fastgen-configure

name: fastgen-configuredescription: >-  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.license: Apache-2.0metadata:   author: "Julius Berner <jberner@nvidia.com>"  domain: ml

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.

1. 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.
2. 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.
3. Compose. Start from the nearest compatible experiment config. If none fits the task, architecture and objective, 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.
4. 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.
5. 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.
6. 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.