- Fix misaligned address error in quad NMS cuda implementation (4ba370611ce8a0cfbc1fa47ade37f3c81c7fc3f8)
- Fix undefined idx variable in quad_non_maximal_suppression callThe for loop was removed but the [idx].unsqueeze(0) indexing wasleft behind, causing a NameError. Pass the full batch tensorsdirectly since the CUDA kernel now supports batch mode natively (5b3a64202d97f2d197204d4d49996117dae3328c)
- Remove model weights from resulting Python wheel (11e8b7fe11bd8496a6f38eb7a585b22321c09fc2)
- Simplify weight download to use hf_hub_download consistently- Remove weight_downloader.py module- Inline hf_hub_download calls in pipeline.py- Remove hf_token and force_download params from NemotronOCR- Simplify example.py (3575cd80133269d2425382b61eb875f9a9ffae3e)
Co-authored-by: Jeremy <jdye64@users.noreply.huggingface.co>
- Rename NemoRetriverOCR to NemotronOCR (9037b7929f55d87745efa222c1a615d23a830589)
Co-authored-by: Oliver Holworthy <nvidia-oliver-holworthy@users.noreply.huggingface.co>