nvidia

/

NVIDIA-Nemotron-Parse-v1.2

like 30

Follow

NVIDIA 55.5k

This PR updates the custom Nemotron Parse decoder to work with both the legacy Transformers 4.x cache API and the newer 4.57/5.x cache interfaces, including MBartDecoderLayer signature differences and generation input handling. It also adds a reproducible uv + Docker test environment and a golden regression suite with captured reference outputs, so future dependency upgrades can be checked against the current 4.51.3 baseline.

What changed

• Added compatibility handling in hf_nemotron_parse_modeling.py for legacy tuple-based KV caches and newer Cache / EncoderDecoderCache flows.
• Added prepare_inputs_for_generation support so cached decoding works correctly during generation.
• Added pyproject.toml, Dockerfile, and docker-compose.yaml for a reproducible GPU-backed test environment.
• Added test_golden.py and golden_outputs.json to validate preprocessing, encoder outputs, decoder logits, generation, and processor behavior against known-good outputs.

Testing

• python test_golden.py --capture
• pytest test_golden.py -v