| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Utilities for downloading and initializing model weights.""" |
| 4 | |
| 5 | import asyncio |
| 6 | import concurrent.futures |
| 7 | import fnmatch |
| 8 | import glob |
| 9 | import hashlib |
| 10 | import json |
| 11 | import os |
| 12 | import tempfile |
| 13 | import threading |
| 14 | import time |
| 15 | from collections.abc import Callable, Generator, Iterable |
| 16 | from contextlib import contextmanager, suppress |
| 17 | from pathlib import Path |
| 18 | from typing import IO, TYPE_CHECKING, Any |
| 19 | |
| 20 | import filelock |
| 21 | import huggingface_hub.constants |
| 22 | import numpy as np |
| 23 | import regex as re |
| 24 | import torch |
| 25 | from safetensors.torch import load, load_file, safe_open |
| 26 | from tqdm.auto import tqdm |
| 27 | from transformers.utils import SAFE_WEIGHTS_INDEX_NAME |
| 28 | |
| 29 | from vllm import envs |
| 30 | from vllm.config import ModelConfig |
| 31 | from vllm.config.load import ( |
| 32 | DEFAULT_SAFETENSORS_PREFETCH_BLOCK_SIZE, |
| 33 | DEFAULT_SAFETENSORS_PREFETCH_NUM_THREADS, |
| 34 | LoadConfig, |
| 35 | ) |
| 36 | from vllm.distributed import get_tensor_model_parallel_rank, get_world_group |
| 37 | from vllm.logger import init_logger |
| 38 | from vllm.model_executor.layers.quantization import ( |
| 39 | QuantizationConfig, |
| 40 | get_quantization_config, |
| 41 | ) |
| 42 | from vllm.model_executor.model_loader.ep_weight_filter import ( |
| 43 | should_skip_weight, |
| 44 | ) |
| 45 | from vllm.platforms import current_platform |
| 46 | from vllm.tracing import instrument |
| 47 | from vllm.transformers_utils.repo_utils import hf_api, hf_fs |
| 48 | from vllm.utils.import_utils import PlaceholderModule |
| 49 | |
| 50 | try: |
| 51 | from runai_model_streamer import SafetensorsStreamer |
| 52 | except ImportError: |
| 53 | runai_model_streamer = PlaceholderModule("runai_model_streamer") # type: ignore[assignment] |
| 54 | SafetensorsStreamer = runai_model_streamer.placeholder_attr("SafetensorsStreamer") |
| 55 | |
| 56 | try: |
| 57 | from fastsafetensors import SingleGroup |
| 58 | except ImportError: |
| 59 | fastsafetensors = PlaceholderModule("fastsafetensors") |
| 60 | SingleGroup = fastsafetensors.placeholder_attr("SingleGroup") |
| 61 | |
| 62 | from vllm.model_executor.layers.quantization.torchao import torchao_version_at_least |
| 63 | |
| 64 | if TYPE_CHECKING: |
| 65 | from vllm.model_executor.models.utils import WeightsMapper |
| 66 | |
| 67 | logger = init_logger(__name__) |
| 68 | |
| 69 | # use system-level temp directory for file locks, so that multiple users |
| 70 | # can share the same lock without error. |
| 71 | # lock files in the temp directory will be automatically deleted when the |
| 72 | # system reboots, so users will not complain about annoying lock files |
| 73 | temp_dir = tempfile.gettempdir() |
| 74 | |
| 75 | |
| 76 | def enable_xet_high_performance(): |
| 77 | """Automatically activates xet high performance mode""" |
| 78 | if "HF_XET_HIGH_PERFORMANCE" not in os.environ: |
| 79 | huggingface_hub.constants.HF_XET_HIGH_PERFORMANCE = True |
| 80 | |
| 81 | |
| 82 | enable_xet_high_performance() |
| 83 | |
| 84 | |
| 85 | class DisabledTqdm(tqdm): |
| 86 | def __init__(self, *args, **kwargs): |
| 87 | kwargs["disable"] = True |
| 88 | super().__init__(*args, **kwargs) |
| 89 | |
| 90 | |
| 91 | def get_lock(model_name_or_path: str | Path, cache_dir: str | None = None): |
| 92 | lock_dir = cache_dir or temp_dir |
| 93 | model_name_or_path = str(model_name_or_path) |
| 94 | os.makedirs(os.path.dirname(lock_dir), exist_ok=True) |
| 95 | model_name = model_name_or_path.replace("/", "-") |
| 96 | hash_name = hashlib.sha256(model_name.encode()).hexdigest() |
| 97 | # add hash to avoid conflict with old users' lock files |
| 98 | lock_file_name = hash_name + model_name + ".lock" |
| 99 | # mode 0o666 is required for the filelock to be shared across users |
| 100 | lock = filelock.FileLock(os.path.join(lock_dir, lock_file_name), mode=0o666) |
| 101 | return lock |
| 102 | |
| 103 | |
| 104 | @contextmanager |
| 105 | def atomic_writer( |
| 106 | filepath: str | Path, mode: str = "w", encoding: str | None = None |
| 107 | ) -> Generator[IO]: |
| 108 | """Context manager that provides an atomic file writing routine. |
| 109 | |
| 110 | The context manager writes to a temporary file and, if successful, |
| 111 | atomically replaces the original file. |
| 112 | |
| 113 | Args: |
| 114 | filepath (str or Path): The path to the file to write. |
| 115 | mode (str): The file mode for the temporary file (e.g., 'w', 'wb'). |
| 116 | encoding (str): The encoding for text mode. |
| 117 | |
| 118 | Yields: |
| 119 | file object: A handle to the temporary file. |
| 120 | |
| 121 | """ |
| 122 | # Create a temporary file in the same directory as the target file |
| 123 | # to ensure it's on the same filesystem for an atomic replace. |
| 124 | temp_dir = os.path.dirname(filepath) |
| 125 | temp_fd, temp_path = tempfile.mkstemp(dir=temp_dir) |
| 126 | |
| 127 | try: |
| 128 | # Open the temporary file for writing |
| 129 | with os.fdopen(temp_fd, mode=mode, encoding=encoding) as temp_file: |
| 130 | yield temp_file |
| 131 | |
| 132 | # If the 'with' block completes successfully, |
| 133 | # perform the atomic replace. |
| 134 | os.replace(temp_path, filepath) |
| 135 | |
| 136 | except Exception: |
| 137 | logger.exception( |
| 138 | "Error during atomic write. Original file '%s' not modified", filepath |
| 139 | ) |
| 140 | raise |
| 141 | finally: |
| 142 | # Clean up the temporary file if it still exists. |
| 143 | if os.path.exists(temp_path): |
| 144 | os.remove(temp_path) |
| 145 | |
| 146 | |
| 147 | def _natural_sort_key(filepath: str) -> list: |
| 148 | """Natural sort key for filenames with numeric components, such as |
| 149 | model-00001-of-00005.safetensors -> ['model-', 1, '-of-', 5, '.safetensors']""" |
| 150 | return [ |
| 151 | int(s) if s.isdigit() else s |
| 152 | for s in re.split(r"(\d+)", os.path.basename(filepath)) |
| 153 | ] |
| 154 | |
| 155 | |
| 156 | def maybe_download_from_modelscope( |
| 157 | model: str, |
| 158 | revision: str | None = None, |
| 159 | download_dir: str | None = None, |
| 160 | ignore_patterns: str | list[str] | None = None, |
| 161 | allow_patterns: list[str] | str | None = None, |
| 162 | ) -> str | None: |
| 163 | """Download model from ModelScope hub if VLLM_USE_MODELSCOPE is True. |
| 164 | |
| 165 | Returns the path to the downloaded model, or None if the model is not |
| 166 | downloaded from ModelScope.""" |
| 167 | if envs.VLLM_USE_MODELSCOPE: |
| 168 | # download model from ModelScope hub, |
| 169 | # lazy import so that modelscope is not required for normal use. |
| 170 | # pylint: disable=C. |
| 171 | from modelscope.hub.snapshot_download import snapshot_download |
| 172 | |
| 173 | # Use file lock to prevent multiple processes from |
| 174 | # downloading the same model weights at the same time. |
| 175 | with get_lock(model, download_dir): |
| 176 | if not os.path.exists(model): |
| 177 | model_path = snapshot_download( |
| 178 | model_id=model, |
| 179 | cache_dir=download_dir, |
| 180 | local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, |
| 181 | revision=revision, |
| 182 | ignore_file_pattern=ignore_patterns, |
| 183 | allow_patterns=allow_patterns, |
| 184 | ) |
| 185 | else: |
| 186 | model_path = model |
| 187 | return model_path |
| 188 | return None |
| 189 | |
| 190 | |
| 191 | # TODO(woosuk): Move this to other place. |
| 192 | def get_quant_config( |
| 193 | model_config: ModelConfig, load_config: LoadConfig |
| 194 | ) -> QuantizationConfig: |
| 195 | if model_config.quantization is None: |
| 196 | raise ValueError("Model quantization method is not specified in the config.") |
| 197 | quant_cls = get_quantization_config(model_config.quantization) |
| 198 | from vllm.config.quantization import ( |
| 199 | _ONLINE_SHORTHANDS, |
| 200 | QuantizationConfigArgs, |
| 201 | resolve_quantization_config, |
| 202 | ) |
| 203 | from vllm.model_executor.layers.quantization.online.base import ( |
| 204 | OnlineQuantizationConfig, |
| 205 | ) |
| 206 | |
| 207 | # resolve_quantization_config does not resolve `quantization_config` from |
| 208 | # ambiguous shorthands as `"mxfp4"`, `"mxfp8"`. |
| 209 | online_args = model_config.quantization_config |
| 210 | |
| 211 | def maybe_compose_online_quantization( |
| 212 | checkpoint_config: QuantizationConfig, |
| 213 | ) -> QuantizationConfig: |
| 214 | if online_args is None: |
| 215 | return checkpoint_config |
| 216 | assert isinstance(online_args, QuantizationConfigArgs) |
| 217 | |
| 218 | checkpoint_config.online_quantization_config = OnlineQuantizationConfig( |
| 219 | online_args |
| 220 | ) |
| 221 | return checkpoint_config |
| 222 | |
| 223 | # Read the quantization config from the HF model config, if available. |
| 224 | hf_quant_config = getattr(model_config.hf_config, "quantization_config", None) |
| 225 | # some vision model may keep quantization_config in their text_config |
| 226 | hf_text_config = getattr(model_config.hf_config, "text_config", None) |
| 227 | if hf_quant_config is None and hf_text_config is not None: |
| 228 | hf_quant_config = getattr(hf_text_config, "quantization_config", None) |
| 229 | if hf_quant_config is None: |
| 230 | # compressed-tensors uses a compressions_config |
| 231 | hf_quant_config = getattr(model_config.hf_config, "compression_config", None) |
| 232 | |
| 233 | # Pipe information about heads to enable TP-aware loading of attn_head scales |
| 234 | if ( |
| 235 | hf_quant_config is not None |
| 236 | and hf_quant_config.get("quant_method") == "compressed-tensors" |
| 237 | and "config_groups" in hf_quant_config |
| 238 | ): |
| 239 | n_heads = model_config.model_arch_config.total_num_attention_heads |
| 240 | n_kv_heads = model_config.model_arch_config.total_num_kv_heads |
| 241 | |
| 242 | hf_quant_config["total_num_heads"] = n_heads |
| 243 | hf_quant_config["total_num_kv_heads"] = ( |
| 244 | n_kv_heads if n_kv_heads is not None else n_heads |
| 245 | ) |
| 246 | |
| 247 | if hf_quant_config is not None: |
| 248 | # `model_config.quantization_config` may be set alongside a checkpoint |
| 249 | # quant config: the checkpoint determines `quant_cls`, and the user's |
| 250 | # QuantizationConfigArgs is consulted by individual quant methods |
| 251 | # (e.g. for activation overrides via the MXFP4 oracle). |
| 252 | |
| 253 | # For modelopt_mixed, config.json's quantization_config may or may |
| 254 | # not contain the per-layer quantized_layers map. Newer checkpoints |
| 255 | # embed it directly; older ones keep it only in hf_quant_config.json. |
| 256 | # If it is missing, fall through to the file-based loading path. |
| 257 | if ( |
| 258 | model_config.quantization == "modelopt_mixed" |
| 259 | and "quantized_layers" not in hf_quant_config |
| 260 | ): |
| 261 | pass # fall through to file-based loading below |
| 262 | else: |
| 263 | return maybe_compose_online_quantization( |
| 264 | quant_cls.from_config(hf_quant_config) |
| 265 | ) |
| 266 | |
| 267 | # if hf_quant_config is None, we will try to get config from |
| 268 | # hf_overrides |
| 269 | hf_overrides = model_config.hf_overrides |
| 270 | if callable(hf_overrides): |
| 271 | # A callable hf_overrides is a config-to-config transform (e.g. the |
| 272 | # one SpeculativeConfig installs on draft model configs); it cannot |
| 273 | # carry quantization config entries, so treat it as no overrides. |
| 274 | hf_overrides = {} |
| 275 | if not isinstance(hf_overrides, dict): |
| 276 | raise ValueError( |
| 277 | "hf_overrides must be a dict for get_quant_config " |
| 278 | "to get the quantization config from it." |
| 279 | ) |
| 280 | quantization_config_file = hf_overrides.get("quantization_config_file") |
| 281 | if quantization_config_file is not None: |
| 282 | if hasattr(quant_cls, "from_config_file"): |
| 283 | return maybe_compose_online_quantization( |
| 284 | quant_cls.from_config_file(quantization_config_file) |
| 285 | ) |
| 286 | else: |
| 287 | raise NotImplementedError( |
| 288 | "from_config_file is specified in hf_override config, " |
| 289 | "but quant_cls.from_config_file is not implemented in " |
| 290 | f"{quant_cls}" |
| 291 | ) |
| 292 | quantization_config_json = hf_overrides.get("quantization_config_dict_json") |
| 293 | if quantization_config_json is not None: |
| 294 | if hasattr(quant_cls, "from_config_dict_json"): |
| 295 | return maybe_compose_online_quantization( |
| 296 | quant_cls.from_config_dict_json(quantization_config_json) |
| 297 | ) |
| 298 | else: |
| 299 | raise NotImplementedError( |
| 300 | "from_config_dict_json is specified in hf_override config, " |
| 301 | "but quant_cls.from_config_dict_json is not implemented in " |
| 302 | f"{quant_cls}" |
| 303 | ) |
| 304 | |
| 305 | # Raw unambiguous online quantization doesn't read from checkpoint configs |
| 306 | # We must continue to load/read the config below in two cases: |
| 307 | # 1. composed online quantization, before `maybe_compose_online_quantization`, |
| 308 | # 2. Ambiguous online shorthands as `"mxpf4"`, `"mxfp8"`, for which `online_args` |
| 309 | # is not set yet. |
| 310 | if quant_cls is OnlineQuantizationConfig and online_args is not None: |
| 311 | assert isinstance(online_args, QuantizationConfigArgs) |
| 312 | return OnlineQuantizationConfig(args=online_args) |
| 313 | |
| 314 | model_name_or_path = ( |
| 315 | maybe_download_from_modelscope( |
| 316 | model_config.model, |
| 317 | revision=model_config.revision, |
| 318 | download_dir=load_config.download_dir, |
| 319 | allow_patterns=["*.json"], |
| 320 | ) |
| 321 | or model_config.model |
| 322 | ) |
| 323 | is_local = os.path.isdir(model_name_or_path) |
| 324 | if not is_local: |
| 325 | # Download the config files. |
| 326 | with get_lock(model_config.model, load_config.download_dir): |
| 327 | hf_folder = hf_api().snapshot_download( |
| 328 | model_config.model, |
| 329 | revision=model_config.revision, |
| 330 | allow_patterns="*.json", |
| 331 | cache_dir=load_config.download_dir, |
| 332 | local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, |
| 333 | tqdm_class=DisabledTqdm, |
| 334 | ) |
| 335 | else: |
| 336 | hf_folder = model_name_or_path |
| 337 | |
| 338 | possible_config_filenames = quant_cls.get_config_filenames() |
| 339 | |
| 340 | # If the quantization config is not found, use the default config. |
| 341 | if not possible_config_filenames: |
| 342 | if model_config.quantization == "fp8": |
| 343 | logger.warning_once( |
| 344 | "--quantization fp8 is deprecated for online quantization; " |
| 345 | "use --quantization fp8_per_tensor instead." |
| 346 | ) |
| 347 | fp8_args = resolve_quantization_config("fp8_per_tensor", online_args) |
| 348 | assert fp8_args is not None |
| 349 | return OnlineQuantizationConfig(args=fp8_args) |
| 350 | if model_config.quantization in _ONLINE_SHORTHANDS: |
| 351 | args = online_args or _ONLINE_SHORTHANDS[model_config.quantization] |
| 352 | assert isinstance(args, QuantizationConfigArgs) |
| 353 | return OnlineQuantizationConfig(args=args) |
| 354 | return maybe_compose_online_quantization(quant_cls()) |
| 355 | |
| 356 | config_files = glob.glob(os.path.join(hf_folder, "*.json")) |
| 357 | |
| 358 | quant_config_files = [ |
| 359 | f for f in config_files if any(f.endswith(x) for x in possible_config_filenames) |
| 360 | ] |
| 361 | if len(quant_config_files) == 0: |
| 362 | if model_config.quantization in _ONLINE_SHORTHANDS: |
| 363 | args = online_args or _ONLINE_SHORTHANDS[model_config.quantization] |
| 364 | assert isinstance(args, QuantizationConfigArgs) |
| 365 | return OnlineQuantizationConfig(args=args) |
| 366 | raise ValueError(f"Cannot find the config file for {model_config.quantization}") |
| 367 | if len(quant_config_files) > 1: |
| 368 | raise ValueError( |
| 369 | f"Found multiple config files for {model_config.quantization}: " |
| 370 | f"{quant_config_files}" |
| 371 | ) |
| 372 | |
| 373 | quant_config_file = quant_config_files[0] |
| 374 | with open(quant_config_file) as f: |
| 375 | config = json.load(f) |
| 376 | |
| 377 | if model_config.quantization in ("modelopt", "modelopt_mixed"): |
| 378 | if config.get("producer", {}).get("name") == "modelopt": |
| 379 | return maybe_compose_online_quantization(quant_cls.from_config(config)) |
| 380 | else: |
| 381 | raise ValueError( |
| 382 | f"Unsupported quantization config" |
| 383 | f" found for {model_config.quantization} in {f}." |
| 384 | ) |
| 385 | |
| 386 | return maybe_compose_online_quantization(quant_cls.from_config(config)) |
| 387 | |
| 388 | |
| 389 | def get_sparse_attention_config( |
| 390 | model_config: ModelConfig, |
| 391 | load_config: LoadConfig, |
| 392 | sparse_attention_config_filename: str = "sparse_attention_config.json", |
| 393 | ) -> dict[str, Any]: |
| 394 | model_name_or_path = model_config.model |
| 395 | is_local = os.path.isdir(model_name_or_path) |
| 396 | if not is_local: |
| 397 | # Download the config files. |
| 398 | with get_lock(model_name_or_path, load_config.download_dir): |
| 399 | hf_folder = hf_api().snapshot_download( |
| 400 | model_name_or_path, |
| 401 | revision=model_config.revision, |
| 402 | allow_patterns="*.json", |
| 403 | cache_dir=load_config.download_dir, |
| 404 | local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, |
| 405 | tqdm_class=DisabledTqdm, |
| 406 | ) |
| 407 | else: |
| 408 | hf_folder = model_name_or_path |
| 409 | |
| 410 | config_file = os.path.join(hf_folder, sparse_attention_config_filename) |
| 411 | if not os.path.exists(config_file): |
| 412 | return {} |
| 413 | |
| 414 | # Load the sparse attention config. |
| 415 | with open(config_file) as f: |
| 416 | config = json.load(f) |
| 417 | logger.info("Loaded sparse attention config from %s", config_file) |
| 418 | |
| 419 | return config |
| 420 | |
| 421 | |
| 422 | @instrument(span_name="Download weights - HF") |
| 423 | def download_weights_from_hf( |
| 424 | model_name_or_path: str, |
| 425 | cache_dir: str | None, |
| 426 | allow_patterns: list[str], |
| 427 | revision: str | None = None, |
| 428 | subfolder: str | None = None, |
| 429 | ignore_patterns: str | list[str] | None = None, |
| 430 | ) -> str: |
| 431 | """Download model weights from Hugging Face Hub. |
| 432 | |
| 433 | Args: |
| 434 | model_name_or_path (str): The model name or path. |
| 435 | cache_dir (Optional[str]): The cache directory to store the model |
| 436 | weights. If None, will use HF defaults. |
| 437 | allow_patterns (list[str]): The allowed patterns for the |
| 438 | weight files. Files matched by any of the patterns will be |
| 439 | downloaded. |
| 440 | revision (Optional[str]): The revision of the model. |
| 441 | subfolder (Optional[str]): The subfolder within the model repository |
| 442 | to download weights from. |
| 443 | ignore_patterns (Optional[Union[str, list[str]]]): The patterns to |
| 444 | filter out the weight files. Files matched by any of the patterns |
| 445 | will be ignored. |
| 446 | |
| 447 | Returns: |
| 448 | str: The path to the downloaded model weights. |
| 449 | |
| 450 | """ |
| 451 | assert len(allow_patterns) > 0 |
| 452 | local_only = huggingface_hub.constants.HF_HUB_OFFLINE |
| 453 | if not local_only: |
| 454 | # Attempt to reduce allow_patterns to a single pattern |
| 455 | # so we only have to call snapshot_download once. |
| 456 | try: |
| 457 | fs = hf_fs() |
| 458 | file_list = fs.ls( |
| 459 | os.path.join(model_name_or_path, subfolder or ""), |
| 460 | detail=False, |
| 461 | revision=revision, |
| 462 | ) |
| 463 | |
| 464 | # If downloading safetensors and an index file exists, use the |
| 465 | # specific file names from the index to avoid downloading |
| 466 | # unnecessary files (e.g., from subdirectories like "original/"). |
| 467 | index_file = f"{model_name_or_path}/{SAFE_WEIGHTS_INDEX_NAME}" |
| 468 | if "*.safetensors" in allow_patterns and index_file in file_list: |
| 469 | index_path = hf_api().hf_hub_download( |
| 470 | repo_id=model_name_or_path, |
| 471 | filename=SAFE_WEIGHTS_INDEX_NAME, |
| 472 | cache_dir=cache_dir, |
| 473 | revision=revision, |
| 474 | subfolder=subfolder, |
| 475 | ) |
| 476 | with open(index_path) as f: |
| 477 | weight_map = json.load(f)["weight_map"] |
| 478 | if weight_map: |
| 479 | # Extra [] so that weight_map files are treated as a |
| 480 | # single allow_pattern in the loop below |
| 481 | allow_patterns = [list(set(weight_map.values()))] # type: ignore[list-item] |
| 482 | else: |
| 483 | allow_patterns = ["*.safetensors"] |
| 484 | else: |
| 485 | # Use the first pattern found in the HF repo's files. |
| 486 | for pattern in allow_patterns: |
| 487 | if fnmatch.filter(file_list, pattern): |
| 488 | allow_patterns = [pattern] |
| 489 | break |
| 490 | except Exception as e: |
| 491 | logger.warning( |
| 492 | "Failed to get file list for '%s'. Trying each pattern in " |
| 493 | "allow_patterns individually until weights have been " |
| 494 | "downloaded. Error: %s", |
| 495 | model_name_or_path, |
| 496 | e, |
| 497 | ) |
| 498 | |
| 499 | logger.debug("Using model weights format %s", allow_patterns) |
| 500 | # Use file lock to prevent multiple processes from |
| 501 | # downloading the same model weights at the same time. |
| 502 | with get_lock(model_name_or_path, cache_dir): |
| 503 | start_time = time.perf_counter() |
| 504 | for allow_pattern in allow_patterns: |
| 505 | hf_folder = hf_api().snapshot_download( |
| 506 | model_name_or_path, |
| 507 | allow_patterns=allow_pattern, |
| 508 | ignore_patterns=ignore_patterns, |
| 509 | cache_dir=cache_dir, |
| 510 | tqdm_class=DisabledTqdm, |
| 511 | revision=revision, |
| 512 | local_files_only=local_only, |
| 513 | ) |
| 514 | # If we have downloaded weights for this allow_pattern, |
| 515 | # we don't need to check the rest. |
| 516 | # allow_pattern can be a list (from weight_map) or str (glob) |
| 517 | if isinstance(allow_pattern, list): |
| 518 | break |
| 519 | if any(Path(hf_folder).glob(allow_pattern)): |
| 520 | break |
| 521 | time_taken = time.perf_counter() - start_time |
| 522 | if time_taken > 0.5: |
| 523 | logger.info( |
| 524 | "Time spent downloading weights for %s: %.6f seconds", |
| 525 | model_name_or_path, |
| 526 | time_taken, |
| 527 | ) |
| 528 | return hf_folder |
| 529 | |
| 530 | |
| 531 | def download_safetensors_index_file_from_hf( |
| 532 | model_name_or_path: str, |
| 533 | index_file: str, |
| 534 | cache_dir: str | None, |
| 535 | subfolder: str | None = None, |
| 536 | revision: str | None = None, |
| 537 | ) -> None: |
| 538 | """Download hf safetensors index file from Hugging Face Hub. |
| 539 | |
| 540 | Args: |
| 541 | model_name_or_path (str): The model name or path. |
| 542 | index_file (str): The safetensors index file name |
| 543 | cache_dir (Optional[str]): The cache directory to store the model |
| 544 | weights. If None, will use HF defaults. |
| 545 | subfolder (Optional[str]): The subfolder within the model repository |
| 546 | to download weights from. |
| 547 | revision (Optional[str]): The revision of the model. |
| 548 | |
| 549 | """ |
| 550 | # Use file lock to prevent multiple processes from |
| 551 | # downloading the same model weights at the same time. |
| 552 | with get_lock(model_name_or_path, cache_dir): |
| 553 | try: |
| 554 | # Download the safetensors index file. |
| 555 | hf_api().hf_hub_download( |
| 556 | repo_id=model_name_or_path, |
| 557 | filename=index_file, |
| 558 | cache_dir=cache_dir, |
| 559 | revision=revision, |
| 560 | subfolder=subfolder, |
| 561 | local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, |
| 562 | ) |
| 563 | # If file not found on remote or locally, we should not fail since |
| 564 | # only some models will have index_file. |
| 565 | except huggingface_hub.utils.LocalEntryNotFoundError: |
| 566 | logger.info("No %s found in local cache.", index_file) |
| 567 | except huggingface_hub.utils.EntryNotFoundError: |
| 568 | logger.info("No %s found in remote.", index_file) |
| 569 | |
| 570 | |
| 571 | # For models like Mistral-7B-v0.3, there are both sharded |
| 572 | # safetensors files and a consolidated safetensors file. |
| 573 | # Passing both of these to the weight loader functionality breaks. |
| 574 | # So, we use the index_file to |
| 575 | # look up which safetensors files should be used. |
| 576 | def filter_duplicate_safetensors_files( |
| 577 | hf_weights_files: list[str], hf_folder: str, index_file: str |
| 578 | ) -> list[str]: |
| 579 | # model.safetensors.index.json is a mapping from keys in the |
| 580 | # torch state_dict to safetensors file holding that weight. |
| 581 | index_file_name = os.path.join(hf_folder, index_file) |
| 582 | if not os.path.isfile(index_file_name): |
| 583 | return hf_weights_files |
| 584 | |
| 585 | # Iterate through the weight_map (weight_name: safetensors files) |
| 586 | # to identify weights that we should use. |
| 587 | with open(index_file_name) as f: |
| 588 | weight_map = json.load(f)["weight_map"] |
| 589 | weight_files_in_index = set() |
| 590 | for weight_name in weight_map: |
| 591 | weight_files_in_index.add(os.path.join(hf_folder, weight_map[weight_name])) |
| 592 | # Check if files referenced in model.safetensors.index.json actually exist. |
| 593 | # Raise error if any file is missing. |
| 594 | hf_weights_files_set = set(hf_weights_files) |
| 595 | missing_files = weight_files_in_index - hf_weights_files_set |
| 596 | if missing_files: |
| 597 | raise FileNotFoundError( |
| 598 | f"Weight files referenced in index but missing: {missing_files}" |
| 599 | ) |
| 600 | # Filter out any fields that are not found in the index file. |
| 601 | hf_weights_files = [f for f in hf_weights_files if f in weight_files_in_index] |
| 602 | return hf_weights_files |
| 603 | |
| 604 | |
| 605 | def filter_files_not_needed_for_inference(hf_weights_files: list[str]) -> list[str]: |
| 606 | """Exclude files that are not needed for inference. |
| 607 | |
| 608 | See https://github.com/huggingface/transformers/blob/v4.34.0/src/transformers/trainer.py#L227-L233 |
| 609 | """ |
| 610 | blacklist = [ |
| 611 | "training_args.bin", |
| 612 | "optimizer.bin", |
| 613 | "optimizer.pt", |
| 614 | "scheduler.pt", |
| 615 | "scaler.pt", |
| 616 | ] |
| 617 | hf_weights_files = [ |
| 618 | f for f in hf_weights_files if not any(f.endswith(x) for x in blacklist) |
| 619 | ] |
| 620 | return hf_weights_files |
| 621 | |
| 622 | |
| 623 | def _normalize_module_prefixes( |
| 624 | language_model_names: Iterable[str] | None, |
| 625 | ) -> tuple[str, ...]: |
| 626 | names = tuple(language_model_names or ()) |
| 627 | return tuple(n if n.endswith(".") else f"{n}." for n in names) |
| 628 | |
| 629 | |
| 630 | def _mapped_weight_name(weights_mapper: "WeightsMapper | None", key: str) -> str | None: |
| 631 | """Apply ``WeightsMapper.map_name`` when present; else identity.""" |
| 632 | if weights_mapper is None: |
| 633 | return key |
| 634 | return weights_mapper.map_name(key) |
| 635 | |
| 636 | |
| 637 | def _hf_prefix_maps_to_lm( |
| 638 | new_name: str | None, lm_module_prefixes: tuple[str, ...] |
| 639 | ) -> bool: |
| 640 | if not new_name: |
| 641 | return False |
| 642 | new_p = new_name if new_name.endswith(".") else f"{new_name}." |
| 643 | return any(new_p.startswith(p) or p.startswith(new_p) for p in lm_module_prefixes) |
| 644 | |
| 645 | |
| 646 | def _shared_hf_root_module_prefixes( |
| 647 | lm_module_prefixes: tuple[str, ...], |
| 648 | weights_mapper: "WeightsMapper | None", |
| 649 | ) -> tuple[str, ...]: |
| 650 | """Module prefixes that are also an HF root for non-LM weights. |
| 651 | |
| 652 | E.g. Molmo/Phi-4-MM/Muse mark LM as ``model`` while vision HF keys stay |
| 653 | under ``model.vision_*`` / ``model.embed_tokens_extend.*``. Detected via |
| 654 | ``WeightsMapper.orig_to_new_prefix`` (no hard-coded name list). |
| 655 | """ |
| 656 | if weights_mapper is None: |
| 657 | return () |
| 658 | mapping = weights_mapper.orig_to_new_prefix |
| 659 | if not mapping: |
| 660 | return () |
| 661 | |
| 662 | non_lm_origs: list[str] = [] |
| 663 | for orig, new in mapping.items(): |
| 664 | orig_p = orig if orig.endswith(".") else f"{orig}." |
| 665 | if not _hf_prefix_maps_to_lm(new, lm_module_prefixes): |
| 666 | non_lm_origs.append(orig_p) |
| 667 | |
| 668 | if not non_lm_origs: |
| 669 | return () |
| 670 | |
| 671 | shared = [ |
| 672 | p |
| 673 | for p in lm_module_prefixes |
| 674 | if any(orig.startswith(p) for orig in non_lm_origs) |
| 675 | ] |
| 676 | return tuple(shared) |
| 677 | |
| 678 | |
| 679 | def resolve_mm_encoder_only_lm_prefixes( |
| 680 | language_model_names: Iterable[str] | None, |
| 681 | *, |
| 682 | weights_mapper: "WeightsMapper | None" = None, |
| 683 | ) -> tuple[str, ...] | None: |
| 684 | """Resolve vLLM LM *module* prefixes for ``--mm-encoder-only`` shard skip. |
| 685 | |
| 686 | Prefixes come from ``_language_model_names``. Classification of HF index |
| 687 | keys is done later via optional ``weights_mapper`` (see |
| 688 | ``filter_mm_encoder_only_safetensors_files``), so this helper does **not** |
| 689 | hard-code Qwen/HF nest lists. |
| 690 | |
| 691 | Returns ``None`` (leave the safetensors file list unchanged) when |
| 692 | ``_language_model_names`` is empty/missing, or when a module prefix is a |
| 693 | shared HF checkpoint root for both LM and non-LM weights — fail-closed so |
| 694 | Molmo / Phi-4-MM / Muse cannot under-load the encoder. |
| 695 | """ |
| 696 | prefixes = _normalize_module_prefixes(language_model_names) |
| 697 | if not prefixes: |
| 698 | return None |
| 699 | shared = _shared_hf_root_module_prefixes(prefixes, weights_mapper) |
| 700 | if shared: |
| 701 | logger.warning_once( |
| 702 | "mm-encoder-only whole-shard filter not applied: LM module " |
| 703 | "prefix(es) %s share an HF checkpoint root with non-LM weights " |
| 704 | "(via WeightsMapper). Encoder-only load keeps the full " |
| 705 | "safetensors file list; finer HF prefixes or key-level " |
| 706 | "filtering can enable skip later.", |
| 707 | shared, |
| 708 | ) |
| 709 | return None |
| 710 | return prefixes |
| 711 | |
| 712 | |
| 713 | def filter_mm_encoder_only_safetensors_files( |
| 714 | hf_weights_files: list[str], |
| 715 | hf_folder: str, |
| 716 | index_file: str, |
| 717 | language_model_prefixes: Iterable[str], |
| 718 | *, |
| 719 | weights_mapper: "WeightsMapper | None" = None, |
| 720 | ) -> list[str]: |
| 721 | """Drop safetensors shards that only contain language-model weights. |
| 722 | |
| 723 | Used with ``--mm-encoder-only`` so Encoder-only EPD instances avoid reading |
| 724 | pure LM shards from disk/DRAM. |
| 725 | |
| 726 | Each HF index key is classified by mapping through ``weights_mapper`` (when |
| 727 | provided) and testing the *vLLM* name against ``language_model_prefixes``. |
| 728 | Without a mapper, HF names are compared directly (identity checkpoint |
| 729 | layout, e.g. Kimi ``language_model.*``). |
| 730 | |
| 731 | A shard is kept if it contains any non-LM key. Without an index file, |
| 732 | returns ``hf_weights_files`` unchanged. |
| 733 | """ |
| 734 | prefixes = tuple(language_model_prefixes) |
| 735 | if not prefixes: |
| 736 | return hf_weights_files |
| 737 | |
| 738 | index_path = os.path.join(hf_folder, index_file) |
| 739 | if not os.path.isfile(index_path): |
| 740 | logger.warning_once( |
| 741 | "mm-encoder-only shard filter skipped: index file %s not found", |
| 742 | index_path, |
| 743 | ) |
| 744 | return hf_weights_files |
| 745 | |
| 746 | with open(index_path) as f: |
| 747 | weight_map: dict[str, str] = json.load(f)["weight_map"] |
| 748 | |
| 749 | keys_by_file: dict[str, list[str]] = {} |
| 750 | for weight_name, weight_file in weight_map.items(): |
| 751 | keys_by_file.setdefault(weight_file, []).append(weight_name) |
| 752 | |
| 753 | def _is_lm_key(key: str) -> bool: |
| 754 | mapped = _mapped_weight_name(weights_mapper, key) |
| 755 | if mapped is None: |
| 756 | return False |
| 757 | return any(mapped.startswith(p) for p in prefixes) |
| 758 | |
| 759 | def _is_lm_only(weight_file: str) -> bool: |
| 760 | keys = keys_by_file.get(weight_file) |
| 761 | if not keys: |
| 762 | # Not listed in the index (e.g. extra matched glob) — keep. |
| 763 | return False |
| 764 | return all(_is_lm_key(key) for key in keys) |
| 765 | |
| 766 | kept: list[str] = [] |
| 767 | skipped = 0 |
| 768 | for path in hf_weights_files: |
| 769 | rel = os.path.relpath(path, hf_folder) |
| 770 | # Index weight_map values are usually basenames; tolerate either form. |
| 771 | basename = os.path.basename(path) |
| 772 | if _is_lm_only(rel) or _is_lm_only(basename): |
| 773 | skipped += 1 |
| 774 | continue |
| 775 | kept.append(path) |
| 776 | |
| 777 | if skipped: |
| 778 | logger.info_once( |
| 779 | "mm-encoder-only: skipped %d/%d safetensors shard(s) that only " |
| 780 | "contain language-model weights (prefixes=%s)", |
| 781 | skipped, |
| 782 | skipped + len(kept), |
| 783 | prefixes, |
| 784 | ) |
| 785 | return kept |
| 786 | |
| 787 | |
| 788 | # explicitly use pure text format, with a newline at the end |
| 789 | # this makes it impossible to see the animation in the progress bar |
| 790 | # but will avoid messing up with ray or multiprocessing, which wraps |
| 791 | # each line of output with some prefix. |
| 792 | _BAR_FORMAT = "{desc}: {percentage:3.0f}% Completed | {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}]\n" # noqa: E501 |
| 793 | |
| 794 | |
| 795 | def enable_tqdm(use_tqdm_on_load: bool): |
| 796 | return use_tqdm_on_load and ( |
| 797 | not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0 |
| 798 | ) |
| 799 | |
| 800 | |
| 801 | def np_cache_weights_iterator( |
| 802 | model_name_or_path: str, |
| 803 | cache_dir: str | None, |
| 804 | hf_folder: str, |
| 805 | hf_weights_files: list[str], |
| 806 | use_tqdm_on_load: bool, |
| 807 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 808 | """Iterate over the weights in the model np files. |
| 809 | |
| 810 | Will dump the model weights to numpy files if they are not already dumped. |
| 811 | """ |
| 812 | # Convert the model weights from torch tensors to numpy arrays for |
| 813 | # faster loading. |
| 814 | np_folder = os.path.join(hf_folder, "np") |
| 815 | os.makedirs(np_folder, exist_ok=True) |
| 816 | weight_names_file = os.path.join(np_folder, "weight_names.json") |
| 817 | # Use file lock to prevent multiple processes from |
| 818 | # dumping the same model weights to numpy at the same time. |
| 819 | with get_lock(model_name_or_path, cache_dir): |
| 820 | if not os.path.exists(weight_names_file): |
| 821 | weight_names: list[str] = [] |
| 822 | for bin_file in tqdm( |
| 823 | hf_weights_files, |
| 824 | desc="Loading np_cache checkpoint shards", |
| 825 | disable=not enable_tqdm(use_tqdm_on_load), |
| 826 | bar_format=_BAR_FORMAT, |
| 827 | ): |
| 828 | state = torch.load(bin_file, map_location="cpu", weights_only=True) |
| 829 | for name, param in state.items(): |
| 830 | param_path = os.path.join(np_folder, name) |
| 831 | with open(param_path, "wb") as f: |
| 832 | np.save(f, param.cpu().detach().numpy()) |
| 833 | weight_names.append(name) |
| 834 | with open(weight_names_file, "w") as f: |
| 835 | json.dump(weight_names, f) |
| 836 | |
| 837 | with open(weight_names_file) as f: |
| 838 | weight_names = json.load(f) |
| 839 | |
| 840 | for name in weight_names: |
| 841 | param_path = os.path.join(np_folder, name) |
| 842 | with open(param_path, "rb") as f: |
| 843 | param = np.load(f) |
| 844 | yield name, torch.from_numpy(param) |
| 845 | |
| 846 | |
| 847 | def drop_checkpoint_cache( |
| 848 | model_name_or_path: str, |
| 849 | revision: str | None = None, |
| 850 | cache_dir: str | None = None, |
| 851 | ) -> None: |
| 852 | checkpoint_dir: Path | None = None |
| 853 | if os.path.isdir(model_name_or_path): |
| 854 | checkpoint_dir = Path(model_name_or_path) |
| 855 | else: |
| 856 | with suppress(OSError, ValueError): |
| 857 | checkpoint_dir = Path( |
| 858 | hf_api().snapshot_download( |
| 859 | model_name_or_path, |
| 860 | revision=revision, |
| 861 | cache_dir=cache_dir, |
| 862 | local_files_only=True, |
| 863 | ) |
| 864 | ) |
| 865 | if checkpoint_dir is None: |
| 866 | logger.info_once( |
| 867 | "No local checkpoint found for %s; skipping page-cache eviction", |
| 868 | model_name_or_path, |
| 869 | ) |
| 870 | return |
| 871 | for path in checkpoint_dir.glob("*.safetensors"): |
| 872 | try: |
| 873 | with path.open("rb") as file: |
| 874 | os.posix_fadvise(file.fileno(), 0, 0, os.POSIX_FADV_DONTNEED) |
| 875 | except OSError as exc: |
| 876 | logger.warning( |
| 877 | "Could not release checkpoint page cache for %s: %s", path, exc |
| 878 | ) |
| 879 | |
| 880 | |
| 881 | def _get_checkpoints_size_bytes(files: list[str]) -> int: |
| 882 | """Return the total size of the checkpoint files in bytes.""" |
| 883 | if not files: |
| 884 | return 0 |
| 885 | return sum(os.path.getsize(f) for f in files) |
| 886 | |
| 887 | |
| 888 | def _get_available_ram_bytes() -> int: |
| 889 | """Return available RAM, honoring cgroup limits.""" |
| 890 | import psutil |
| 891 | |
| 892 | host_available = psutil.virtual_memory().available |
| 893 | |
| 894 | from vllm.utils.cpu_resource_utils import ( |
| 895 | get_cgroup_memory_limit, |
| 896 | get_cgroup_memory_usage, |
| 897 | ) |
| 898 | |
| 899 | cgroup_limit = get_cgroup_memory_limit() |
| 900 | if cgroup_limit is None: |
| 901 | return host_available |
| 902 | cgroup_usage = get_cgroup_memory_usage() |
| 903 | cgroup_available = ( |
| 904 | cgroup_limit if cgroup_usage is None else max(0, cgroup_limit - cgroup_usage) |
| 905 | ) |
| 906 | return min(host_available, cgroup_available) |
| 907 | |
| 908 | |
| 909 | def _get_fs_type(files: list[str]) -> str: |
| 910 | """Get the filesystem type of the first file in *files* (Linux only).""" |
| 911 | if not files: |
| 912 | return "" |
| 913 | try: |
| 914 | # Only the first file is checked — all checkpoint shards reside |
| 915 | # in the same directory and therefore on the same filesystem. |
| 916 | resolved = os.path.realpath(files[0]) |
| 917 | best_mount = "" |
| 918 | best_fstype = "" |
| 919 | # /proc/mounts may contain nested mount points (e.g. "/" -> ext4, |
| 920 | # "/data" -> nfs4, "/data/local" -> ext4). We pick the entry with |
| 921 | # the longest matching mount_point — the same "longest prefix match" |
| 922 | # rule the kernel uses to decide which filesystem serves a path. |
| 923 | with open("/proc/mounts") as f: |
| 924 | for line in f: |
| 925 | parts = line.split() |
| 926 | if len(parts) < 3: |
| 927 | continue |
| 928 | mount_point, fstype = parts[1], parts[2] |
| 929 | if ( |
| 930 | resolved == mount_point |
| 931 | or resolved.startswith(os.path.join(mount_point, "")) |
| 932 | ) and len(mount_point) > len(best_mount): |
| 933 | best_mount = mount_point |
| 934 | best_fstype = fstype |
| 935 | return best_fstype |
| 936 | except Exception: |
| 937 | # /proc/mounts is Linux-specific; on other OSes (or if the read |
| 938 | # fails for any reason) we fall back to an empty string. |
| 939 | return "" |
| 940 | |
| 941 | |
| 942 | def _prefetch_checkpoint( |
| 943 | file_path: str, |
| 944 | block_size: int = DEFAULT_SAFETENSORS_PREFETCH_BLOCK_SIZE, |
| 945 | ) -> None: |
| 946 | """Prefetch a checkpoint file into the OS page cache. |
| 947 | |
| 948 | Reads the file in blocks so the kernel caches its pages before workers load |
| 949 | the same file. |
| 950 | """ |
| 951 | if block_size < 1: |
| 952 | raise ValueError("safetensors prefetch block size must be >= 1") |
| 953 | |
| 954 | with open(file_path, "rb") as f: |
| 955 | while f.read(block_size): |
| 956 | pass |
| 957 | |
| 958 | |
| 959 | def _prefetch_all_checkpoints( |
| 960 | sorted_files: list[str], |
| 961 | num_prefetch_threads: int = DEFAULT_SAFETENSORS_PREFETCH_NUM_THREADS, |
| 962 | block_size: int = DEFAULT_SAFETENSORS_PREFETCH_BLOCK_SIZE, |
| 963 | ) -> None: |
| 964 | """Start prefetching checkpoint files into page cache in a background thread.""" |
| 965 | if num_prefetch_threads < 1: |
| 966 | raise ValueError("safetensors prefetch num threads must be >= 1") |
| 967 | if block_size < 1: |
| 968 | raise ValueError("safetensors prefetch block size must be >= 1") |
| 969 | |
| 970 | if torch.distributed.is_initialized(): |
| 971 | rank = torch.distributed.get_rank() |
| 972 | world_size = torch.distributed.get_world_size() |
| 973 | else: |
| 974 | rank = 0 |
| 975 | world_size = 1 |
| 976 | paths_to_prefetch = sorted_files[rank::world_size] |
| 977 | total_for_rank = len(paths_to_prefetch) |
| 978 | |
| 979 | async def _prefetch_all() -> None: |
| 980 | loop = asyncio.get_running_loop() |
| 981 | completed = 0 |
| 982 | next_log_pct = 10 |
| 983 | |
| 984 | async def prefetch_one( |
| 985 | path: str, |
| 986 | executor: concurrent.futures.ThreadPoolExecutor, |
| 987 | ) -> None: |
| 988 | nonlocal completed, next_log_pct |
| 989 | try: |
| 990 | await loop.run_in_executor( |
| 991 | executor, _prefetch_checkpoint, path, block_size |
| 992 | ) |
| 993 | completed += 1 |
| 994 | if total_for_rank > 0 and next_log_pct <= 100: |
| 995 | pct = 100 * completed / total_for_rank |
| 996 | if pct >= next_log_pct: |
| 997 | logger.info( |
| 998 | "Prefetching checkpoint files: %d%% (%d/%d)", |
| 999 | next_log_pct, |
| 1000 | completed, |
| 1001 | total_for_rank, |
| 1002 | ) |
| 1003 | next_log_pct += 10 |
| 1004 | except Exception: |
| 1005 | logger.warning( |
| 1006 | "Failed to prefetch checkpoint file %r.", path, exc_info=True |
| 1007 | ) |
| 1008 | |
| 1009 | with concurrent.futures.ThreadPoolExecutor( |
| 1010 | max_workers=num_prefetch_threads |
| 1011 | ) as executor: |
| 1012 | await asyncio.gather( |
| 1013 | *(prefetch_one(p, executor) for p in paths_to_prefetch) |
| 1014 | ) |
| 1015 | |
| 1016 | def _run_prefetch() -> None: |
| 1017 | start = time.perf_counter() |
| 1018 | asyncio.run(_prefetch_all()) |
| 1019 | elapsed = time.perf_counter() - start |
| 1020 | logger.info( |
| 1021 | "Prefetching checkpoint files into page cache finished in %.2fs", |
| 1022 | elapsed, |
| 1023 | ) |
| 1024 | |
| 1025 | logger.info( |
| 1026 | "Prefetching checkpoint files into page cache started " |
| 1027 | "(in background, num_threads=%d, block_size=%d bytes)", |
| 1028 | num_prefetch_threads, |
| 1029 | block_size, |
| 1030 | ) |
| 1031 | threading.Thread(target=_run_prefetch, daemon=True).start() |
| 1032 | |
| 1033 | |
| 1034 | def safetensors_weights_iterator( |
| 1035 | hf_weights_files: list[str], |
| 1036 | use_tqdm_on_load: bool, |
| 1037 | safetensors_load_strategy: str | None = None, |
| 1038 | local_expert_ids: set[int] | None = None, |
| 1039 | *, |
| 1040 | safetensors_prefetch_num_threads: int = DEFAULT_SAFETENSORS_PREFETCH_NUM_THREADS, |
| 1041 | safetensors_prefetch_block_size: int = DEFAULT_SAFETENSORS_PREFETCH_BLOCK_SIZE, |
| 1042 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1043 | """Iterate over the weights in the model safetensor files. |
| 1044 | |
| 1045 | When *local_expert_ids* is provided, expert weights not belonging to |
| 1046 | this rank are skipped **before** reading from disk, which drastically |
| 1047 | reduces storage I/O for MoE models under EP. |
| 1048 | """ |
| 1049 | loading_desc = "Loading safetensors checkpoint shards" |
| 1050 | if safetensors_load_strategy == "eager": |
| 1051 | loading_desc += " (eager)" |
| 1052 | |
| 1053 | sorted_files = sorted(hf_weights_files, key=_natural_sort_key) |
| 1054 | |
| 1055 | fs_type = _get_fs_type(sorted_files) |
| 1056 | is_net_fs = fs_type in ("nfs", "nfs4", "lustre") |
| 1057 | total_bytes = _get_checkpoints_size_bytes(sorted_files) |
| 1058 | avail_bytes = _get_available_ram_bytes() |
| 1059 | ram_threshold_pct = 90 |
| 1060 | fits_in_ram = total_bytes <= (ram_threshold_pct / 100.0) * avail_bytes |
| 1061 | fs_name = fs_type.upper() if fs_type else "unknown" |
| 1062 | |
| 1063 | logger.info_once( |
| 1064 | "Filesystem type for checkpoints: %s. Checkpoint size: %.2f GiB. " |
| 1065 | "Available RAM: %.2f GiB.", |
| 1066 | fs_name, |
| 1067 | total_bytes / 1024**3, |
| 1068 | avail_bytes / 1024**3, |
| 1069 | ) |
| 1070 | |
| 1071 | should_prefetch = safetensors_load_strategy == "prefetch" |
| 1072 | if safetensors_load_strategy is None: |
| 1073 | if is_net_fs and fits_in_ram: |
| 1074 | should_prefetch = True |
| 1075 | elif is_net_fs and not fits_in_ram: |
| 1076 | logger.warning_once( |
| 1077 | "Network filesystem (%s) detected but checkpoint total size " |
| 1078 | "(%.2f GiB) exceeds %d%% of available RAM (%.2f GiB). " |
| 1079 | "Skipping auto-prefetch.", |
| 1080 | fs_name, |
| 1081 | total_bytes / 1024**3, |
| 1082 | ram_threshold_pct, |
| 1083 | avail_bytes / 1024**3, |
| 1084 | ) |
| 1085 | elif not is_net_fs and fits_in_ram: |
| 1086 | logger.info_once( |
| 1087 | "Auto-prefetch is disabled because the filesystem (%s) is not a " |
| 1088 | "recognized network FS (NFS/Lustre). If you want to force " |
| 1089 | "prefetching, start vLLM with --safetensors-load-strategy=prefetch.", |
| 1090 | fs_name, |
| 1091 | ) |
| 1092 | elif not is_net_fs and not fits_in_ram: |
| 1093 | logger.info_once( |
| 1094 | "Auto-prefetch is disabled because the filesystem (%s) is not a " |
| 1095 | "recognized network FS (NFS/Lustre) and the checkpoint size " |
| 1096 | "(%.2f GiB) exceeds %d%% of available RAM (%.2f GiB).", |
| 1097 | fs_name, |
| 1098 | total_bytes / 1024**3, |
| 1099 | ram_threshold_pct, |
| 1100 | avail_bytes / 1024**3, |
| 1101 | ) |
| 1102 | elif should_prefetch and not fits_in_ram: |
| 1103 | logger.warning_once( |
| 1104 | "safetensors_load_strategy='prefetch' was explicitly specified, but " |
| 1105 | "checkpoint total size (%.2f GiB) exceeds %d%% of available RAM " |
| 1106 | "(%.2f GiB). This may cause out-of-memory errors.", |
| 1107 | total_bytes / 1024**3, |
| 1108 | ram_threshold_pct, |
| 1109 | avail_bytes / 1024**3, |
| 1110 | ) |
| 1111 | |
| 1112 | if should_prefetch: |
| 1113 | _prefetch_all_checkpoints( |
| 1114 | sorted_files, |
| 1115 | num_prefetch_threads=safetensors_prefetch_num_threads, |
| 1116 | block_size=safetensors_prefetch_block_size, |
| 1117 | ) |
| 1118 | |
| 1119 | leftover_state_dict: dict[str, torch.Tensor] = {} |
| 1120 | for st_file in tqdm( |
| 1121 | sorted_files, |
| 1122 | desc=loading_desc, |
| 1123 | disable=not enable_tqdm(use_tqdm_on_load), |
| 1124 | bar_format=_BAR_FORMAT, |
| 1125 | ): |
| 1126 | if safetensors_load_strategy == "eager": |
| 1127 | with open(st_file, "rb") as f: |
| 1128 | state_dict = load(f.read()) |
| 1129 | for name, param in state_dict.items(): |
| 1130 | if not should_skip_weight(name, local_expert_ids): |
| 1131 | yield name, param |
| 1132 | elif safetensors_load_strategy == "torchao": |
| 1133 | # we can't load flattened torchao tensor subclasses directly into the model |
| 1134 | # instead we reconstruct the subclasses here before returning |
| 1135 | if not torchao_version_at_least("0.15.0"): |
| 1136 | raise ValueError( |
| 1137 | "Please use torchao version >= 0.15.0 " |
| 1138 | "to load torchao safetensors checkpoint" |
| 1139 | ) |
| 1140 | from torchao.prototype.safetensors.safetensors_support import ( |
| 1141 | unflatten_tensor_state_dict, |
| 1142 | ) |
| 1143 | |
| 1144 | with safe_open(st_file, framework="pt") as f: |
| 1145 | state_dict = {} |
| 1146 | for name in f.keys(): # noqa: SIM118 |
| 1147 | if should_skip_weight(name, local_expert_ids): |
| 1148 | continue |
| 1149 | state_dict[name] = f.get_tensor(name) |
| 1150 | |
| 1151 | # update with leftover tensor data from previous iteration, if any |
| 1152 | state_dict.update(leftover_state_dict) |
| 1153 | metadata = f.metadata() |
| 1154 | # due to sharded checkpoints, we are not guaranteed that we have all |
| 1155 | # tensor subclass data on one file |
| 1156 | # state_dict has the leftover data from this step and we wait for |
| 1157 | # missing information to be provided in a future iteration |
| 1158 | unflattened_state_dict, leftover_state_dict = ( |
| 1159 | unflatten_tensor_state_dict(state_dict, metadata) |
| 1160 | ) |
| 1161 | yield from unflattened_state_dict.items() |
| 1162 | else: |
| 1163 | with safe_open(st_file, framework="pt") as f: |
| 1164 | for name in f.keys(): # noqa: SIM118 |
| 1165 | if should_skip_weight(name, local_expert_ids): |
| 1166 | continue |
| 1167 | param = f.get_tensor(name) |
| 1168 | yield name, param |
| 1169 | |
| 1170 | |
| 1171 | def multi_thread_safetensors_weights_iterator( |
| 1172 | hf_weights_files: list[str], |
| 1173 | use_tqdm_on_load: bool, |
| 1174 | max_workers: int = 4, |
| 1175 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1176 | """Multi-Thread iterate over the weights in the model safetensor files.""" |
| 1177 | |
| 1178 | def _load_file(st_file: str): |
| 1179 | result = load_file(st_file, device="cpu") |
| 1180 | return result |
| 1181 | |
| 1182 | with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor: |
| 1183 | # Note to use generator here so we do not store all the loaded files in memory |
| 1184 | # at the same time, which can cause OOM for large models. |
| 1185 | futures = (executor.submit(_load_file, st_file) for st_file in hf_weights_files) |
| 1186 | futures_iter = tqdm( |
| 1187 | concurrent.futures.as_completed(futures), |
| 1188 | total=len(hf_weights_files), |
| 1189 | desc="Multi-thread loading shards", |
| 1190 | disable=not enable_tqdm(use_tqdm_on_load), |
| 1191 | bar_format=_BAR_FORMAT, |
| 1192 | ) |
| 1193 | |
| 1194 | for future in futures_iter: |
| 1195 | state_dict = future.result() |
| 1196 | del future |
| 1197 | for key in list(state_dict): |
| 1198 | yield key, state_dict.pop(key) |
| 1199 | |
| 1200 | |
| 1201 | def runai_safetensors_weights_iterator( |
| 1202 | hf_weights_files: list[str], |
| 1203 | use_tqdm_on_load: bool, |
| 1204 | is_distributed: bool = False, |
| 1205 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1206 | """Iterate over the weights in the model safetensor files.""" |
| 1207 | with SafetensorsStreamer() as streamer: |
| 1208 | is_cuda_alike = current_platform.is_cuda_alike() |
| 1209 | device = ( |
| 1210 | f"cuda:{current_platform.current_device()}" |
| 1211 | if is_distributed and is_cuda_alike |
| 1212 | else "cpu" |
| 1213 | ) |
| 1214 | |
| 1215 | streamer.stream_files( |
| 1216 | hf_weights_files, |
| 1217 | device=device, |
| 1218 | is_distributed=is_distributed, |
| 1219 | ) |
| 1220 | total_tensors = sum( |
| 1221 | len(tensors_meta) |
| 1222 | for tensors_meta in streamer.files_to_tensors_metadata.values() |
| 1223 | ) |
| 1224 | |
| 1225 | tensor_iter = tqdm( |
| 1226 | streamer.get_tensors(), |
| 1227 | total=total_tensors, |
| 1228 | desc="Loading safetensors using Runai Model Streamer", |
| 1229 | bar_format=_BAR_FORMAT, |
| 1230 | disable=not enable_tqdm(use_tqdm_on_load), |
| 1231 | mininterval=2, |
| 1232 | ) |
| 1233 | |
| 1234 | for name, tensor in tensor_iter: |
| 1235 | yield name, tensor.clone() |
| 1236 | |
| 1237 | |
| 1238 | def fastsafetensors_weights_iterator( |
| 1239 | hf_weights_files: list[str], |
| 1240 | use_tqdm_on_load: bool, |
| 1241 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1242 | """Iterate over the weights in the model safetensor files |
| 1243 | using fastsafetensor library. |
| 1244 | |
| 1245 | Uses ParallelLoader for pipelined loading: the producer thread |
| 1246 | prepares metadata for the next shard while the consumer yields |
| 1247 | tensors from the current shard. |
| 1248 | """ |
| 1249 | from fastsafetensors.parallel_loader import ParallelLoader |
| 1250 | |
| 1251 | if torch.distributed.is_initialized(): |
| 1252 | pg = torch.distributed.group.WORLD |
| 1253 | else: |
| 1254 | pg = SingleGroup() |
| 1255 | |
| 1256 | device = torch.device(f"cuda:{current_platform.current_device()}") |
| 1257 | hf_weights_files = sorted(hf_weights_files, key=_natural_sort_key) |
| 1258 | |
| 1259 | # Use nogds=True for TP > 1 to avoid cuFileDriverOpen() which |
| 1260 | # initializes the GDS DMA subsystem for all visible GPUs, creating |
| 1261 | # unwanted CUDA contexts on every device. |
| 1262 | nogds = pg.size() > 1 |
| 1263 | |
| 1264 | queue_size = envs.VLLM_FASTSAFETENSORS_QUEUE_SIZE |
| 1265 | tqdm_enabled = enable_tqdm(use_tqdm_on_load) |
| 1266 | |
| 1267 | def _make_loader(nogds: bool) -> "ParallelLoader": |
| 1268 | return ParallelLoader( |
| 1269 | pg=pg, |
| 1270 | hf_weights_files=hf_weights_files, |
| 1271 | queue_size=queue_size, |
| 1272 | use_tqdm_on_load=tqdm_enabled, |
| 1273 | device=str(device), |
| 1274 | nogds=nogds, |
| 1275 | ) |
| 1276 | |
| 1277 | # GDS can fail either at construction or lazily inside the producer |
| 1278 | # thread during iteration (e.g. cuFileHandleRegister returning |
| 1279 | # CU_FILE_HANDLE_NOT_REGISTERED on a filesystem without GDS support). |
| 1280 | # Catch both and fall back to nogds, but only before yielding any |
| 1281 | # tensor -- restarting mid-stream would reload earlier shards. |
| 1282 | pl = None |
| 1283 | yielded = False |
| 1284 | try: |
| 1285 | try: |
| 1286 | pl = _make_loader(nogds) |
| 1287 | for name, tensor in pl.iterate_weights(): |
| 1288 | yielded = True |
| 1289 | yield name, tensor |
| 1290 | except RuntimeError as e: |
| 1291 | if nogds or yielded or "gds" not in str(e): |
| 1292 | raise |
| 1293 | logger.warning_once( |
| 1294 | "GDS not enabled, setting `nogds=True`.\n" |
| 1295 | "For more information, see: https://github.com/foundation-model-stack/" |
| 1296 | "fastsafetensors?tab=readme-ov-file#basic-api-usages" |
| 1297 | ) |
| 1298 | if pl is not None: |
| 1299 | pl.close() |
| 1300 | pl = _make_loader(nogds=True) |
| 1301 | yield from pl.iterate_weights() |
| 1302 | finally: |
| 1303 | if pl is not None: |
| 1304 | pl.close() |
| 1305 | |
| 1306 | |
| 1307 | def instanttensor_weights_iterator( |
| 1308 | hf_weights_files: list[str], |
| 1309 | use_tqdm_on_load: bool, |
| 1310 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1311 | """Iterate over the weights in the model safetensor files |
| 1312 | using instanttensor library.""" |
| 1313 | try: |
| 1314 | import instanttensor |
| 1315 | except ImportError as e: |
| 1316 | raise ImportError( |
| 1317 | "Please install instanttensor via `pip install vllm[instanttensor]`" |
| 1318 | ) from e |
| 1319 | |
| 1320 | if not current_platform.is_cuda(): |
| 1321 | raise ValueError("InstantTensor requires NVIDIA GPUs") |
| 1322 | |
| 1323 | try: |
| 1324 | world_group = get_world_group() |
| 1325 | except AssertionError: |
| 1326 | # Entering here only in unit tests where the world group is not initialized. |
| 1327 | process_group = None |
| 1328 | else: |
| 1329 | process_group = world_group.device_group if world_group.world_size > 1 else None |
| 1330 | |
| 1331 | device = current_platform.current_device() |
| 1332 | |
| 1333 | # copy=True yields tensors that own their memory, staying valid after the |
| 1334 | # context exits or InstantTensor reuses its buffer. |
| 1335 | with instanttensor.safe_open( |
| 1336 | hf_weights_files, |
| 1337 | framework="pt", |
| 1338 | device=device, |
| 1339 | process_group=process_group, |
| 1340 | copy=True, |
| 1341 | ) as f: |
| 1342 | # Track bytes so the bar reports load throughput (GB/s). |
| 1343 | pbar = tqdm( |
| 1344 | total=f.total_tensor_size, |
| 1345 | desc="Loading safetensors using InstantTensor loader", |
| 1346 | disable=not enable_tqdm(use_tqdm_on_load), |
| 1347 | bar_format=_BAR_FORMAT, |
| 1348 | position=tqdm._get_free_pos(), |
| 1349 | unit="B", |
| 1350 | unit_scale=True, |
| 1351 | unit_divisor=1024, |
| 1352 | mininterval=1.0, |
| 1353 | ) |
| 1354 | try: |
| 1355 | for name, tensor in f.tensors(): |
| 1356 | pbar.update(tensor.numel() * tensor.element_size()) |
| 1357 | yield name, tensor |
| 1358 | finally: |
| 1359 | pbar.close() |
| 1360 | |
| 1361 | |
| 1362 | def pt_weights_iterator( |
| 1363 | hf_weights_files: list[str], |
| 1364 | use_tqdm_on_load: bool, |
| 1365 | pt_load_map_location: str | dict[str, str] = "cpu", |
| 1366 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1367 | """Iterate over the weights in the model bin/pt files.""" |
| 1368 | for bin_file in tqdm( |
| 1369 | hf_weights_files, |
| 1370 | desc="Loading pt checkpoint shards", |
| 1371 | disable=not enable_tqdm(use_tqdm_on_load), |
| 1372 | bar_format=_BAR_FORMAT, |
| 1373 | ): |
| 1374 | state = torch.load( |
| 1375 | bin_file, map_location=pt_load_map_location, weights_only=True |
| 1376 | ) |
| 1377 | yield from state.items() |
| 1378 | del state |
| 1379 | |
| 1380 | |
| 1381 | def multi_thread_pt_weights_iterator( |
| 1382 | hf_weights_files: list[str], |
| 1383 | use_tqdm_on_load: bool, |
| 1384 | pt_load_map_location: str | dict[str, str] = "cpu", |
| 1385 | max_workers: int = 4, |
| 1386 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1387 | """Multi-Thread iterate over the weights in the model bin/pt files.""" |
| 1388 | |
| 1389 | def _load_file(bin_file: str): |
| 1390 | return torch.load( |
| 1391 | bin_file, map_location=pt_load_map_location, weights_only=True |
| 1392 | ) |
| 1393 | |
| 1394 | with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor: |
| 1395 | futures = [ |
| 1396 | executor.submit(_load_file, bin_file) for bin_file in hf_weights_files |
| 1397 | ] |
| 1398 | futures_iter = tqdm( |
| 1399 | concurrent.futures.as_completed(futures), |
| 1400 | total=len(hf_weights_files), |
| 1401 | desc="Multi-thread loading pt checkpoint shards", |
| 1402 | disable=not enable_tqdm(use_tqdm_on_load), |
| 1403 | bar_format=_BAR_FORMAT, |
| 1404 | ) |
| 1405 | |
| 1406 | for future in futures_iter: |
| 1407 | state = future.result() |
| 1408 | yield from state.items() |
| 1409 | del state |
| 1410 | |
| 1411 | |
| 1412 | def default_weight_loader(param: torch.Tensor, loaded_weight: torch.Tensor) -> None: |
| 1413 | """Default weight loader.""" |
| 1414 | try: |
| 1415 | if param.numel() == 1 and loaded_weight.numel() == 1: |
| 1416 | # Sometimes scalar values aren't considered tensors with shapes |
| 1417 | # so if both param and loaded_weight are a scalar, |
| 1418 | # reshape to match before copying |
| 1419 | param.data.copy_(loaded_weight.view(param.shape)) |
| 1420 | else: |
| 1421 | assert param.size() == loaded_weight.size(), ( |
| 1422 | f"Attempted to load weight ({loaded_weight.size()}) " |
| 1423 | f"into parameter ({param.size()})" |
| 1424 | ) |
| 1425 | |
| 1426 | param.data.copy_(loaded_weight) |
| 1427 | except Exception: |
| 1428 | # NOTE: This exception is added for the purpose of setting breakpoint to |
| 1429 | # debug weight loading issues. |
| 1430 | raise |
| 1431 | |
| 1432 | |
| 1433 | def row_parallel_weight_loader( |
| 1434 | param: torch.Tensor, loaded_weight: torch.Tensor |
| 1435 | ) -> None: |
| 1436 | """Load weights that are row-parallelized.""" |
| 1437 | tp_rank = get_tensor_model_parallel_rank() |
| 1438 | shard_dim = 0 if param.dim() != 1 else None |
| 1439 | |
| 1440 | if shard_dim is not None: |
| 1441 | shard_size = param.data.shape[shard_dim] |
| 1442 | start_idx = tp_rank * shard_size |
| 1443 | loaded_weight = loaded_weight.narrow(shard_dim, start_idx, shard_size) |
| 1444 | |
| 1445 | return default_weight_loader(param, loaded_weight) |
| 1446 | |
| 1447 | |
| 1448 | LoaderFunction = Callable[[torch.Tensor, torch.Tensor], None] |
| 1449 | |
| 1450 | |
| 1451 | def sharded_weight_loader(shard_axis: int) -> LoaderFunction: |
| 1452 | """Create a weight loader that shards the weights along the given axis""" |
| 1453 | |
| 1454 | def loader(param: torch.Tensor, loaded_weight: torch.Tensor) -> None: |
| 1455 | tp_rank = get_tensor_model_parallel_rank() |
| 1456 | |
| 1457 | shard_size = param.data.shape[shard_axis] |
| 1458 | start_idx = tp_rank * shard_size |
| 1459 | loaded_weight = loaded_weight.narrow(shard_axis, start_idx, shard_size) |
| 1460 | |
| 1461 | return default_weight_loader(param, loaded_weight) |
| 1462 | |
| 1463 | return loader |
| 1464 | |
| 1465 | |
| 1466 | def composed_weight_loader( |
| 1467 | loader: LoaderFunction, fn: Callable[[torch.Tensor], torch.Tensor] |
| 1468 | ) -> LoaderFunction: |
| 1469 | """Create a weight loader that post-processes the weights after loading""" |
| 1470 | |
| 1471 | def composed_loader(param: torch.Tensor, loaded_weight: torch.Tensor) -> None: |
| 1472 | loader(param, loaded_weight) |
| 1473 | param.data.copy_(fn(param)) |
| 1474 | return |
| 1475 | |
| 1476 | return composed_loader |
| 1477 | |
| 1478 | |
| 1479 | @torch.no_grad() |
| 1480 | def initialize_single_dummy_weight( |
| 1481 | param: torch.Tensor, |
| 1482 | low: float = -1e-3, |
| 1483 | high: float = 1e-3, |
| 1484 | seed: int = 1234, |
| 1485 | ) -> None: |
| 1486 | if param.device.type == "meta": |
| 1487 | return # deferred to finalize_layerwise_processing (e.g. online quant) |
| 1488 | |
| 1489 | if (dummy_weight_value := getattr(param, "dummy_weight_value", None)) is not None: |
| 1490 | param.fill_(dummy_weight_value) |
| 1491 | return |
| 1492 | |
| 1493 | if not torch.is_floating_point(param): |
| 1494 | if current_platform.is_rocm(): |
| 1495 | # On ROCm, integer params (e.g. GPTQ qweight/qzeros) are left |
| 1496 | # as torch.empty() by default, giving non-deterministic values |
| 1497 | # across processes. Zero them for reproducibility. |
| 1498 | param.zero_() |
| 1499 | return |
| 1500 | |
| 1501 | if current_platform.is_tpu(): |
| 1502 | generator = torch.Generator(device="cpu") |
| 1503 | generator.manual_seed(seed) |
| 1504 | # Note: The param.uniform_ function cannot be used in this |
| 1505 | # context because it demands more TPU HBM than directly copying |
| 1506 | # from a CPU tensor. |
| 1507 | # Note: We avoid using torch.rank_like as it doesn't currently |
| 1508 | # support the generator argument. |
| 1509 | param.copy_( |
| 1510 | (high - low) |
| 1511 | * torch.rand( |
| 1512 | param.shape, |
| 1513 | generator=generator, |
| 1514 | dtype=param.dtype, |
| 1515 | layout=param.layout, |
| 1516 | requires_grad=param.requires_grad, |
| 1517 | device="cpu", |
| 1518 | ) |
| 1519 | + low |
| 1520 | ) |
| 1521 | torch._sync(param) |
| 1522 | return |
| 1523 | |
| 1524 | generator = torch.Generator(device=param.data.device) |
| 1525 | generator.manual_seed(seed) |
| 1526 | if torch.finfo(param.data.dtype).bits < 16: |
| 1527 | # uniform_ doesn't support < 16-bit datatypes (FP8) |
| 1528 | tmp_param = torch.empty_like(param, dtype=torch.float16) |
| 1529 | tmp_param.uniform_(low, high, generator=generator) |
| 1530 | param.copy_(tmp_param) |
| 1531 | else: |
| 1532 | param.uniform_(low, high, generator=generator) |
| 1533 | |
| 1534 | |
| 1535 | def maybe_remap_kv_scale_name(name: str, params_dict: dict) -> str | None: |
| 1536 | """Remap the name of FP8 k/v_scale parameters. |
| 1537 | |
| 1538 | This function handles the remapping of FP8 k/v_scale parameter names. |
| 1539 | It detects if the given name ends with a suffix and attempts to remap |
| 1540 | it to the expected name format in the model. If the remapped name is not |
| 1541 | found in the params_dict, a warning is printed and None is returned. |
| 1542 | |
| 1543 | Args: |
| 1544 | name (str): The original loaded checkpoint parameter name. |
| 1545 | params_dict (dict): Dictionary containing the model's named parameters. |
| 1546 | |
| 1547 | Returns: |
| 1548 | str: The remapped parameter name if successful, or the original name |
| 1549 | if no remapping is needed. |
| 1550 | None: If the remapped name is not found in params_dict. |
| 1551 | |
| 1552 | """ |
| 1553 | # Already in vLLM's expected form (e.g. weights pre-renamed by a |
| 1554 | # `WeightsMapper` from the quant config). Skip the regex remap, which |
| 1555 | # would otherwise double-apply the `.attn` prefix and drop the weight. |
| 1556 | if name in params_dict: |
| 1557 | return name |
| 1558 | if name.endswith(".kv_scale"): |
| 1559 | logger.warning_once( |
| 1560 | "DEPRECATED. Found kv_scale in the checkpoint. " |
| 1561 | "This format is deprecated in favor of separate k_scale and " |
| 1562 | "v_scale tensors and will be removed in a future release. " |
| 1563 | "Functionally, we will remap kv_scale to k_scale and duplicate " |
| 1564 | "k_scale to v_scale" |
| 1565 | ) |
| 1566 | # NOTE: we remap the deprecated kv_scale to k_scale |
| 1567 | remapped_name = name.replace(".kv_scale", ".attn.k_scale") |
| 1568 | if remapped_name not in params_dict: |
| 1569 | logger.warning_once( |
| 1570 | "Found kv_scale in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). kv_scale is not loaded.", # noqa: E501 |
| 1571 | name, |
| 1572 | remapped_name, |
| 1573 | ) |
| 1574 | return None |
| 1575 | return remapped_name |
| 1576 | |
| 1577 | if any("mla_attn" in key for key in params_dict): |
| 1578 | attn_str = "mla_attn.mla_attn" |
| 1579 | logger.debug_once( |
| 1580 | f"Found mla_attn with k_scale and v_scale in " |
| 1581 | f"the checkpoint, using {attn_str} as attn_str" |
| 1582 | ) |
| 1583 | else: |
| 1584 | attn_str = "attn" |
| 1585 | # Define scale name mapping patterns in order of precedence |
| 1586 | scale_mapping_patterns = [ |
| 1587 | # ModelOpt format: .self_attn.{k,v}_proj.{k,v}_scale -> |
| 1588 | # .self_attn.attn.{k,v}_scale |
| 1589 | ( |
| 1590 | r"\.self_attn\.([kv])_proj\.([kv])_scale$", |
| 1591 | rf".self_attn.{attn_str}.\2_scale", |
| 1592 | ), |
| 1593 | # QKV proj format: .self_attn.qkv_proj.{k,v}_scale -> |
| 1594 | # .self_attn.attn.{k,v}_scale |
| 1595 | (r"\.self_attn\.qkv_proj\.([kv])_scale$", r".self_attn.attn.\1_scale"), |
| 1596 | # Qwen3 MoE format: .self_attn.qkqkv_proj.{k,v}_scale -> |
| 1597 | # .self_attn.attn.{k,v}_scale |
| 1598 | (r"\.self_attn\.qkqkv_proj\.([kv])_scale$", r".self_attn.attn.\1_scale"), |
| 1599 | # NemotronH format: .mixer.{k,v}_proj.{k,v}_scale -> |
| 1600 | # .mixer.attn.{k,v}_scale |
| 1601 | (r"\.mixer\.[kv]_proj\.([kv])_scale$", r".mixer.attn.\1_scale"), |
| 1602 | # HYV3 format: .self_attn.q.scale -> .self_attn.attn.q_scale |
| 1603 | (r"\.self_attn\.q\.scale$", r".self_attn.attn.q_scale"), |
| 1604 | # HYV3 format: .self_attn.{k,v}_cache.scale -> |
| 1605 | # .self_attn.attn.{k,v}_scale |
| 1606 | (r"\.self_attn\.([kv])_cache\.scale$", r".self_attn.attn.\1_scale"), |
| 1607 | # Default format: .{k,v}_scale -> .attn.{k,v}_scale |
| 1608 | (r"\.([qkv])_scale$", r".attn.\1_scale"), |
| 1609 | (r"\.([qkv])_zero_point$", r".attn.\1_zero_point"), |
| 1610 | ] |
| 1611 | |
| 1612 | # Check if name ends with k_scale or v_scale |
| 1613 | if name.endswith( |
| 1614 | ( |
| 1615 | ".k_scale", |
| 1616 | ".v_scale", |
| 1617 | ".q_scale", |
| 1618 | ".k_zero_point", |
| 1619 | ".v_zero_point", |
| 1620 | ".q_zero_point", |
| 1621 | ".q.scale", |
| 1622 | ".k_cache.scale", |
| 1623 | ".v_cache.scale", |
| 1624 | ) |
| 1625 | ): |
| 1626 | import regex as re |
| 1627 | |
| 1628 | for pattern, replacement in scale_mapping_patterns: |
| 1629 | if re.search(pattern, name): |
| 1630 | remapped_name = re.sub(pattern, replacement, name) |
| 1631 | if remapped_name not in params_dict: |
| 1632 | scale_type = name.split(".")[-1] |
| 1633 | logger.warning_once( |
| 1634 | "Found %s in the checkpoint (e.g. %s), but not found the expected name in the model (e.g. %s). %s is not loaded.", # noqa: E501 |
| 1635 | scale_type, |
| 1636 | name, |
| 1637 | remapped_name, |
| 1638 | scale_type, |
| 1639 | ) |
| 1640 | return None |
| 1641 | return remapped_name |
| 1642 | |
| 1643 | # If there were no matches, return the untouched param name |
| 1644 | return name |
| 1645 | |
| 1646 | |
| 1647 | def maybe_remap_moe_expert_param_name( |
| 1648 | name: str, |
| 1649 | params_dict: dict[str, torch.nn.Parameter], |
| 1650 | ) -> str: |
| 1651 | """Remap MoE expert parameter names to account for routed_experts hierarchy. |
| 1652 | |
| 1653 | This handles the transition from the old FusedMoE structure where weights |
| 1654 | were directly in the experts module, to the new MoERunner → RoutedExperts |
| 1655 | structure. |
| 1656 | |
| 1657 | Checkpoint weights have names like: |
| 1658 | layers.0.mlp.experts.w13_weight |
| 1659 | layers.0.feed_forward.experts.w2_input_scale |
| 1660 | But actual parameters are now: |
| 1661 | layers.0.mlp.experts.routed_experts.w13_weight |
| 1662 | layers.0.feed_forward.experts.routed_experts.w2_input_scale |
| 1663 | |
| 1664 | This function inserts 'routed_experts.' into the path when needed. |
| 1665 | |
| 1666 | Args: |
| 1667 | name: Parameter name from checkpoint |
| 1668 | params_dict: Dictionary of model parameters (from named_parameters()) |
| 1669 | |
| 1670 | Returns: |
| 1671 | Remapped parameter name if routed_experts hierarchy exists, |
| 1672 | otherwise the original name |
| 1673 | |
| 1674 | """ |
| 1675 | # Only remap if this looks like an expert parameter |
| 1676 | if ".experts." not in name: |
| 1677 | return name |
| 1678 | |
| 1679 | # Skip if already has routed_experts |
| 1680 | if ".experts.routed_experts." in name: |
| 1681 | return name |
| 1682 | |
| 1683 | # Expert parameter patterns to check |
| 1684 | expert_param_suffixes = [ |
| 1685 | "w13_weight", |
| 1686 | "w2_weight", |
| 1687 | "w13_weight_scale", |
| 1688 | "w2_weight_scale", |
| 1689 | "w13_input_scale", |
| 1690 | "w2_input_scale", |
| 1691 | "w13_bias", |
| 1692 | "w2_bias", |
| 1693 | "w13_scale", |
| 1694 | "w2_scale", |
| 1695 | "w13_qweight", |
| 1696 | "w2_qweight", |
| 1697 | "w13_qzeros", |
| 1698 | "w2_qzeros", |
| 1699 | "w13_weight_shape", |
| 1700 | "w2_weight_shape", |
| 1701 | ] |
| 1702 | |
| 1703 | # Check if this is an expert weight parameter |
| 1704 | is_expert_param = any( |
| 1705 | f".{suffix}" in name or name.endswith(suffix) |
| 1706 | for suffix in expert_param_suffixes |
| 1707 | ) |
| 1708 | |
| 1709 | if not is_expert_param: |
| 1710 | return name |
| 1711 | |
| 1712 | # Try inserting routed_experts after .experts. |
| 1713 | new_name = name.replace(".experts.", ".experts.routed_experts.", 1) |
| 1714 | |
| 1715 | # Only use the new name if it exists in the model |
| 1716 | if new_name in params_dict: |
| 1717 | return new_name |
| 1718 | |
| 1719 | # Otherwise return original name (old checkpoint format or different structure) |
| 1720 | return name |
| 1721 | |
| 1722 | |
| 1723 | def remap_moe_expert_weights( |
| 1724 | weights: Iterable[tuple[str, torch.Tensor]], |
| 1725 | params_dict: dict[str, torch.nn.Parameter], |
| 1726 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 1727 | """Remap MoE expert parameter names for backward compatibility. |
| 1728 | |
| 1729 | This allows models with custom weight loading to automatically handle both old |
| 1730 | and new checkpoint formats without needing model-specific remapping code. |
| 1731 | |
| 1732 | Usage: |
| 1733 | params_dict = dict(model.named_parameters()) |
| 1734 | for name, weight in remap_moe_expert_weights(weights, params_dict): |
| 1735 | # name is automatically remapped if needed |
| 1736 | param = params_dict[name] |
| 1737 | ... |
| 1738 | |
| 1739 | Args: |
| 1740 | weights: Iterator of (name, tensor) tuples from checkpoint |
| 1741 | params_dict: Dictionary of model parameters (from named_parameters()) |
| 1742 | |
| 1743 | Yields: |
| 1744 | (remapped_name, tensor) tuples |
| 1745 | |
| 1746 | """ |
| 1747 | for name, weight in weights: |
| 1748 | remapped_name = maybe_remap_moe_expert_param_name(name, params_dict) |
| 1749 | yield (remapped_name, weight) |