| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import dataclasses |
| 4 | import os |
| 5 | import traceback |
| 6 | from collections.abc import Callable |
| 7 | from typing import Any, Concatenate |
| 8 | |
| 9 | import torch |
| 10 | from torch.multiprocessing import spawn # pyright: ignore[reportPrivateImportUsage] |
| 11 | from typing_extensions import ParamSpec |
| 12 | |
| 13 | import vllm.envs as envs |
| 14 | from vllm.config import VllmConfig, set_current_vllm_config |
| 15 | from vllm.distributed import ( |
| 16 | cleanup_dist_env_and_memory, |
| 17 | init_distributed_environment, |
| 18 | initialize_model_parallel, |
| 19 | ) |
| 20 | from vllm.utils.network_utils import get_open_port |
| 21 | |
| 22 | ## Parallel Processes Utils |
| 23 | |
| 24 | P = ParamSpec("P") |
| 25 | |
| 26 | |
| 27 | @dataclasses.dataclass |
| 28 | class ProcessGroupInfo: |
| 29 | world_size: int |
| 30 | world_local_size: int |
| 31 | rank: int |
| 32 | node_rank: int |
| 33 | local_rank: int |
| 34 | device: torch.device |
| 35 | |
| 36 | |
| 37 | def _set_vllm_config( |
| 38 | vllm_config: VllmConfig, world_size: int, rank: int, local_rank: int |
| 39 | ): |
| 40 | import tempfile |
| 41 | |
| 42 | temp_file = tempfile.mkstemp()[1] |
| 43 | |
| 44 | # When DP is enabled, processes are organized as: |
| 45 | # rank = dp_rank * tp_pp_world_size + tp_pp_rank |
| 46 | tp_pp_world_size = vllm_config.parallel_config.world_size |
| 47 | vllm_config.parallel_config.data_parallel_rank = rank // tp_pp_world_size |
| 48 | tp_pp_rank = rank % tp_pp_world_size |
| 49 | vllm_config.parallel_config.rank = tp_pp_rank |
| 50 | |
| 51 | with set_current_vllm_config(vllm_config): |
| 52 | init_distributed_environment( |
| 53 | world_size=tp_pp_world_size, |
| 54 | rank=tp_pp_rank, |
| 55 | distributed_init_method=f"file://{temp_file}", |
| 56 | local_rank=local_rank, |
| 57 | backend="nccl", |
| 58 | ) |
| 59 | |
| 60 | initialize_model_parallel( |
| 61 | tensor_model_parallel_size=vllm_config.parallel_config.tensor_parallel_size, |
| 62 | pipeline_model_parallel_size=vllm_config.parallel_config.pipeline_parallel_size, |
| 63 | ) |
| 64 | if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP: |
| 65 | cpu_group = torch.distributed.split_group( |
| 66 | split_ranks=[list(range(world_size))], |
| 67 | group_desc="moe_test_cpu", |
| 68 | ) |
| 69 | else: |
| 70 | cpu_group = torch.distributed.new_group( |
| 71 | list(range(world_size)), backend="gloo" |
| 72 | ) |
| 73 | return cpu_group |
| 74 | |
| 75 | |
| 76 | def _worker_parallel_launch( |
| 77 | local_rank: int, |
| 78 | world_size: int, |
| 79 | world_local_size: int, |
| 80 | node_rank: int, |
| 81 | init_method: str, |
| 82 | worker: Callable[..., None], |
| 83 | vllm_config: VllmConfig | None, |
| 84 | env_dict: dict | None, |
| 85 | worker_kwargs: dict[str, Any], |
| 86 | *args: Any, |
| 87 | ) -> None: |
| 88 | rank = node_rank * world_local_size + local_rank |
| 89 | device = torch.device("cuda", local_rank) |
| 90 | torch.accelerator.set_device_index(device) |
| 91 | torch.set_default_device(device) |
| 92 | torch.distributed.init_process_group( |
| 93 | backend="cpu:gloo,cuda:nccl", |
| 94 | init_method=init_method, |
| 95 | rank=rank, |
| 96 | world_size=world_size, |
| 97 | device_id=device, |
| 98 | ) |
| 99 | barrier = torch.tensor([rank], device=device) |
| 100 | torch.distributed.all_reduce(barrier) |
| 101 | |
| 102 | if env_dict is not None: |
| 103 | os.environ.update(env_dict) |
| 104 | |
| 105 | cpu_group = None |
| 106 | if vllm_config is not None: |
| 107 | cpu_group = _set_vllm_config(vllm_config, world_size, rank, local_rank) |
| 108 | |
| 109 | def _run_worker(): |
| 110 | worker( |
| 111 | ProcessGroupInfo( |
| 112 | world_size=world_size, |
| 113 | world_local_size=world_local_size, |
| 114 | rank=rank, |
| 115 | node_rank=node_rank, |
| 116 | local_rank=local_rank, |
| 117 | device=device, |
| 118 | ), |
| 119 | vllm_config, |
| 120 | cpu_group, |
| 121 | *args, |
| 122 | **worker_kwargs, |
| 123 | ) |
| 124 | |
| 125 | try: |
| 126 | if vllm_config is not None: |
| 127 | with set_current_vllm_config(vllm_config): |
| 128 | _run_worker() |
| 129 | else: |
| 130 | _run_worker() |
| 131 | except Exception as ex: |
| 132 | print(ex) |
| 133 | traceback.print_exc() |
| 134 | raise |
| 135 | finally: |
| 136 | torch.accelerator.synchronize() |
| 137 | if vllm_config is not None: |
| 138 | cleanup_dist_env_and_memory() |
| 139 | else: |
| 140 | torch.distributed.destroy_process_group() |
| 141 | |
| 142 | |
| 143 | def parallel_launch_with_config( |
| 144 | world_size: int, |
| 145 | worker: Callable[Concatenate[ProcessGroupInfo, VllmConfig, Any, P], None], |
| 146 | vllm_config: VllmConfig, |
| 147 | env_dict: dict[Any, Any] | None, |
| 148 | *args: P.args, |
| 149 | **kwargs: P.kwargs, |
| 150 | ) -> None: |
| 151 | spawn( |
| 152 | _worker_parallel_launch, |
| 153 | args=( |
| 154 | world_size, |
| 155 | world_size, |
| 156 | 0, |
| 157 | f"tcp://{os.getenv('LOCALHOST', 'localhost')}:{get_open_port()}", |
| 158 | worker, |
| 159 | vllm_config, |
| 160 | env_dict, |
| 161 | kwargs, |
| 162 | ) |
| 163 | + args, |
| 164 | nprocs=world_size, |
| 165 | join=True, |
| 166 | ) |