Quick Start: LingBot-VA Training

This guide walks you through launching LingBot-VA post-training in the LoongForge framework. The examples provide two 8-GPU training entrypoints for RoboTwin and LIBERO.

0. Resource Preparation

The LingBot-VA post-training weights and LeRobot datasets are published at the following locations:

Resource	Link
RoboTwin Post-training Weights	https://huggingface.co/robbyant/lingbot-va-posttrain-robotwin
LIBERO-Long Post-training Weights	https://huggingface.co/robbyant/lingbot-va-posttrain-libero-long
RoboTwin LeRobot Dataset	https://huggingface.co/datasets/robbyant/robotwin-clean-and-aug-lerobot
LIBERO-Long LeRobot Dataset	https://huggingface.co/datasets/robbyant/libero-long-lerobot

This project follows the same offline preprocessing scheme as the official LingBot-VA community. Please stage the weights and data under the local /workspace directory in advance. The data directory must contain pre-generated latents and the corresponding empty_emb.pt. When using your own dataset, first convert the raw robot data to LeRobot format, then augment each episode in meta/episodes.jsonl with action_config (start/end frames of action clips and text descriptions), map the actions to the standard 30-dimensional format (pad missing dimensions with 0), and extract video latents with Wan2.2 VAE, placing them under a latents/ directory that mirrors videos/. Videos should be preprocessed according to the official data spec. For detailed fields, directory layout, and preprocessing commands, refer to the Custom Dataset Preparation section of the official README.md.

1. Launch Training

cd /workspace/LoongForge

export LOONGFORGE_PATH=/workspace/LoongForge
export NCCL_DEBUG=WARN
export GPUS_PER_NODE=8

RoboTwin:

export CHECKPOINT_PATH=/workspace/models/lingbot-va-posttrain-robotwin
export DATA_PATH=/workspace/datasets/robotwin-clean-and-aug-lerobot
export OUTPUT_DIR=/workspace/outputs/lingbot_va_robotwin

bash examples/embodied/lingbot_va/run_lingbot_va_robotwin_finetune.sh

LIBERO-Long:

export CHECKPOINT_PATH=/workspace/models/lingbot-va-posttrain-libero-long
export DATA_PATH=/workspace/datasets/libero-long-lerobot
export OUTPUT_DIR=/workspace/outputs/lingbot_va_libero

bash examples/embodied/lingbot_va/run_lingbot_va_libero_finetune.sh

2. Performance Optimization Options

2.1 Baked-in Optimizations

The following optimizations are baked into the LingBot implementation and have no toggle:

Optimization	Effect	Bottleneck Addressed
Self-Flex mask cache	Reuse pre-computed attention masks.	Avoid re-constructing masks and the associated CPU/GPU overhead.
LeRobot repo discovery cache	Cache the list of repos in the data directory.	Avoid repeatedly scanning large volumes of metadata each epoch, reducing startup and dataloader initialization overhead.
Transformer Engine Q/K RMSNorm	Use the fused Q/K RMSNorm implementation.	Reduce normalization kernel launches, memory access, and intermediate tensor overhead.
Self-Flex block64	Fix the block64 configuration for Flex Attention’s Triton fused forward/backward kernels (FWD=64,64,4,3, BWD=32,32,4,1).	Avoid falling back to dense attention, reducing attention compute cost and memory pressure.
Cross attention optimized implementation	Use an efficient cross-attention path.	Reduce kernel and memory overhead of cross-attention.

2.2 Optional Optimizations

The following performance optimizations are user-controllable and are enabled by default in the example scripts:

Switch	Default	Effect
LINGBOT_BALANCED_SAMPLER	1	Perform deterministic load balancing across DP ranks based on per-sample compute cost, reducing slow-rank stragglers; this also changes sample loading order, so disable it if you need step-by-step alignment of loss with the official community.
LINGBOT_LAYERWISE_COMPILE	1	Compile the layerwise norm, modulation, and residual gate paths.

For example, to disable the optional optimizations:

The command below uses RoboTwin as an example; for LIBERO, use run_lingbot_va_libero_finetune.sh.

LINGBOT_BALANCED_SAMPLER=0 \
LINGBOT_LAYERWISE_COMPILE=0 \
bash examples/embodied/lingbot_va/run_lingbot_va_robotwin_finetune.sh

The following switches control logging, diagnostics, or artifact management and do not change model training semantics:

Switch	Default	Effect
LINGBOT_BASELINE_LOSS_LOG	1	Emit loss logs compatible with the upstream baseline, which facilitates comparison of training curves.
LINGBOT_SAMPLE_META_EXPORT	0	Export sample/frame/CFG metadata, for diagnosing sample order and reproducing experiments.
LINGBOT_SKIP_FINAL_CHECKPOINT	0	Skip final checkpoint saving to reduce disk footprint and shutdown time.

2.3 Correctness Verification

To ensure the optimizations do not affect training accuracy, we performed a step-by-step action-loss comparison between LoongForge’s LingBot-VA adaptation and the official implementation under identical data, weights, and training configurations. The results show that LoongForge’s performance optimizations do not affect training accuracy: