Chunk relational model for consistent CUDA performance

fixed batched relational

updated pipeline

updated pipeline relational to match v1

Added config to the downloaded artifacts

Pad recognizer classifier to 64-aligned output for tensor-core efficiency

Vectorize post-processing format step and default to white padding

Remove unused code: training-only methods, dead model blocks, and test files

Accept torch.Tensor inputs, remove non-release scripts

Clean up docs, make syncs optional, update examples for new modes

Make per-batch CUDA syncs and logging optional via verbose_post

Add detector_only and skip_relational inference modes

Lower default detector batch size to 8 for 16GB GPU compatibility

Reduce recognizer VRAM by discarding logits per chunk

Add pre-NMS centerness + peak filter for faster inference

merged english and multilingual and updated model card

Duplicate from nvidia/nemotron-ocr-v2-multilingual