LibreWeVisOneVL - third-party attribution
=========================================

The WeVisOne-VL family loads its Qwen3-VL backbone through the native
Hugging Face transformers classes and ports the inference path of the
upstream pixel (alpha-matte) decoder.

    Project:  WeVisOne-VL
    Upstream: https://github.com/Tencent/WeVisOne-VL
    Commit:   9a6bcd85404505df0d05cfb5d97512ccacf35dee
    License:  Apache License 2.0 (the repository LICENSE prefixes the
              Apache-2.0 text with a Tencent preamble; it adds no terms)
    Copyright (C) 2026 Tencent. All rights reserved.

Ported or adapted files (each keeps an SPDX header with this provenance):

  * mask_decoder.py   <- wevisone_vl/mask_decoder.py, config.py, decoder.py
  * feature_bridge.py <- wevisone_vl/feature_bridge.py
  * prompts.py        <- wevisone_vl/prompts.py, engine.py (system prompt)
  * parsing.py        <- evaluation/utils/parser.py,
                         engine.py (parse_seg_boxes)

LibreYOLO adaptations: inference-only (dropout and training-time
initializers removed), single-image decoding, and the decoder's deformable
cross-attention reuses LibreYOLO's Apache-2.0 Deformable-DETR port
(libreyolo/models/deformable_detr/ms_deform_attn.py) instead of vendoring
upstream wevisone_vl/msda.py. All instances of a response share one
query-transformer pass, and only the per-instance refinement is chunked to
bound memory. Module and parameter names are unchanged so the
released mask_decoder.pt loads strictly. The vLLM backend, CLI, Gradio demo,
visualization and evaluation scripts are not ported.

Model weights are not redistributed by LibreYOLO. They download at runtime
from the upstream Hugging Face repositories, pinned to:

    tencent/WeVisOne-VL-2B  798ffc7c87ad6104410812b589e240ba32d92315
    tencent/WeVisOne-VL-4B  4a7bed92750ac9d8b8a124cede14759036932dc2

Both are Apache-2.0, Copyright (C) 2026 Tencent, and are fine-tuned from
Qwen3-VL-2B-Instruct / Qwen3-VL-4B-Instruct (Apache-2.0, Copyright 2024
Alibaba Cloud), as stated in the LICENSE file of each repository. The Qwen3-VL
architecture runs from transformers (Apache-2.0); no Qwen code is copied into
LibreYOLO.
