[Dana57740]

Change Summary

Add LoongForge-VLDA with two-stage LIBERO training and policy-server evaluation. Stage 1 initializes from OpenPI π₀.₅ base and V-JEPA 2 weights and learns the latent world model; Stage 2 initializes from Stage 1 and jointly learns world-conditioned action prediction.

• Add the VLDA model, configuration, three-stream Gemma integration, V-JEPA 2 encoder, teacher EMA, and checkpoint/statistics integration.
• Add a LeRobot v2 reader for embedded images, future frames, negative actions, and quantile normalization.
• Integrate VLDA with the shared dataset, trainer, model configuration, evaluation factory, and payload registries.
• Add an optional cosine-decay start step for Stage 2's warmup/hold/decay schedule. Its default of zero retains decay immediately after warmup.
• Add the cookbook/loongforge-vlda/ guide, two 8-GPU training launchers, and LIBERO evaluation configuration. Keep internal and evaluation identifiers as vlda.

Validation

• Completed the full two-stage training workflow: Stage 1 training and checkpoint saving, followed by Stage 2 initialization from the Stage 1 checkpoint, training, and checkpoint saving.
• Completed checkpoint-based LIBERO evaluation, verifying that the trained Stage 2 checkpoint and normalization statistics load correctly through the policy server and simulator workflow.
• Shell syntax and launcher/configuration checks: 25/25 checks passed, including invocation outside the repository, argument overrides, and invalid-input cases.
• Stage 1: 8 GPUs, global batch 96, 3 optimizer updates; compilation, FlexAttention, and gradient checkpointing enabled.
• Stage 2: 8 GPUs, global batch 128, 3 optimizer updates; compilation, FlexAttention, and gradient checkpointing enabled.
• Both stages exited successfully with finite losses/gradient norms and zero NaN or skipped iterations.
• LIBERO closed-loop smoke test: 1 task, 1 episode, 20 steps using OSMesa; 20 finite 7-dimensional actions, 2 inference calls, 18 action-cache hits, and no recorded incidents.
• All 32 changed Python files passed the CI-pinned Ruff 0.9.10 check and AST syntax checks. Changed files passed the sensitive-information scan with zero findings.
• Full upstream Static Checks suite and regression coverage for other models.

CI, GPU, and Image Impact

• This change documents model and simulator environment requirements; it does not modify dependency manifests, Dockerfiles, or patches.
• GPU validation is required. Suggested suite: embodied; the completed VLDA training, checkpoint-based evaluation, and additional smoke tests supplement the standard regression suite.
• A local candidate image should be built during GPU validation.
• No GPU or image validation is required.

Reviewer Notes

VLDA training currently supports DDP and DDP with ZeRO1; FSDP is not supported. Review the optional shared LR-scheduler argument and validation alongside the new model-specific paths. V-JEPA 2 code retains its upstream MIT license and attribution.

Before marking this draft ready: synchronize with the latest upstream master, resolve SPDX-header checks while retaining third-party notices, remove the README's trailing whitespace, and replace or remove its paper/framework placeholders. Local test artifacts are kept outside the committed source tree.