The configuration system uses the YAML format. To simplify the configuration, we made some extensions on top of the
basic YAML syntax. import
, default
, overwrite
are our extended keywords.
The import
keyword is used to import configurations from other files. For example, the following two ways are
equivalent:
Method one:
# config.yaml
definition:
def1: something...
def2: something...
Method two:
# def1.yaml
def1: something...
# def2.yaml
def2: something...
# config.yaml
definition:
import:
- def1.yaml
- def2.yaml
The import
keyword supports using a string or list as its value, corresponding to importing a single file or multiple
files respectively.
During the import process, if there's an import
keyword in the imported file, the import
in the imported file will
be executed first. The same applies to the other two keywords.
During the import process, if a key conflict occurs, the system will try to recursively merge the values of the conflicting keys. If the merge is not possible, the later value will overwrite the earlier one.
The default
keyword is used to specify default values. The following two ways are equivalent:
Method one:
definition:
def1:
type: int
value: 1
def2:
type: int
value: 2
def3:
type: float
value: 1.1
Method two:
definition:
default:
type: int
def1:
value: 1
def2:
value: 2
def3:
type: float
value: 1.1
The default
keyword supports a string, list, or dictionary as its value. The config parser will try to merge
the default
value with the value of the keys at the same level as default
. In the event of a conflict, the default
keyword value has a lower priority.
The overwrite
keyword works similarly to default
. However, in the event of a conflict, the overwrite
keyword value
has a higher priority. This keyword is often used with import
to set the required values under this configuration
file.
The main directory structure of the configuration file is as follows:
configs
├── assignments
│ ├── definition.yaml
│ ├── default.yaml
│ └── ...
├── agents
├── tasks
│ ├── task_assembly.yaml
│ └── ...
└── start_task.yaml
The assignments
directory contains all task configuration files. The definition.yaml
collects all task definitions
and model definitions.
A single task configuration file mainly requires the following fields:
definition
: usually imported fromdefinition.yaml
, for defining tasks and models.concurrency
: defines the maximum concurrency of the model.assignments
: accepts multipleassignment
, defining the specific allocation of tasks.output
: defines the path of the output file.
A single assignment
requires two fields:
agent
: the name of the agent required for this task.task
: the name of the task required for this task.
The agents
directory contains all agent configuration files. The key in the configuration is the agent's name, and the
value is the agent's configuration. A single agent configuration requires the following fields:
module
: defines the corresponding agent client module.parameters
: defines the parameters to be passed to the corresponding module.
The tasks
directory contains all task configuration files. The task_assembly.yaml
collects all task definitions. If
you only want to run existing tasks, you generally do not need to modify the files in this directory.
Similar to the agent configuration, the key is the task's name, and the value is the task's configuration. A single task configuration requires the following fields:
module
: defines the corresponding task module.parameters
: defines the parameters to be passed to the corresponding module.
This configuration file is used in conjunction with src.start_task
to automate the bulk launch of task_workers. This
file's fields are as follows:
definition
: used to define tasks, usually imported fromtask_assembly.yaml
.start(Optional)
: used to specify the tasks to start, the key is the task name, and the value is the number of workers to start.controller_address(Optional)
: used to specify the address of the controller, default is http://localhost:5000/api/