LibreSenseNovaVision - third-party attribution
==============================================

The SenseNova-Vision family vendors an inference-only port of the upstream
model implementation. All ported files retain their upstream copyright
headers and per-file provenance notes.

    Project:  SenseNova-Vision
    Upstream: https://github.com/OpenSenseNova/SenseNova-Vision
    Commit:   12ccd96e32b32967a11cacb6c5bd5fe3a555fc0c
    License:  Apache License 2.0 (repository code)
    Copyright (c) 2026 SenseTime Group Inc. and/or its affiliates

The upstream implementation itself derives from Apache-2.0 projects, and the
ported files keep those lineage headers:

  * Bagel (Qwen2-NaViT MoT decoder, SigLIP-NaViT tower, inferencer):
    Copyright (c) 2025 Bytedance Ltd., Apache-2.0
    https://github.com/ByteDance-Seed/Bagel
  * Qwen2 / SigLIP leaf modules and configurations:
    Copyright (c) 2024 The Qwen Team and The HuggingFace Inc. team, Apache-2.0
    https://github.com/huggingface/transformers
  * FLUX autoencoder (modeling/autoencoder.py):
    Copyright (c) 2024 Black Forest Labs, Apache-2.0
    https://github.com/black-forest-labs/flux

LibreYOLO adaptations (documented in each file header): the training-only
forward paths are removed, flash-attn is optional behind a numerically
equivalent SDPA fallback, einops rearranges are rewritten as plain torch ops,
and imports are made relative.

NOT ported: upstream modeling/bagel/modeling_utils.py is marked
CC BY-NC 4.0 (it derives from facebookresearch/DiT) and is incompatible with
LibreYOLO. The small standard components it provided are re-derived in
libreyolo/models/sensenova/modeling/layers.py from the original permissive
sources instead:

  * 2D sine-cosine position table: Hugging Face transformers
    (models/vit_mae/modeling_vit_mae.py, Apache-2.0, Meta copyright under
    Apache in that file), which predates DiT's non-commercial fork of the
    same routine.
  * Sinusoidal timestep embedding: openai/guided-diffusion
    (guided_diffusion/nn.py, MIT), the source DiT itself credits.
  * The MLP connector and frozen-buffer wrappers follow the parameter
    names/shapes fixed by the released checkpoint.

Model weights are CC BY-NC 4.0 (non-commercial use only) and are mirrored,
byte-identical and with attribution, at:

    Mirror:     LibreYOLO/SenseNovaVision7b (Hugging Face)
    Upstream:   sensenova/SenseNova-Vision-7B-MoT
    Revision:   79548fcc5b954598799b9317f8d3ec5e347d5c0e
    SHA-256:    ema.safetensors
                96f29abd98791288c5a24087322e964bb9bcabfc2f185ece71543f827bc2b11e
                ae.safetensors
                afc8e28272cd15db3919bacdb6918ce9c1ed22e96cb12c4d5ed0fba823529e38

Mirroring does not change the license; the loader prints the
non-commercial notice before every automatic download, and users are
responsible for complying with the upstream terms.
