Configure a FastGen experiment
Read the configuration README for composition, override
grammar and run directories, and the method index for the
chosen method's documentation. Read config contracts
when changing inherited values, geometry or supervision. For an algorithm the library
does not implement, use fastgen-add-method.
- Task. Infer the complete task from request, network and data: conditioning type
(none, class, text, image, video, camera, audio), generation versus editing, image or
video, joint audio-video, causal generation; checkpoint variant, resolution,
duration/FPS and target sampling budget (NFE). A model family name does not establish
that every mode is supported; check the
supported networks and the existing
experiment configs.
- Method. Keep an explicitly chosen method. Otherwise use SFT for fine-tuning and the
baseline guidance for
distillation, including data-free training and the
causal pipeline. Check the
method's prerequisite checkpoints and supervision in its category README.
- Compose. Start with hyperparameters from the experiment config closest in network
architecture, task and method. If none is compatible, compose the method config's
create_config() with compatible presets. Copy network, data, optimizer and callback
presets from fastgen/configs/*.py and derive loader geometry from the network where needed
(model_pixel_shape). Change only justified values; do not carry precision, guidance,
schedule, learning-rate or batch settings over from an unrelated recipe.
- Supervision. Decide data needs from the active losses, not the method name: SFT
needs targets, KD needs cached pairs or trajectories in the
KD loader format,
backward simulation replaces clean targets in student inputs but not in other enabled
losses, and image- or video-conditioned tasks always need their conditioning media.
Route missing data to fastgen-dataset.
- Alignment. Match loader keys and latent geometry to the network; align teacher and
student weights, prediction type, noise schedule, guidance and negative conditioning;
check that every enabled loss has its modules and trainable parameters, routing
missing capabilities to fastgen-add-network.
Set iterations, optimizer, effective batch, logging, checkpoints and
validation limits
to the budget.
- Resolve. Consult fastgen-runtime when writing
the command, including its override syntax. Run
--dryrun and inspect the resolved
values for placeholder paths, unintended weights, auxiliary networks and conflicting
distributed modes. Hand execution to
fastgen-train; a config-only task ends
with the reviewed config.
Report the config path, departures from the nearest recipe, the method and data rationale
and how far the config was validated. For Unknown config fields, inspect the typed
schema before adding keys; for unsupported method/network combinations explain the
missing capability instead of silently disabling a loss.
1---2name: fastgen-configure3description: >-4 Create or adapt an NVIDIA FastGen experiment config for a network, task, dataset and training objective, and pick a distillation method when none was chosen. Use it whenever a FastGen run needs a new or modified config file or non-trivial overrides.5license: Apache-2.06metadata:7 source: https://skillmd.com/skills/nvidia/fastgen-configure8 author: "Julius Berner <jberner@nvidia.com>"9 domain: ml10---1112# Configure a FastGen experiment1314Read the [configuration README](https://github.com/NVlabs/FastGen/blob/main/fastgen/configs/README.md) for composition, override15grammar and run directories, and the [method index](https://github.com/NVlabs/FastGen/blob/main/fastgen/methods/README.md) for the16chosen method's documentation. Read [config contracts](references/config-contract.md)17when changing inherited values, geometry or supervision. For an algorithm the library18does not implement, use [fastgen-add-method](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-add-method/SKILL.md).19201. **Task.** Infer the complete task from request, network and data: conditioning type21 (none, class, text, image, video, camera, audio), generation versus editing, image or22 video, joint audio-video, causal generation; checkpoint variant, resolution,23 duration/FPS and target sampling budget (NFE). A model family name does not establish24 that every mode is supported; check the25 [supported networks](https://github.com/NVlabs/FastGen/blob/main/fastgen/networks/README.md#supported-networks) and the existing26 experiment configs.272. **Method.** Keep an explicitly chosen method. Otherwise use SFT for fine-tuning and the28 [baseline guidance](https://github.com/NVlabs/FastGen/blob/main/fastgen/methods/README.md#choosing-a-distillation-baseline) for29 distillation, including data-free training and the30 [causal pipeline](https://github.com/NVlabs/FastGen/blob/main/fastgen/methods/README.md#causal-distillation-pipeline). Check the31 method's prerequisite checkpoints and supervision in its category README.323. **Compose.** Start with hyperparameters from the experiment config closest in network33 architecture, task and method. If none is compatible, compose the method config's34 `create_config()` with compatible presets. Copy network, data, optimizer and callback35 presets from `fastgen/configs/*.py` and derive loader geometry from the network where needed36 (`model_pixel_shape`). Change only justified values; do not carry precision, guidance,37 schedule, learning-rate or batch settings over from an unrelated recipe.384. **Supervision.** Decide data needs from the active losses, not the method name: SFT39 needs targets, KD needs cached pairs or trajectories in the40 [KD loader format](https://github.com/NVlabs/FastGen/blob/main/fastgen/datasets/README.md#knowledge-distillation-loaders),41 backward simulation replaces clean targets in student inputs but not in other enabled42 losses, and image- or video-conditioned tasks always need their conditioning media.43 Route missing data to [fastgen-dataset](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-dataset/SKILL.md).445. **Alignment.** Match loader keys and latent geometry to the network; align teacher and45 student weights, prediction type, noise schedule, guidance and negative conditioning;46 check that every enabled loss has its modules and trainable parameters, routing47 missing capabilities to [fastgen-add-network](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-add-network/SKILL.md).48 Set iterations, optimizer, effective batch, logging, checkpoints and49 [validation limits](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-train/references/run-contract.md#what-an-iteration-does)50 to the budget.516. **Resolve.** Consult [fastgen-runtime](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-runtime/SKILL.md) when writing52 the command, including its override syntax. Run `--dryrun` and inspect the resolved53 values for placeholder paths, unintended weights, auxiliary networks and conflicting54 distributed modes. Hand execution to55 [fastgen-train](https://github.com/NVlabs/FastGen/blob/main/skills/fastgen-train/SKILL.md); a config-only task ends56 with the reviewed config.5758Report the config path, departures from the nearest recipe, the method and data rationale59and how far the config was validated. For `Unknown config fields`, inspect the typed60schema before adding keys; for unsupported method/network combinations explain the61missing capability instead of silently disabling a loss.