| 884 | pipeline parallelism into account. |
| 885 | |
| 886 | Args: |
| 887 | num_hidden_layers: Total number of hidden layers in the model. |
| 888 | layer_fn: Function to create a layer given its index. |
| 889 | prefix: Prefix for layer names. |
| 890 | |
| 891 | Returns: |
| 892 | Tuple of (start_layer, end_layer, modules). |
| 893 | |
| 894 | """ |
| 895 | from vllm.distributed.parallel_state import get_pp_group |
| 896 | from vllm.distributed.utils import get_pp_indices |
| 897 | from vllm.model_executor.offloader import get_offloader |
| 898 | |
| 899 | start_layer, end_layer = get_pp_indices( |
| 900 | num_hidden_layers, get_pp_group().rank_in_group, get_pp_group().world_size |
| 901 | ) |
| 902 | |
| 903 | modules = torch.nn.ModuleList( |
| 904 | [PPMissingLayer() for _ in range(start_layer)] |
| 905 | + get_offloader().wrap_modules( |
| 906 | layer_fn(prefix=f"{prefix}.{idx}") for idx in range(start_layer, end_layer) |
| 907 | ) |
| 908 | + [PPMissingLayer() for _ in range(end_layer, num_hidden_layers)] |
| 909 | ) |
| 910 | |
| 911 | return start_layer, end_layer, modules |
| 912 | |
| 913 | |
| 914 | # NOTE: don't use lru_cache here because it can prevent garbage collection |
| 915 | _model_to_pp_missing_layer_names: dict[int, list[str]] = {} |
| 916 | |
| 917 | |
| 918 | def get_pp_missing_layer_names(model: torch.nn.Module) -> list[str]: |
| 919 | """Get the names of the missing layers in a pipeline parallel model.""" |
| 920 | model_id = id(model) |
| 921 | if model_id in _model_to_pp_missing_layer_names: |
| 922 | return _model_to_pp_missing_layer_names[model_id] |
| 923 | |
| 924 | missing_layer_names = [] |
| 925 | for name, module in model.named_modules(): |
| 926 | if isinstance(module, (StageMissingLayer, PPMissingLayer)): |
| 927 | # NOTE: the trailing dot is used to match the prefix of the layer. |
| 928 | # without the dot, we could match a layer that is not missing, |
| 929 | # e.g., 'encoder.layer.1' would match 'encoder.layer.11' |
| 930 | missing_layer_names.append(name + ".") |
| 931 | _model_to_pp_missing_layer_names[model_id] = missing_layer_names |
| 932 | |
| 933 | return missing_layer_names |
| 934 | |
| 935 | |
| 936 | def is_pp_missing_parameter(name: str, model: torch.nn.Module) -> bool: |
| 937 | """Check if a parameter is missing in a pipeline parallel model.""" |
| 938 | if isinstance(model, (StageMissingLayer, PPMissingLayer)): |
| 939 | return True |
| 940 | |
| 941 | return any( |
| 942 | name.startswith(missing_layer_name) |
| 943 | for missing_layer_name in get_pp_missing_layer_names(model) |
| 944 | ) |
| 945 | |
| 946 | |
| 947 | def make_empty_intermediate_tensors_factory(keys: list[str], hidden_size: int): |
| 948 | def make_empty_intermediate_tensors( |
| 949 | batch_size: int, |
| 950 | dtype: torch.dtype, |
| 951 | device: torch.device, |
| 952 | ) -> IntermediateTensors: |
| 953 | return IntermediateTensors( |
| 954 | { |
| 955 | key: torch.zeros((batch_size, hidden_size), dtype=dtype, device=device) |
| 956 | for key in keys |
| 957 | } |
| 958 | ) |
| 959 | |
| 960 | return make_empty_intermediate_tensors |
| 961 | |
| 962 | |
| 963 | def maybe_prefix(prefix: str, name: str) -> str: |
| 964 | """Add a prefix to a name if the prefix is non-empty. |
| 965 | |
| 966 | Args: |
| 967 | prefix: The prefix to add. If empty, no prefix will be added. |
| 968 | name: The name to potentially prefix. |
| 969 | |
| 970 | Returns: |
| 971 | The string "prefix.name" if prefix was non-empty, otherwise just "name". |
| 972 | |
| 973 | """ |
| 974 | return name if not prefix else f"{prefix}.{name}" |
| 975 | |
| 976 | |
| 977 | def get_draft_quant_config(vllm_config: VllmConfig) -> "QuantizationConfig | None": |
| 978 | """Get quantization config for Draft models. |
| 979 | |
| 980 | Draft models should use their own quantization config instead of the verifier/target |
| 981 | model's config. This helper retrieves the draft model's quantization config. |
| 982 | |
| 983 | Args: |
| 984 | vllm_config: The vLLM configuration object. |
| 985 | |
| 986 | Returns: |
| 987 | The draft model's config if available, None otherwise. |
| 988 | |
| 989 | """ |
| 990 | speculative_config = vllm_config.speculative_config |
| 991 | assert speculative_config is not None |
| 992 | draft_model_config = speculative_config.draft_model_config |
| 993 | draft_load_config = vllm_config.load_config |
| 994 | |
| 995 | return ( |
| 996 | VllmConfig.get_quantization_config(draft_model_config, draft_load_config) |
| 997 | if draft_model_config |
| 998 | else None |
| 999 | ) |
| 1000 | |
| 1001 | |
| 1002 | def extract_layer_index(layer_name: str, num_attn_module: int = 1) -> int: |
| 1003 | """Extract the layer index from the module name. |
| 1004 | |
| 1005 | Examples: |
| 1006 | - "encoder.layers.0" -> 0 |
| 1007 | - "encoder.layers.1.self_attn" -> 1 |
| 1008 | - "2.self_attn" -> 2 |
| 1009 | - "model.encoder.layers.0.sub.1" -> ValueError if num_attn_module == 1 |
| 1010 | |
| 1011 | """ |
| 1012 | subnames = layer_name.split(".") |
| 1013 | int_vals: list[int] = [] |
| 1014 | for subname in subnames: |
| 1015 | try: |
| 1016 | int_vals.append(int(subname)) |
| 1017 | except ValueError: |
| 1018 | continue |
| 1019 | if num_attn_module == 1 or "attn" not in layer_name: |
| 1020 | assert len(int_vals) == 1, ( |
| 1021 | f"layer name {layer_name} should only contain one integer" |
| 1022 | ) |
| 1023 | |
| 1024 | return int_vals[0] |
| 1025 | else: |
| 1026 | assert len(int_vals) <= 2, ( |
| 1027 | f"layer name {layer_name} should contain most two integers" |
| 1028 | ) |
| 1029 | layer_index = ( |
| 1030 | int_vals[0] * num_attn_module + int_vals[1] |
| 1031 | if len(int_vals) == 2 |
| 1032 | else int_vals[0] |
| 1033 | ) |
| 1034 | return layer_index |
| 1035 | |
| 1036 | |
| 1037 | def cast_overflow_tensors(tensors: torch.Tensor, offset: float = 1000) -> torch.Tensor: |
| 1038 | clamp_value = torch.finfo(tensors.dtype).max - offset |
| 1039 | return torch.clamp(tensors, min=-clamp_value, max=clamp_value) |
| 1040 | |
| 1041 | |
| 1042 | def fast_topk( |
| 1043 | values: torch.Tensor, topk: int, dim: int |
| 1044 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 1045 | """Optimized topk implementation that uses torch.max for k=1 case. |
| 1046 | |
| 1047 | This function provides better performance for the common case of k=1 |
| 1048 | by using torch.max instead of the more general torch.topk. |
| 1049 | |
| 1050 | Args: |
| 1051 | values: Input tensor to find top-k values from |
| 1052 | topk: Number of top values to return (k). Must be > 0. |
| 1053 | dim: Dimension along which to compute topk |
| 1054 | |
| 1055 | Returns: |
| 1056 | Tuple of (values, indices) where values are the top-k values |
| 1057 | and indices are their corresponding indices in the input tensor |
| 1058 | |
| 1059 | """ |
| 1060 | if topk == 1: |
| 1061 | # Use max along the specified dimension to get both value and index |
| 1062 | return torch.max(values, dim=dim, keepdim=True) |
| 1063 | else: |
| 1064 | # Use topk for efficiency with larger k values |
| 1065 | return torch.topk(values, topk, dim=dim) |
| 1066 | |
| 1067 | |
| 1068 | # Chunk x along the num_tokens axis for sequence parallelism |
| 1069 | # NOTE: This is wrapped in a torch custom op to work around the following issue: |
| 1070 | # The output tensor can have a sequence length 0 at small input sequence lengths |
| 1071 | # even though we explicitly pad to avoid this. |
| 1072 | def sequence_parallel_chunk(x: torch.Tensor) -> torch.Tensor: |
| 1073 | return torch.ops.vllm.sequence_parallel_chunk_impl(x) |
| 1074 | |
| 1075 | |
| 1076 | def sequence_parallel_chunk_impl(x: torch.Tensor) -> torch.Tensor: |
| 1077 | tp_size = get_tensor_model_parallel_world_size() |
| 1078 | tp_rank = get_tensor_model_parallel_rank() |
| 1079 | |
| 1080 | # all_gather needs the sequence length to be divisible by tp_size |
| 1081 | seq_len = x.size(0) |
| 1082 | remainder = seq_len % tp_size |
| 1083 | if remainder != 0: |
| 1084 | pad_len = tp_size - remainder |
| 1085 | y = nn.functional.pad(x, (0, 0, 0, pad_len)) |
| 1086 | else: |
| 1087 | y = x |
| 1088 | |
| 1089 | chunk = y.shape[0] // tp_size |
| 1090 | start = tp_rank * chunk |
| 1091 | out = torch.narrow(y, 0, start, chunk) |
| 1092 | # narrow() returns a view; clone when it aliases the input (no-pad case), |
| 1093 | # since a functional custom op must not return a view of an input. |
| 1094 | return out.clone() if y is x else out |
| 1095 | |
| 1096 | |
| 1097 | def sequence_parallel_chunk_impl_fake(x: torch.Tensor) -> torch.Tensor: |
| 1098 | tp_size = get_tensor_model_parallel_world_size() |
| 1099 | seq_len = cdiv(x.size(0), tp_size) |
| 1100 | shape = list(x.shape) |
| 1101 | shape[0] = seq_len |
| 1102 | out = torch.empty(shape, dtype=x.dtype, device=x.device) |
| 1103 | return out |
| 1104 | |
| 1105 | |
| 1106 | direct_register_custom_op( |
| 1107 | op_name="sequence_parallel_chunk_impl", |
| 1108 | op_func=sequence_parallel_chunk_impl, |
| 1109 | fake_impl=sequence_parallel_chunk_impl_fake, |
| 1110 | tags=(torch.Tag.needs_fixed_stride_order,), |
| 1111 | ) |
| 1112 | |
| 1113 | |
| 1114 | def process_eagle_weight( |
| 1115 | model: nn.Module, |
| 1116 | name: str, |
| 1117 | ) -> None: |
| 1118 | """Update EAGLE model flags based on loaded weight name. |
| 1119 | This should be called during weight loading to detect if a model |
| 1120 | has its own lm_head or embed_tokens weight. |
| 1121 | |
| 1122 | Args: |
| 1123 | model: The model instance (must support EAGLE) |
| 1124 | name: The name of the weight to process |
| 1125 | |
| 1126 | """ |
| 1127 | if not supports_any_eagle(model): |
| 1128 | return |
| 1129 | |
| 1130 | # To prevent overriding with target model's layers |
| 1131 | if "lm_head" in name: |
| 1132 | model.has_own_lm_head = True |
| 1133 | if "embed_tokens" in name: |
| 1134 | model.has_own_embed_tokens = True |
| 1135 | |
| 1136 | |
| 1137 | def get_layer_index(feature_layer_index: int, num_hidden_layers: int) -> int: |
| 1138 | """Given a signed vision feature layer, get the number of hidden layers |
| 1139 | needed to leverage it. |
| 1140 | |
| 1141 | Args: |
| 1142 | feature_layer_index: Index of a required layer in the visual encoder. |
| 1143 | num_hidden_layers: The total number of hidden layers in the visual encoder. |
| 1144 | |
| 1145 | """ |
| 1146 | if feature_layer_index < 0: |
| 1147 | return num_hidden_layers + feature_layer_index + 1 |
| 1148 | return feature_layer_index |
| 1149 | |
| 1150 | |
| 1151 | def scatter_output_slices( |
| 1152 | output: torch.Tensor, |
| 1153 | indices: list[int], |
| 1154 | per_item_out_tokens: list[int], |
| 1155 | dest: dict[int, torch.Tensor] | list[torch.Tensor | None], |
| 1156 | clone: bool = False, |
| 1157 | ) -> None: |
| 1158 | """Slice a concatenated output tensor and scatter into dest by index.""" |
| 1159 | offset = 0 |
| 1160 | for idx in indices: |
| 1161 | n_tok = per_item_out_tokens[idx] |
| 1162 | sliced = output[offset : offset + n_tok] |
| 1163 | dest[idx] = sliced.clone() if clone else sliced |
| 1164 | offset += n_tok |
| 1165 | |
| 1166 | |
| 1167 | def parse_diarized_timestamp(marker: str) -> float | None: |
| 1168 | if ( |
| 1169 | not marker |
| 1170 | or not marker.isascii() |
| 1171 | or marker.count(".") > 1 |
| 1172 | or not marker.replace(".", "").isdigit() |
| 1173 | ): |
| 1174 | return None |
| 1175 | return float(marker) |
| 1176 | |
| 1177 | |
| 1178 | def parse_diarized_speaker(speaker: str) -> str | None: |
| 1179 | if ( |
| 1180 | len(speaker) < 2 |
| 1181 | or speaker[0] != "S" |
| 1182 | or not speaker[1:].isascii() |
| 1183 | or not speaker[1:].isdigit() |
| 1184 | ): |
| 1185 | return None |
| 1186 | return speaker |