Add timestep-aware sparse KD weighting

Filter unusable Cauldron image payloads

Force cached KD masks during Stage 3b

Guarantee masked assistant tokens in diffusion training

Fix HF Sequence type alias in Stage 2 loader

Make Stage 2 Cauldron loading cost-safe

Fix The Cauldron aokvqa config name

Filter invalid LLaVA image records for real runs

Fix LLaVA fallback image loading in Stage 1

Implement stage-aware real-run training pipeline

Fix VRAM property access in training script to use total_memory

Add Trackio integration for enhanced logging in training process

Enhance LLaVAPretrainDataset to support fallback loading from JSON and improve image path resolution

Add dropout parameter to ProjConfig for improved model regularization

Update model configuration and training scripts with new vision backbone support and dependencies

Upload model_config.py

Upload vil_dlm_model.py

Upload vision_xlstm.py

Upload train_production.py