[Bug] 在无gpu的机器上执行case，运行时报错数据集未注册（其实已经注册）

[Caeser-SONG]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version.

Type

I'm evaluating with the officially supported tasks/models/datasets.

Environment

1. conda python 10
2. cpu模式运行设置 (num_gpus=0)
3. 运行环境无gpu
4. dataset 已经注册在LOAD_DATASET 里
5. 报错位置

opencompass/task/openical_infer.py

def run(self, cur_model=None, cur_model_abbr=None):
        self.logger.info(f'Task {task_abbr_from_cfg(self.cfg)}')
        for model_cfg, dataset_cfgs in zip(self.model_cfgs, self.dataset_cfgs):
            self.max_out_len = model_cfg.get('max_out_len', None)
            self.batch_size = model_cfg.get('batch_size', None)
            self.min_out_len = model_cfg.get('min_out_len', None)
            if cur_model and cur_model_abbr == model_abbr_from_cfg(model_cfg):
                self.model = cur_model
            else:
                self.model = build_model_from_cfg(model_cfg)

            for dataset_cfg in dataset_cfgs:
                self.model_cfg = model_cfg
                self.dataset_cfg = dataset_cfg
                self.infer_cfg = self.dataset_cfg['infer_cfg']
                self.dataset = build_dataset_from_cfg(self.dataset_cfg)

Reproduces the problem - code/configuration sample

执行的配置代码

from docs.en.conf import project
from opencompass.datasets.mymodel.load_data import MyDataset
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.runners import LocalRunner
from opencompass.tasks import OpenICLInferTask
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.partitioners import NaivePartitioner

from opencompass.metrics.mymodel.eval_metrics import ObjectiveEvaluator
from opencompass.models.mymodel.llama_cpp_model import LlamaCppModel
from opencompass.models import HuggingFacewithChatTemplate,HuggingFaceBaseModel

reader_cfg = dict(
    input_columns=['question', 'prompt'],
    output_column='reference',
)

infer_cfg = dict(
    # Prompt 生成配置
    prompt_template=dict(
        type=PromptTemplate,
        # Prompt 模板，模板形式与后续指定的 inferencer 类型相匹配
        # 这里为了计算 PPL，需要指定每个答案对应的 Prompt 模板
        template=dict(
            begin=[
                dict(role='system', prompt='{prompt}\n')
            ],
            round=[
                dict(role='user', prompt='{question}\n'),
            ])),
    retriever=dict(type=ZeroRetriever),
    # 推理方式配置
    #   - PPLInferencer 使用 PPL（困惑度）获取答案
    #   - GenInferencer 使用模型的生成结果获取答案
    inferencer=dict(type=GenInferencer),
    partitioner=dict(type=NaivePartitioner),
    runner=dict(type=LocalRunner, max_num_workers=16, task=dict(type=OpenICLInferTask)),
)

project_ch=['文本扩写']
project_en=['text_expansion']

datasets = []

for ch,en in zip(project_ch,project_en):
    eval_cfg = dict(evaluator=dict(type=ObjectiveEvaluator, project=ch), num_gpus=0)
    datasets.append(dict(
            type=MyDataset,
            abbr=en,
            filename=en,
            path=f'./data/{en}',
            reader_cfg=reader_cfg,
            infer_cfg=infer_cfg,
            eval_cfg=eval_cfg)
    )

models = [
    dict(
        abbr='qwen2.5-1.5B-instruct',
        type=HuggingFacewithChatTemplate,
        path='opencompass/models/mymodel/modelscope/qwen/Qwen2___5-1___5B-instruct',
        max_out_len=5000,
        batch_size=16,
        run_cfg=dict(num_gpus=0),
        stop_words=['<|im_end|>', '<|im_start|>'],
        max_seq_len=5000
        # model_kwargs=dict(tensor_parallel_size=1, gpu_memory_utilization=0.8),
    )
]

@LOAD_DATASET.register_module()
class MyDataset(BaseDataset):

Reproduces the problem - command or script

python3 run.py configs/objective_testcase/eval_all.py

Reproduces the problem - error message

标准输出：

launch OpenICLInfer[qwen2.5-1.5B-instruct/schedule_extraction] on CPU                                                                   
  0%|                                                                                                             | 0/1 [00:00<?, ?it/s]11/29 14:19:04 - OpenCompass - ERROR - /home/caesar/codes/opencompass/opencompass/runners/local.py - _launch - 236 - task OpenICLInfer[qwen2.5-1.5B-instruct/schedule_extraction] fail, see
outputs/all/20241129_141900/logs/infer/qwen2.5-1.5B-instruct/schedule_extraction.out
100%| 1/1 [00:04<00:00,  4.80s/it]
11/29 14:19:04 - OpenCompass - ERROR - /home/caesar/codes/opencompass/opencompass/runners/base.py - summarize - 64 - OpenICLInfer[qwen2.5-1.5B-instruct/schedule_extraction] failed with code 1
11/29 14:19:04 - OpenCompass - INFO - Partitioned into 1 tasks.

output输出：

11/29 15:08:35 - OpenCompass - INFO - Task [qwen2.5-1.5B-instruct/schedule_extraction]
WARNING 11-29 15:08:37 _custom_ops.py:19] Failed to import from vllm._C with ImportError('libcuda.so.1: cannot open shared object file: No such file or directory')
11/29 15:08:40 - OpenCompass - INFO - using stop words: ['<|endoftext|>', '<|im_start|>', '<|im_end|>']
Traceback (most recent call last):
  File "/home/caesar/codes/opencompass/opencompass/tasks/openicl_infer.py", line 161, in <module>
    inferencer.run()
  File "/home/caesar/codes/opencompass/opencompass/tasks/openicl_infer.py", line 79, in run
    self.dataset = build_dataset_from_cfg(self.dataset_cfg)
  File "/home/caesar/miniconda3/envs/opencompass/lib/python3.10/site-packages/opencompass/utils/build.py", line 13, in build_dataset_from_cfg
    return LOAD_DATASET.build(dataset_cfg)
  File "/home/caesar/miniconda3/envs/opencompass/lib/python3.10/site-packages/mmengine/registry/registry.py", line 570, in build
    return self.build_func(cfg, *args, **kwargs, registry=self)
  File "/home/caesar/miniconda3/envs/opencompass/lib/python3.10/site-packages/mmengine/registry/build_functions.py", line 100, in build_from_cfg
    raise KeyError(
KeyError: 'opencompass.datasets.mymodel.mymodel.MyDataset is not in the opencompass::load_dataset registry. Please check whether the value of `opencompass.datasets.mymodel.mymodel.MyDataset` is correct or it was registered as expected. More details can be found at https://mmengine.readthedocs.io/en/latest/advanced_tutorials/config.html#import-the-custom-module'

Other information

在main函数中打印LOAD_DATASET 是一直有MYDATASET的；但是运行报错代码的时候 LOAD_DATASET里就没有MYDATASET
了，但是main打印的在执行task前后都是有的，不知道是不是多线程导致的

[tonysy]

You need import the datasets in opencompass/datasets/__init__.py