Add a FastGen network
Apply fastgen-develop
and fastgen-runtime. Start with the
network README and the interface docstrings in
network.py and module.py,
then inspect the closest maintained adapter for the same task. Read
network contracts before implementing and load its
conditional sections only for the capabilities being added. For a new signal type also
use fastgen-add-modality.
- Source. Verify the upstream model, its revision, code, weights, access and
supported modes. Diffusers implementations are reusable, but check the authors'
reference for precision and preprocessing defaults, and whether the checkpoint's
precision or quantisation supports training. Never substitute a similarly named model.
- Trace upstream. Preprocessing, tokenizer and chat templates, conditioning and
masks, VAE scaling and packing, position encodings, time convention, scheduler grid
and guidance recipe, plus modality geometry and duration/FPS where relevant.
Distinguish native edit or reference conditioning from initialising noisy latents.
- Implement. Add the adapter under
fastgen/networks/<Name>/ and its presets in
net.py. Follow the module lifecycle, preprocessing,
conditioning, output-transformation and sampling contracts; share task-independent
components such as text encoders and VAEs, and do not embed an inference pipeline as
the training forward. Load weights through the documented
pretrained-model conventions so
local mirrors and offline mode keep working. Audit precision, buffer reconstruction,
meta loading and sharding for the network and every preprocessor.
- Capabilities. Trace the requested methods' forward and loss calls; add feature and
discriminator gradients, second-time inputs, PDD, LoRA, multimodal or causal support
as needed and check each independently. SFT support implies none of them.
Add and run small forward, conditioning and capability regressions following
tests/test_network*.py and tests/test_conditioning.py before expensive parity or
training runs.
- Configs. Use fastgen-configure for SFT and
the requested method configs, with negative-prompt presets where the upstream guidance
needs them and none where it does not.
- Parity. Follow validate a new network
with fastgen-validate and
fastgen-inference, and resolve unexplained
differences before training.
- Train and document. Use fastgen-train for a
short SFT run with media every iteration, then the requested methods; compare outputs
before and after and reload a checkpoint. Update the README tables (network list,
pretrained models).
Report implementation and config paths and per-task parity, training and checkpoint
results. If the loss decreases while outputs degrade, inspect input normalisation,
prediction and time convention and guidance before tuning the optimizer; mark
unsupported capabilities explicitly.
1---2name: fastgen-add-network3description: >-4 Integrate a new diffusion or flow model, or a task variant such as I2V, editing or causal generation, into NVIDIA FastGen as a FastGenNetwork adapter with preprocessors, configs, method capabilities, reference parity and trainability. Use it whenever a model family or checkpoint is not yet supported.5license: Apache-2.06metadata:7 source: https://skillmd.com/skills/nvidia/fastgen-add-network8 author: "Julius Berner <jberner@nvidia.com>"9 domain: ml10---1112# Add a FastGen network1314Apply [fastgen-develop](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-develop/SKILL.md)15and [fastgen-runtime](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-runtime/SKILL.md). Start with the16[network README](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/README.md) and the interface docstrings in17[network.py](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/network.py) and [module.py](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/module.py),18then inspect the closest maintained adapter for the same task. Read19[network contracts](references/network-contract.md) before implementing and load its20conditional sections only for the capabilities being added. For a new signal type also21use [fastgen-add-modality](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-add-modality/SKILL.md).22231. **Source.** Verify the upstream model, its revision, code, weights, access and24 supported modes. Diffusers implementations are reusable, but check the authors'25 reference for precision and preprocessing defaults, and whether the checkpoint's26 precision or quantisation supports training. Never substitute a similarly named model.272. **Trace upstream.** Preprocessing, tokenizer and chat templates, conditioning and28 masks, VAE scaling and packing, position encodings, time convention, scheduler grid29 and guidance recipe, plus modality geometry and duration/FPS where relevant.30 Distinguish native edit or reference conditioning from initialising noisy latents.313. **Implement.** Add the adapter under `fastgen/networks/<Name>/` and its presets in32 [net.py](https://github.com/NVlabs/FastGen/blob/main/fastgen/configs/net.py). Follow the module lifecycle, preprocessing,33 conditioning, output-transformation and sampling contracts; share task-independent34 components such as text encoders and VAEs, and do not embed an inference pipeline as35 the training forward. Load weights through the documented36 [pretrained-model conventions](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/README.md#pretrained-models) so37 local mirrors and offline mode keep working. Audit precision, buffer reconstruction,38 meta loading and sharding for the network and every preprocessor.394. **Capabilities.** Trace the requested methods' forward and loss calls; add feature and40 discriminator gradients, second-time inputs, PDD, LoRA, multimodal or causal support41 as needed and check each independently. SFT support implies none of them.42 Add and run small forward, conditioning and capability regressions following43 `tests/test_network*.py` and `tests/test_conditioning.py` before expensive parity or44 training runs.455. **Configs.** Use [fastgen-configure](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-configure/SKILL.md) for SFT and46 the requested method configs, with negative-prompt presets where the upstream guidance47 needs them and none where it does not.486. **Parity.** Follow [validate a new network](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/README.md#validate-a-new-network)49 with [fastgen-validate](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-validate/SKILL.md) and50 [fastgen-inference](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-inference/SKILL.md), and resolve unexplained51 differences before training.527. **Train and document.** Use [fastgen-train](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-train/SKILL.md) for a53 short SFT run with media every iteration, then the requested methods; compare outputs54 before and after and reload a checkpoint. Update the README tables (network list,55 pretrained models).5657Report implementation and config paths and per-task parity, training and checkpoint58results. If the loss decreases while outputs degrade, inspect input normalisation,59prediction and time convention and guidance before tuning the optimizer; mark60unsupported capabilities explicitly.