| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # unit test for `examples/features/torchrun/torchrun_example_offline.py` |
| 5 | import os |
| 6 | import random |
| 7 | |
| 8 | import torch |
| 9 | import torch.distributed as dist |
| 10 | |
| 11 | import vllm.envs as envs |
| 12 | from vllm import LLM, SamplingParams |
| 13 | from vllm.distributed.parallel_state import get_tp_group, get_world_group |
| 14 | |
| 15 | # By default, let PyTorch choose the WORLD backend for the current device |
| 16 | # type (legacy lazy-init path). When VLLM_DISTRIBUTED_USE_SPLIT_GROUP=1, |
| 17 | # use the explicit eager-init pattern required by `split_group` (mixed |
| 18 | # cpu:gloo,cuda:nccl backend + device_id binding). |
| 19 | if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP: |
| 20 | local_rank = int(os.environ["LOCAL_RANK"]) |
| 21 | torch.accelerator.set_device_index(local_rank) |
| 22 | dist.init_process_group( |
| 23 | backend="cpu:gloo,cuda:nccl", |
| 24 | device_id=torch.device(f"cuda:{local_rank}"), |
| 25 | ) |
| 26 | else: |
| 27 | dist.init_process_group() |
| 28 | |
| 29 | # Create prompts |
| 30 | prompts = [ |
| 31 | "Hello, my name is", |
| 32 | "The president of the United States is", |
| 33 | "The capital of France is", |
| 34 | "The future of AI is", |
| 35 | ] * 10 |
| 36 | dp_size = int(os.getenv("DP_SIZE", "1")) |
| 37 | dp_rank = int(os.getenv("DP_RANK", "0")) |
| 38 | |
| 39 | if dp_size > 1: |
| 40 | # distribute the prompts across the data parallel ranks |
| 41 | prompts = [prompt for idx, prompt in enumerate(prompts) if idx % dp_size == dp_rank] |
| 42 | |
| 43 | sampling_params = SamplingParams(temperature=0.8, top_p=0.95) |
| 44 | |
| 45 | # set different `gpu_memory_utilization` for different ranks, |
| 46 | # to test if all ranks agree on the same kv cache configuration. |
| 47 | llm = LLM( |
| 48 | model="microsoft/Phi-mini-MoE-instruct", |
| 49 | tensor_parallel_size=int(os.getenv("TP_SIZE", "1")), |
| 50 | pipeline_parallel_size=int(os.getenv("PP_SIZE", "1")), |
| 51 | enable_expert_parallel=int(os.getenv("ENABLE_EP", "0")) == 1, |
| 52 | distributed_executor_backend="external_launcher", |
| 53 | gpu_memory_utilization=random.uniform(0.8, 0.92), |
| 54 | seed=0, |
| 55 | max_model_len=1024, |
| 56 | max_num_seqs=16, |
| 57 | ) |
| 58 | |
| 59 | outputs = llm.generate(prompts, sampling_params) |
| 60 | |
| 61 | group = get_world_group() if dp_size == 1 else get_tp_group() |
| 62 | cpu_group = group.cpu_group |
| 63 | group_rank = dist.get_rank(group=cpu_group) |
| 64 | |
| 65 | |
| 66 | def test_consistent_across_ranks(obj): |
| 67 | if group_rank == 0: |
| 68 | dist.broadcast_object_list([obj], src=group.ranks[0], group=cpu_group) |
| 69 | else: |
| 70 | container = [None] |
| 71 | dist.broadcast_object_list(container, src=group.ranks[0], group=cpu_group) |
| 72 | assert container[0] == obj |
| 73 | |
| 74 | |
| 75 | test_consistent_across_ranks(llm.llm_engine.vllm_config.cache_config.num_cpu_blocks) |
| 76 | test_consistent_across_ranks(llm.llm_engine.vllm_config.cache_config.num_gpu_blocks) |
| 77 | |
| 78 | # make sure we can access the model parameters from the calling process |
| 79 | # of the `LLM` instance. |
| 80 | params = list( |
| 81 | llm.llm_engine.model_executor.driver_worker.worker.model_runner.model.parameters() |
| 82 | ) |
| 83 | test_consistent_across_ranks(len(params)) |
| 84 | |
| 85 | # all ranks should have the same outputs |
| 86 | for output in outputs: |
| 87 | prompt = output.prompt |
| 88 | generated_text = output.outputs[0].text |
| 89 | test_consistent_across_ranks(prompt) |
| 90 | test_consistent_across_ranks(generated_text) |
| 91 | print(f"Rank {group_rank}, Prompt: {prompt!r}, Generated text: {generated_text!r}") |