Skip to main content

How to Add a Pipeline Config

This page assumes you are not trying to redesign the runtime. The goal is to add a new task config with minimal Python changes.

Decide which layer you actually need to change​

  • If you are only adding a task, a pipeline YAML is usually enough
  • If you are only switching models or prompts, update init_config or prompt_info
  • If you need a new metric, you may need a new module or a new mapping
  • If you need a new input schema, you may also need a new prompt asset and schema class

Minimal template​

_base_:
  pipes_default: "../modules_init/pipes_default.yaml"
  experts_default: "../modules_init/experts_default.yaml"

name: object-centric
support_task:
  - subject_add

parser_grounder_config:
  instruction_parser:
    default_config: ${pipes_default.instruction_parser}
    init_config:
      backend: vllm
      model_name: Qwen3-4B-Instruct-2507
      prompt_info:
        prompt_id: llm/instruction_parsing/basic_object
        version: v1
  general_grounder:
    default_config: ${pipes_default.general_grounder}
    init_config:
      backend: vllm
      model_name: Qwen3-VL-8B-Instruct
      prompt_info:
        prompt_id: vlm/grounding/general_grounding
        version: v1

metric_configs:
  lpips:
    pipe_name: lpips-pipe
    default_config: ${pipes_default.lpips-pipe}
    init_config:
    scope: unedit_area
  1. Copy the closest starter config
  2. Make sure name is an existing registered pipeline name
  3. Update support_task so it matches the CLI --edit-task
  4. Update prompt_info and confirm the prompt asset exists
  5. Check each pipe_name and scope in metric_configs
  6. If the pipeline is human-centric, complete expert_configs
  7. Validate with a small annotation run before scaling up

When YAML is no longer enough​

Python changes are usually necessary only when:

  • no existing module can express the metric you need
  • the prompt output schema requires new fields
  • the dataset loader cannot consume your data format
  • the post-processing logic needs a genuinely new conversion path

Configuration checklist​

  • Does support_task include the task you will pass to the CLI?
  • Does prompt_id/version point to a real prompt asset?
  • Is every pipe_name already registered?
  • Do model_paths and client_config resolve on the current machine?
  • Are all required human-centric experts configured?