Skip to content

adding-a-baseline.md

Latest commit

 

History

History
50 lines (42 loc) · 3.04 KB

adding-a-baseline.md

File metadata and controls

50 lines (42 loc) · 3.04 KB
Raw

Adding a new model baseline

The checklist for supporting a new image-generation architecture ("baseline") end to end. The mechanics of authoring the pipeline itself are in adding-a-pipeline.md; this page covers the baseline-specific extras.

The CI tripwire tests/pipeline/test_baseline_coverage.py fails the moment a new KNOWN_IMAGE_GENERATION_BASELINE enum value appears in horde_model_reference, and stays red until the baseline is consciously classified — either by completing this checklist or by adding it to the test's UNSUPPORTED_BASELINES set.

Where baseline knowledge is allowed to live

Baseline-conditional logic must use KNOWN_IMAGE_GENERATION_BASELINE enum members — never raw strings — and is confined to:

  • hordelib/pipeline/families/image_gen/baselines.py — the BaselineProfile table (loader kind, force-load policy) and the named baseline groupings (FLUX_BASELINES, ...)
  • hordelib/pipeline/families/image_gen/ — per-pipeline selectors and patch steps
  • hordelib/pipeline/context.py — the ModelContext data structure (its baseline field is the typed input every selector reads; no conditional logic lives here)

ModelContext carries no per-baseline boolean properties. Every pipeline selects by declaring a baseline set on its ImageSelector (baselines=frozenset({...})); a family that spans several baselines (as flux does) declares a named frozenset grouping in baselines.py and reuses it.

Checklist

  1. Model reference: the baseline exists as a KNOWN_IMAGE_GENERATION_BASELINE member and the models are present in stable_diffusion.json (horde_model_reference).
  2. Pipeline: follow adding-a-pipeline.md — graph export, a definition module in families/image_gen/ selecting on the baseline at SelectionTier.BASELINE_FAMILY, and a spot in IMAGE_PIPELINES.
  3. Baseline profile: if the baseline loads split files (bare diffusion model + separate CLIP/VAE) or its models must not be force-loaded into VRAM on weaker cards, add a BaselineProfile row in hordelib/pipeline/families/image_gen/baselines.py (loader=LoaderKind.UNET and/or force_load_skip_classes=("<comfy model_base class>",)) and mirror any new comfy class names in FORCE_LOAD_SKIP_CLASS_NAMES (hordelib/execution/comfy_patches.py — kept horde_model_reference-free on purpose; the startup tripwire fails if the two disagree). Consumers then pass the enum member to initialise(models_not_to_force_load=...).
  4. Classify in the tripwire: move the baseline from UNSUPPORTED_BASELINES to BASELINE_EXPECTED_TEMPLATE in tests/pipeline/test_baseline_coverage.py, mapping it to the new pipeline's name.
  5. GPU oracle: add an inference test (tests/test_horde_inference_<name>.py) with reference outputs in images_expected/.
  6. Worker: zero changes required. The worker resolves models via the reference and passes baseline enums through hordelib.api; new baselines flow through automatically once the worker's model list includes them.