/fastgen-add-modality

name: fastgen-add-modality
description: >-
  Extend NVIDIA FastGen to a new generated or conditioning modality beyond the existing image, video, audio, camera and text paths: representation, loader keys, model state, noise-schedule clock, losses, sampling, serialization and validation. Use it when a new signal type must cross the data, model and output interfaces.
license: Apache-2.0
metadata:
  author: "Julius Berner <jberner@nvidia.com>"
  domain: ml

Add a FastGen modality

Apply [fastgen-develop](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-develop/SKILL.md) and [fastgen-runtime](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-runtime/SKILL.md). Read the [network](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/README.md) and [dataset](https://github.com/NVlabs/FastGen/blob/main/fastgen/datasets/README.md) READMEs, then [modality boundaries](references/modality-contract.md). Use [fastgen-add-network](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-add-network/SKILL.md) alone for another model of an already supported task.

1. **Role.** Establish whether the modality is generated state, conditioning or control, or a derived training target, from the upstream model and its source. Record representation, units, normalisation, dimensions, temporal alignment, masks and decode semantics; a checkpoint name or an extra channel is not evidence of support. 2. **Trace.** Follow one sample through loader, `prepare_batch`, network, method loss, generator and output writer; note which boundaries already handle it and which assume image, video, audio or a single tensor. Preserve raw-versus-encoded distinctions and upstream packing, and do not turn a conditioning signal into diffusion state to fit an existing multimodal example. 3. **Model.** Use [fastgen-add-network](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-add-network/SKILL.md) for encoder and backbone changes; add only the needed shape inference, schedule dispatch and conditioning paths. For joint generation check whether modalities share a clock, grid and batch and keep the upstream relationship. 4. **Data.** Use [fastgen-dataset](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-dataset/SKILL.md) for acquisition, serialization, pairing and loader registration; version learned normalisation with the model artifacts and verify encode/decode and alignment before large preprocessing. 5. **Methods.** Use [fastgen-configure](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-configure/SKILL.md) for SFT and the requested distillation recipes; trace each method's tensor operations, since a tree utility accepting a key does not make the method support it. Reject unsupported combinations explicitly. 6. **Outputs.** Extend an inference entry point and serializer where the existing ones assume conventional media; keep native arrays in a usable format with shape and unit metadata, separate from rendered previews, tracker logging and metrics. 7. **Validate.** Use [fastgen-validate](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-validate/SKILL.md) for parity and reviewable comparisons, then [fastgen-train](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-train/SKILL.md) for short SFT and the requested method; check per-modality gradients, conditioning sensitivity, serialization of every generated modality, checkpoint reload and that existing modes still work.

Report supported input and output modes, implementation, data and config paths, native artifacts and evidence per tested combination. For plausible but wrong outputs check units, packing, alignment and normalisation before tuning losses, and state whether the result is a new conditioning path or a newly generatable modality.