SeedOmni Skill

Use this file as a routing entrypoint and load the references matching the task from the routing table below.

First Checks

Confirm the task touches SeedOmni:
- Source code under veomni/models/seed_omni/.
- Configs under configs/seed_omni/.
- SeedOmni data collation or conversation_list preprocessing.
- SeedOmni checkpoint splitting or module registration.
Treat current code as the source of truth (graph/module source and live configs).
Choose the smallest applicable reference set from the routing table below.
Use live fake_model, Janus, and Qwen modules as implementation maps.

Routing

Task	Read
Understand SeedOmni architecture, runtime hierarchy, and hard constraints	`.agents/knowledge/veomni-seed-omni.md`, `references/source-of-truth.md`
Change data flow, `conversation_list`, CPU preprocessing, dummy items, or raw request preprocessing	`references/data-flow.md`
Add or modify a module mixin/model/config/processor	`references/module-contract.md`
Edit training graph or generation FSM YAML	`references/graph-runtime.md`, then the matching `templates/*.yaml`
Add a new model config layout	`references/graph-runtime.md`, `templates/base.template.yaml`, `templates/modules_train.template.yaml`
Add distributed/eager inference module config	`references/per-module-parallel.md`, `templates/modules_infer_eager.template.yaml`, `templates/modules_infer_fsdp.template.yaml`
Update split-checkpoint conversion	`references/checkpoint-splitting.md`
Validate a SeedOmni change	`references/validation.md`

Current Baseline

Entry-time samples use {"conversation_list": [...]}. seedomni_transform runs ordered module CPU preprocessors (OmniProcessor) per sample in the Dataset stage, and SeedOmniCollator groups preprocessed samples into a micro-batch before GPU forward.
ModulePreprocessorBase is a core data-flow contract, shared by training and inference. Training runs it inside data_transform (seedomni_transform); inference runs it once before the FSM.
Modules communicate by mutating and returning the shared conversation_list carrier.
Training graphs are flat edge lists over module[.method] endpoints. to: end is the virtual sink. Execution order is derived by topo sort.
Generation graphs are FSMs. Each state body is an ordered inline edge list. done is framework-injected; use module_signal and default transitions.
Store tokenizers, processors, and other assets in the owning module's subfolder.

Templates

Templates under templates/ are skeletons, not authoritative examples. Use them to start a new config file, then compare against the live configs listed in references/module-contract.md.

Sibling Skills

Use veomni-new-model first when adding a new HuggingFace model family under veomni/models/transformers/.
Use this skill after that model exists, when wrapping it as SeedOmni modules or wiring it into Omni graphs.
Use veomni-develop for unrelated framework work outside SeedOmni.

veomni-seed-omni