| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import collections |
| 5 | import glob |
| 6 | import os |
| 7 | import time |
| 8 | from collections.abc import Generator |
| 9 | from copy import copy |
| 10 | from typing import Any |
| 11 | |
| 12 | import torch |
| 13 | from torch import nn |
| 14 | |
| 15 | from vllm.config import ModelConfig |
| 16 | from vllm.config.load import LoadConfig |
| 17 | from vllm.logger import init_logger |
| 18 | from vllm.model_executor.model_loader.base_loader import BaseModelLoader |
| 19 | from vllm.model_executor.model_loader.weight_utils import ( |
| 20 | download_weights_from_hf, |
| 21 | runai_safetensors_weights_iterator, |
| 22 | ) |
| 23 | from vllm.transformers_utils.s3_utils import glob as s3_glob |
| 24 | from vllm.transformers_utils.utils import is_s3 |
| 25 | |
| 26 | logger = init_logger(__name__) |
| 27 | |
| 28 | |
| 29 | class ShardedStateLoader(BaseModelLoader): |
| 30 | """Model loader that directly loads each worker's model state dict, which |
| 31 | enables a fast load path for large tensor-parallel models where each worker |
| 32 | only needs to read its own shard rather than the entire checkpoint. See |
| 33 | `examples/features/sharded_state/save_sharded_state_offline.py` for creating |
| 34 | a sharded checkpoint. |
| 35 | """ |
| 36 | |
| 37 | DEFAULT_PATTERN = "model-rank-{rank}-part-{part}.safetensors" |
| 38 | |
| 39 | def __init__(self, load_config: LoadConfig): |
| 40 | super().__init__(load_config) |
| 41 | |
| 42 | extra_config = ( |
| 43 | {} |
| 44 | if load_config.model_loader_extra_config is None |
| 45 | else copy(load_config.model_loader_extra_config) |
| 46 | ) |
| 47 | self.pattern = extra_config.pop("pattern", self.DEFAULT_PATTERN) |
| 48 | if extra_config: |
| 49 | raise ValueError( |
| 50 | f"Unexpected extra config keys for load format " |
| 51 | f"{load_config.load_format}: " |
| 52 | f"{load_config.model_loader_extra_config.keys()}" |
| 53 | ) |
| 54 | |
| 55 | @staticmethod |
| 56 | def _filter_subtensors( |
| 57 | tensors: dict[str, torch.Tensor], |
| 58 | ) -> dict[str, torch.Tensor]: |
| 59 | """Filter out all tensors that share the same memory or a subset of the |
| 60 | memory of another tensor. |
| 61 | """ |
| 62 | same_storage_groups: dict[Any, list[tuple[str, torch.Tensor]]] = ( |
| 63 | collections.defaultdict(list) |
| 64 | ) |
| 65 | for key, tensor in tensors.items(): |
| 66 | if tensor.numel(): |
| 67 | ptr = tensor.untyped_storage().data_ptr() |
| 68 | same_storage_groups[tensor.device, ptr].append((key, tensor)) |
| 69 | |
| 70 | def get_end_ptr(tensor: torch.Tensor) -> int: |
| 71 | return tensor.view(-1)[-1].data_ptr() + tensor.element_size() |
| 72 | |
| 73 | result: dict[str, torch.Tensor] = {} |
| 74 | for group in same_storage_groups.values(): |
| 75 | for k, t in group: |
| 76 | a, b = t.data_ptr(), get_end_ptr(t) |
| 77 | for k2, t2 in group: |
| 78 | if not t2.is_contiguous(): |
| 79 | continue |
| 80 | a2, b2 = t2.data_ptr(), get_end_ptr(t2) |
| 81 | if a < a2 or b2 < b: |
| 82 | continue |
| 83 | if a2 < a or b < b2 or not t.is_contiguous(): |
| 84 | break # t2 covers strictly more memory than t. |
| 85 | if k2 < k: |
| 86 | # Same tensors, keep the one with the smaller key. |
| 87 | break |
| 88 | else: |
| 89 | result[k] = t |
| 90 | return result |
| 91 | |
| 92 | def _prepare_weights(self, model_name_or_path: str, revision: str | None): |
| 93 | if is_s3(model_name_or_path) or os.path.isdir(model_name_or_path): |
| 94 | return model_name_or_path |
| 95 | else: |
| 96 | allow_patterns = ["*.safetensors"] |
| 97 | return download_weights_from_hf( |
| 98 | model_name_or_path, |
| 99 | self.load_config.download_dir, |
| 100 | allow_patterns, |
| 101 | revision, |
| 102 | ignore_patterns=self.load_config.ignore_patterns, |
| 103 | ) |
| 104 | |
| 105 | def download_model(self, model_config: ModelConfig) -> None: |
| 106 | self._prepare_weights(model_config.model, model_config.revision) |
| 107 | |
| 108 | def load_weights(self, model: nn.Module, model_config: ModelConfig) -> None: |
| 109 | from vllm.distributed import get_tensor_model_parallel_rank |
| 110 | |
| 111 | model_weights = model_config.model |
| 112 | if model_weights_override := model_config.model_weights: |
| 113 | model_weights = model_weights_override |
| 114 | local_model_path = model_weights |
| 115 | |
| 116 | rank = get_tensor_model_parallel_rank() |
| 117 | pattern = os.path.join( |
| 118 | local_model_path, |
| 119 | self.pattern.format(rank=rank, part="*"), |
| 120 | ) |
| 121 | |
| 122 | filepaths = [] |
| 123 | if is_s3(local_model_path): |
| 124 | file_pattern = f"*{self.pattern.format(rank=rank, part='*')}" |
| 125 | filepaths = s3_glob(path=local_model_path, allow_pattern=[file_pattern]) |
| 126 | else: |
| 127 | filepaths = glob.glob(pattern) |
| 128 | if not filepaths: |
| 129 | # TODO: support un-sharded checkpoints too |
| 130 | raise ValueError( |
| 131 | f"Could not find checkpoint files '{pattern}', only " |
| 132 | f"pre-sharded checkpoints are currently supported!" |
| 133 | ) |
| 134 | state_dict = self._filter_subtensors(model.state_dict()) |
| 135 | counter_before_loading_weights = time.perf_counter() |
| 136 | for key, tensor in self.iterate_over_files(filepaths): |
| 137 | # If loading with LoRA enabled, additional padding may |
| 138 | # be added to certain parameters. We only load into a |
| 139 | # narrowed view of the parameter data. |
| 140 | param_data = state_dict[key].data |
| 141 | param_shape = state_dict[key].shape |
| 142 | for dim, size in enumerate(tensor.shape): |
| 143 | if size < param_shape[dim]: |
| 144 | param_data = param_data.narrow(dim, 0, size) |
| 145 | if tensor.shape != param_shape: |
| 146 | logger.warning( |
| 147 | "loading tensor of shape %s into parameter '%s' of shape %s", |
| 148 | tensor.shape, |
| 149 | key, |
| 150 | param_shape, |
| 151 | ) |
| 152 | param_data.copy_(tensor) |
| 153 | state_dict.pop(key) |
| 154 | counter_after_loading_weights = time.perf_counter() |
| 155 | logger.info_once( |
| 156 | "Loading weights took %.2f seconds", |
| 157 | counter_after_loading_weights - counter_before_loading_weights, |
| 158 | ) |
| 159 | if state_dict: |
| 160 | raise ValueError(f"Missing keys {tuple(state_dict)} in loaded state!") |
| 161 | |
| 162 | def iterate_over_files( |
| 163 | self, paths |
| 164 | ) -> Generator[tuple[str, torch.Tensor], None, None]: |
| 165 | if self.load_config.load_format == "runai_streamer_sharded": |
| 166 | yield from runai_safetensors_weights_iterator(paths, True) |
| 167 | else: |
| 168 | from safetensors.torch import safe_open |
| 169 | |
| 170 | for path in paths: |
| 171 | with safe_open(path, framework="pt") as f: |
| 172 | for key in f.keys(): # noqa: SIM118 |
| 173 | tensor = f.get_tensor(key) |
| 174 | yield key, tensor |
| 175 | |
| 176 | @staticmethod |
| 177 | def save_model( |
| 178 | model: torch.nn.Module, |
| 179 | path: str, |
| 180 | pattern: str | None = None, |
| 181 | max_size: int | None = None, |
| 182 | ) -> None: |
| 183 | from safetensors.torch import save_file |
| 184 | |
| 185 | from vllm.distributed import get_tensor_model_parallel_rank |
| 186 | |
| 187 | if pattern is None: |
| 188 | pattern = ShardedStateLoader.DEFAULT_PATTERN |
| 189 | rank = get_tensor_model_parallel_rank() |
| 190 | part_idx = 0 |
| 191 | total_size = 0 |
| 192 | state_dict = ShardedStateLoader._filter_subtensors(model.state_dict()) |
| 193 | state_dict_part: dict[str, torch.Tensor] = {} |
| 194 | for key, tensor in state_dict.items(): |
| 195 | param_size = tensor.nelement() * tensor.element_size() |
| 196 | if max_size is not None and total_size + param_size > max_size: |
| 197 | filename = pattern.format(rank=rank, part=part_idx) |
| 198 | save_file( |
| 199 | state_dict_part, |
| 200 | os.path.join(path, filename), |
| 201 | ) |
| 202 | part_idx += 1 |
| 203 | total_size = 0 |
| 204 | state_dict_part = {} |
| 205 | state_dict_part[key] = tensor |
| 206 | total_size += param_size |
| 207 | if len(state_dict_part) > 0: |
| 208 | filename = pattern.format(rank=rank, part=part_idx) |
| 209 | save_file( |
| 210 | state_dict_part, |
| 211 | os.path.join(path, filename), |
| 212 | ) |