open-compass
[Bug] 使用vllm加速时无法启动多卡
Prerequisite
- [X] I have searched Issues and Discussions but cannot get the expected help.
- [X] The bug has not been fixed in the latest version.
Type
I'm evaluating with the officially supported tasks/models/datasets.
Environment
{'CUDA available': True, 'CUDA_HOME': '/usr/local/cuda', 'GCC': 'x86_64-linux-gnu-gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0', 'GPU 0,1,2,3,4,5,6,7': 'NVIDIA A800-SXM4-80GB', 'MMEngine': '0.10.5', 'MUSA available': False, 'NVCC': 'Cuda compilation tools, release 12.2, V12.2.140', 'OpenCV': '4.10.0', 'PyTorch': '2.5.1+cu124', 'PyTorch compiling details': 'PyTorch built with:\n' ' - GCC 9.3\n' ' - C++ Version: 201703\n' ' - Intel(R) oneAPI Math Kernel Library Version ' '2024.2-Product Build 20240605 for Intel(R) 64 ' 'architecture applications\n' ' - Intel(R) MKL-DNN v3.5.3 (Git Hash ' '66f0cb9eb66affd2da3bf5f8d897376f04aae6af)\n' ' - OpenMP 201511 (a.k.a. OpenMP 4.5)\n' ' - LAPACK is enabled (usually provided by ' 'MKL)\n' ' - NNPACK is enabled\n' ' - CPU capability usage: AVX512\n' ' - CUDA Runtime 12.4\n' ' - NVCC architecture flags: ' '-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_90,code=sm_90\n' ' - CuDNN 90.1\n' ' - Magma 2.6.1\n' ' - Build settings: BLAS_INFO=mkl, ' 'BUILD_TYPE=Release, CUDA_VERSION=12.4, ' 'CUDNN_VERSION=9.1.0, ' 'CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, ' 'CXX_FLAGS= -D_GLIBCXX_USE_CXX11_ABI=0 ' '-fabi-version=11 -fvisibility-inlines-hidden ' '-DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO ' '-DLIBKINETO_NOROCTRACER -DLIBKINETO_NOXPUPTI=ON ' '-DUSE_FBGEMM -DUSE_PYTORCH_QNNPACK ' '-DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE ' '-O2 -fPIC -Wall -Wextra -Werror=return-type ' '-Werror=non-virtual-dtor -Werror=bool-operation ' '-Wnarrowing -Wno-missing-field-initializers ' '-Wno-type-limits -Wno-array-bounds ' '-Wno-unknown-pragmas -Wno-unused-parameter ' '-Wno-strict-overflow -Wno-strict-aliasing ' '-Wno-stringop-overflow -Wsuggest-override ' '-Wno-psabi -Wno-error=old-style-cast ' '-Wno-missing-braces -fdiagnostics-color=always ' '-faligned-new -Wno-unused-but-set-variable ' '-Wno-maybe-uninitialized -fno-math-errno ' '-fno-trapping-math -Werror=format ' '-Wno-stringop-overflow, LAPACK_INFO=mkl, ' 'PERF_WITH_AVX=1, PERF_WITH_AVX2=1, ' 'TORCH_VERSION=2.5.1, USE_CUDA=ON, USE_CUDNN=ON, ' 'USE_CUSPARSELT=1, USE_EXCEPTION_PTR=1, ' 'USE_GFLAGS=OFF, USE_GLOG=OFF, USE_GLOO=ON, ' 'USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, ' 'USE_NCCL=1, USE_NNPACK=ON, USE_OPENMP=ON, ' 'USE_ROCM=OFF, USE_ROCM_KERNEL_ASSERT=OFF, \n', 'Python': '3.10.12 (main, Jun 11 2023, 05:26:28) [GCC 11.4.0]', 'TorchVision': '0.20.1+cu124', 'lmdeploy': "not installed:No module named 'lmdeploy'", 'numpy_random_seed': 2147483648, 'opencompass': '0.3.9+b26c76a', 'sys.platform': 'linux', 'transformers': '4.46.1'}
Reproduces the problem - code/configuration sample
【模型配置】使用两张卡 1 from opencompass.models import HuggingFacewithChatTemplate 2 3 models = [ 4 dict( 5 type=HuggingFacewithChatTemplate, 6 abbr='qwen2.5-1.5b-instruct-hf', 7 #path='Qwen/Qwen2.5-1.5B-Instruct', 8 path='/workspace/models/hf_models/Qwen2.5-1.5B-Instruct', 9 max_out_len=4096, 10 batch_size=8, 11 run_cfg=dict(num_gpus=2), 12 ) 13 ]
Reproduces the problem - command or script
python run.py --datasets ceval_gen --models hf_qwen2_5_1_5b_instruct --debug -a vllm
Reproduces the problem - error message
01/08 11:34:07 - OpenCompass - INFO - Loading hf_qwen2_5_1_5b_instruct: configs/models/hf_qwen2_5_1_5b_instruct.py 01/08 11:34:07 - OpenCompass - INFO - Transforming qwen2.5-1.5b-instruct-hf to vllm 01/08 11:34:07 - OpenCompass - INFO - Loading example: /usr/local/lib/python3.10/dist-packages/opencompass/configs/./summarizers/example.py 01/08 11:34:07 - OpenCompass - INFO - Current exp folder: outputs/default/20250108_113407 01/08 11:34:07 - OpenCompass - WARNING - SlurmRunner is not used, so the partition argument is ignored. 01/08 11:34:07 - OpenCompass - INFO - Partitioned into 2 tasks. 01/08 11:34:11 - OpenCompass - WARNING - Only use 2 GPUs for total 8 available GPUs in debug mode. 01/08 11:34:11 - OpenCompass - INFO - Task [qwen2.5-1.5b-instruct-vllm/ceval-computer_network_0,qwen2.5-1.5b-instruct-vllm/ceval-operating_system_0,qwen2.5-1.5b-instruct-vllm/ceval-computer_architecture_0,qwen2.5-1.5b-instruct-vllm/ceval-college_programming_0,qwen2.5-1.5b-instruct-vllm/ceval-college_physics_0,qwen2.5-1.5b-instruct-vllm/ceval-college_chemistry_0,qwen2.5-1.5b-instruct-vllm/ceval-advanced_mathematics_0,qwen2.5-1.5b-instruct-vllm/ceval-probability_and_statistics_0,qwen2.5-1.5b-instruct-vllm/ceval-discrete_mathematics,qwen2.5-1.5b-instruct-vllm/ceval-electrical_engineer_1,qwen2.5-1.5b-instruct-vllm/ceval-metrology_engineer_1,qwen2.5-1.5b-instruct-vllm/ceval-high_school_mathematics_1,qwen2.5-1.5b-instruct-vllm/ceval-high_school_physics_1,qwen2.5-1.5b-instruct-vllm/ceval-high_school_chemistry_1,qwen2.5-1.5b-instruct-vllm/ceval-high_school_biology_1,qwen2.5-1.5b-instruct-vllm/ceval-middle_school_mathematics_1,qwen2.5-1.5b-instruct-vllm/ceval-middle_school_biology_1,qwen2.5-1.5b-instruct-vllm/ceval-middle_school_physics_1,qwen2.5-1.5b-instruct-vllm/ceval-middle_school_chemistry_1,qwen2.5-1.5b-instruct-vllm/ceval-veterinary_medicine_1,qwen2.5-1.5b-instruct-vllm/ceval-college_economics_1,qwen2.5-1.5b-instruct-vllm/ceval-business_administration_1,qwen2.5-1.5b-instruct-vllm/ceval-marxism_1,qwen2.5-1.5b-instruct-vllm/ceval-mao_zedong_thought_1,qwen2.5-1.5b-instruct-vllm/ceval-education_science_1,qwen2.5-1.5b-instruct-vllm/ceval-teacher_qualification_1,qwen2.5-1.5b-instruct-vllm/ceval-high_school_politics_1,qwen2.5-1.5b-instruct-vllm/ceval-high_school_geography_1,qwen2.5-1.5b-instruct-vllm/ceval-middle_school_politics_1,qwen2.5-1.5b-instruct-vllm/ceval-modern_chinese_history_0,qwen2.5-1.5b-instruct-vllm/ceval-ideological_and_moral_cultivation_0,qwen2.5-1.5b-instruct-vllm/ceval-logic_0,qwen2.5-1.5b-instruct-vllm/ceval-law_0,qwen2.5-1.5b-instruct-vllm/ceval-chinese_language_and_literature_0,qwen2.5-1.5b-instruct-vllm/ceval-art_studies_0,qwen2.5-1.5b-instruct-vllm/ceval-professional_tour_guide_0,qwen2.5-1.5b-instruct-vllm/ceval-legal_professional_0,qwen2.5-1.5b-instruct-vllm/ceval-high_school_chinese_0,qwen2.5-1.5b-instruct-vllm/ceval-high_school_history_0,qwen2.5-1.5b-instruct-vllm/ceval-middle_school_history_0,qwen2.5-1.5b-instruct-vllm/ceval-civil_servant_0,qwen2.5-1.5b-instruct-vllm/ceval-sports_science_0,qwen2.5-1.5b-instruct-vllm/ceval-plant_protection_0,qwen2.5-1.5b-instruct-vllm/ceval-basic_medicine_0,qwen2.5-1.5b-instruct-vllm/ceval-clinical_medicine_0,qwen2.5-1.5b-instruct-vllm/ceval-urban_and_rural_planner_0,qwen2.5-1.5b-instruct-vllm/ceval-accountant_0,qwen2.5-1.5b-instruct-vllm/ceval-fire_engineer_0,qwen2.5-1.5b-instruct-vllm/ceval-environmental_impact_assessment_engineer_0,qwen2.5-1.5b-instruct-vllm/ceval-tax_accountant_0,qwen2.5-1.5b-instruct-vllm/ceval-physician_0] INFO 01-08 11:34:15 config.py:378] This model supports multiple tasks: {'embedding', 'generate'}. Defaulting to 'generate'. INFO 01-08 11:34:15 config.py:1048] Defaulting to use mp for distributed inference INFO 01-08 11:34:15 llm_engine.py:249] Initializing an LLM engine (vv1.3) with config: model='/workspace/models/hf_models/Qwen2.5-1.5B-Instruct', speculative_config=None, tokenizer='/workspace/models/hf_models/Qwen2.5-1.5B-Instruct', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=2, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=/workspace/models/hf_models/Qwen2.5-1.5B-Instruct, num_scheduler_steps=1, chunked_prefill_enabled=False multi_step_stream_outputs=True, enable_prefix_caching=False, use_async_output_proc=True, use_cached_outputs=False, chat_template_text_format=string, mm_processor_kwargs=None, pooler_config=None) INFO 01-08 11:34:16 custom_cache_manager.py:17] Setting Triton cache manager to: coca_vllm.triton_utils.custom_cache_manager:CustomCacheManager INFO 01-08 11:34:16 selector.py:135] Using Flash Attention backend. (VllmWorkerProcess pid=69399) INFO 01-08 11:34:16 selector.py:135] Using Flash Attention backend. (VllmWorkerProcess pid=69399) INFO 01-08 11:34:16 multiproc_worker_utils.py:215] Worker ready; awaiting tasks (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] Exception in worker VllmWorkerProcess while processing method init_device. (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] Traceback (most recent call last): (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] File "/usr/local/lib/python3.10/dist-packages/coca_vllm/executor/multiproc_worker_utils.py", line 223, in _run_worker_process (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] output = executor(*args, **kwargs) (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] File "/usr/local/lib/python3.10/dist-packages/coca_vllm/worker/worker.py", line 135, in init_device (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] torch.cuda.set_device(self.device) (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] File "/usr/local/lib/python3.10/dist-packages/torch/cuda/init.py", line 478, in set_device (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] torch._C._cuda_setDevice(device) (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] File "/usr/local/lib/python3.10/dist-packages/torch/cuda/init.py", line 305, in _lazy_init (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] raise RuntimeError( (VllmWorkerProcess pid=69399) ERROR 01-08 11:34:16 multiproc_worker_utils.py:229] RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method ^C(VllmWorkerProcess pid=69399) INFO 01-08 11:34:30 multiproc_worker_utils.py:240] Worker exiting
Other information
不使用vllm时可以多卡评测，使用vllm时报错：RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
export VLLM_WORKER_MULTIPROC_METHOD=spawn 请在使用多卡进行vllm评测前导入此方法
Where can I put this line of code cause I am using python run.py at the begining.
export VLLM_WORKER_MULTIPROC_METHOD=spawn 请在使用多卡进行vllm体育前导入此方法
我应该把这行代码放在哪里，因为我一开始使用的是 python run.py。
Import before running python run.py
CUDA_VISIBLE_DEVICES=0,1 opencompass --datasets demo_gsm8k_chat_gen --hf-type chat --hf-path internlm/internlm2_5-1_8b-chat --max-num-worker 2 没有设置max-num-worker,这个问题我也遇到
