Kohaku architecture implementation adapted from KohakUwULLM.
Upstream: https://github.com/KohakuBlueleaf/KohakUwULLM
Source revision: d4a52f77c683e6a0622d76a4f42752d97ced97ba
Source: src/kohakuwullm/export/hf/modeling_kohaku.py
Copyright 2024 Shih-Ying Yeh; distributed under Apache License 2.0.

This project modifies the architecture to remove Transformers dependencies,
provide an explicit request-owned KV cache, integrate ComfyUI model management,
and expose native text generation. Upstream model weights are not bundled.

Additional changes include input-span/owner attention control in prefill and
decode, owner-aware MoE dispatch and repetition history, independent character
expansion with a shared caption, source-preserving tag guards, and RS pipe/UI
integration. The learned caption-to-tag format was consulted in KGen
(KohakuBlueleaf/KGen, Apache-2.0); KGen is not bundled as an inference dependency.

The model card for the pinned TIPO-v2.1 checkpoint identifies weight terms as
Kohaku License 1.0. Architecture attribution does not license the model weights.
See docs/references.md for exact sources, revisions and adaptation boundaries.

Gemma TIPO runtime and adapter/quantized loading:
RS-Seasonal-Prompt-Generator by raspie10032, revision 354cc9b.
Apache-2.0, see Seasonal-APACHE-LICENSE. Native inference and joint prompt
routing are modified for this extension. Packed GGUF decoding retains its
separate attribution in vendor/gguf_torch/NOTICE.md.

Gemma presampling-fresh-r8-v1 (independent RS adapter, 2026-11-05):
TIPO text presampling: TIPO authors, KohakuBlueleaf and KGen contributors.
Gemma: Google DeepMind; modified base by p-e-w.
Training prompts: KBlueLeaf/titpop-dataset (source card CC BY-NC-ND 4.0).
Korean localization: mengod-hf / 이오리네 and contributors (Korean contributions
CC BY-SA 4.0, with separate upstream notices retained on the model repository).
RS adapter contributions: CC BY-NC 4.0. Base and source terms remain applicable
to their materials; this code's Apache-2.0 license does not relicense them.
No weights or source datasets are bundled. Model card and retained notices:
https://huggingface.co/raspie/gemma4-tipo-lora-presampling-fresh-r8-v1
