--debug命令可以正常推理，但是不使用debug则无法推理

[ZzzybEric]

我们将本地模型部署在服务端，包装成openai接口，在用opencompass推理时，使用如下指令
python3 run.py configs/eval_gpt3.5.py -w outputs/demo --debug，此时接口能正常接收prompt

但是使用如下指令时python3 run.py configs/eval_gpt3.5.py -w outputs/demo，也就是去掉debug，outputs/demo/2023135631/logs/eval/GPT-3.5-turbo/ceval-computer_network.out里则显示prediction no found，接口也不会被调用

另外outputs/demo/2023135631/logs/infer/GPT-3.5-turbo/ceval-computer_network.out里只有一行内容/bin/sh: 1: python: not found

[gaotongxiao]

In non-debugging mode, tasks are performed in independent processes via commands starting with python. Apparently, your environment does not have python.
You may set up a global soft link in /usr/bin or somewhere else and link python to python3.

[ZzzybEric]

Thank you for your reply, it has solved my problem. But I would like to know why I used a local model for inference before, using Python 3 directly without debugging, and can still infer the results normally.
"python3 run.py --models hf_opt_125m --datasets ceval_gen"
opt_125m is downloaded locally, and the model address in the configuration file is also changed to local
Please help me answer this question. Thank you

[gaotongxiao]

The differences come from here:

opencompass/opencompass/tasks/openicl_infer.py

Lines 45 to 53 in 681d301

	if self.num_gpus > 0:
	port = random.randint(12000, 32000)
	command = (f'torchrun --master_port={port} '
	f'--nproc_per_node {self.num_procs} '
	f'{script_path} {cfg_path}')
	else:
	command = f'python {script_path} {cfg_path}'
	return template.format(task_cmd=command)

API model does not require GPU, and the execution command starts with "python". The commands used to run GPU-based models start with torchrun, and therefore bypass the error you encountered.

[ZzzybEric]

Thanks !