Configuration Schema#

This page organizes the common configuration structures in FluxVLA by top-level keys, enabling quick navigation from field locations back to the corresponding tutorials.

Purpose#

Quickly understand the composition of top-level blocks in configuration files
Clarify field modification priorities (required / recommended / optional)
Provide minimal configuration paths for training, evaluation, and inference

Top-Level Structure#

model = dict(...)
# inference_model = dict(...)  # 可选
train_dataloader = dict(...)
runner = dict(...)
inference = dict(...)
eval = dict(...)

Core Field Quick Reference#

`model`#

Field	Description
`type`	Model type
`pretrained_name_or_path`	Pretrained weights
`vlm_backbone` / `vla_head`	Submodule configuration
`name_mapping`	Weight mapping

`train_dataloader`#

Field	Description
`per_device_batch_size`	Batch size per device
`per_device_num_workers`	Number of workers per device
`dataset`	Dataset and transforms configuration

`runner`#

Field	Description
`type`	Training Runner type
`max_epochs` / `learning_rate`	Core training hyperparameters
`metric`	Metrics and logging configuration
`lr_scheduler_type` / `warmup_ratio`	Learning rate schedule

`inference` / `eval`#

Field	Description
`type`	Inference or evaluation Runner
`dataset`	Input processing configuration
`denormalize_action`	Action denormalization
`operator` (inference)	Robot communication interface

Minimal Example#

model = dict(type='LlavaVLA', ...)
train_dataloader = dict(per_device_batch_size=8, ...)
runner = dict(type='FSDPTrainRunner', max_epochs=6, learning_rate=2e-5, ...)