Building around Highway
Model card · ComfyUI · Downloads · License and attribution
Highway is intended to support a community of workflows, integrations, adapters and research tools. The project does not need to maintain every trainer itself. The aim is to document stable compatibility boundaries so that independent developers can build around them.
Available interfaces
| Area | Status |
|---|---|
| Example ComfyUI workflows | Three supplied JSON workflows |
| Compatible LoRA inference | Implemented in the matching runtime; loading support is not a promise that every H3 adapter improves quality |
| LoRA interchange documentation | Interface kit: exact logical targets, shapes, alpha/rank and QKV rules, plus a CPU-only header checker |
| Community trainer or fine-tune implementation | Not supplied or qualified by this release candidate |
| Full fine-tuned model export/import | A complete public route is not yet established |
| Model/runtime delivery | Versioned checkpoint/runtime pair in Downloads |
The interface kit documents compatibility; component licenses remain applicable. It describes 368 invoked linear targets and 52 packed QKV aliases for the identified rc.3.1 Safetensors checkpoint. The catalog was paired against that checkpoint's saved model definition; names, shapes, computation dtypes and packed aliases are unchanged from the previous interface. Its serialization and actual generation qualification are reported separately. It contains no model payload, executable generation graph or inference kernels.
LoRA: a usable interchange boundary
The native importer accepts paired <target>.lora_A.weight and <target>.lora_B.weight tensors, with an optional scalar <target>.alpha, in a local safetensors file. The kit documents the accepted names, dimensions and coordinate basis. It also provides a command-line checker that reads the header and scalar alpha values without loading Highway or using a GPU.
The checker helps diagnose wrongly named or wrongly shaped tensors and packed/split target conflicts before an expensive run. It does not read the A/B payloads, verify their actual Q/K/V row order, prove the adapter was trained correctly, predict visual quality or grant distribution rights. PEFT, Kohya and other formats may require explicit conversion.
Community training
The shipped executor is optimized for inference and explicitly rejects gradient-enabled execution. An external trainer can use a different differentiable implementation; it still needs the correct base, forward/conditioning/noise contract and parameter mapping. A complete training route has not been qualified. A file-format specification alone is not enough to claim training support.
The training guide identifies four requirements: accessible base/computation, learning and conditioning conventions, a trained-adapter round trip, and a separate export path for full fine-tuned weights. Community authors can develop these independently under the applicable terms. No official trainer, training-memory claim or universal compatibility promise is made here.
Contributions and reproducibility
Useful contributions include workflows, UI translations, hardware compatibility reports, adapter conversion tools, evaluation protocols and independently implemented training research. State the tested model/runtime hashes and applicable terms; distinguish a CPU format check from an actual generation or training result. Share reproducible prompts and authorized test inputs rather than private source material.
Please follow the applicable licenses and attribution requirements. The project profile is DucHaiten.