[Remyx Recommendation] Add DiG (Gated Linear Attention) transformer backbone alongside DiT #13

Model/Pipeline/Scheduler description

DiG (Diffusion models with Gated Linear Attention) swaps DiT's quadratic self-attention for Gated Linear Attention (GLA), giving the diffusion transformer sub-quadratic cost on long sequences while keeping the same 2D backbone and forward contract as DiTTransformer2DModel. This lines up directly with two active themes in the repo's trajectory: efficient attention (diffusers already ships SanaLinearAttnProcessor2_0 as a linear-attention backbone, an implicit acceptance signal for this direction) and inference efficiency / scalability. It is a clean architectural sibling of DiTTransformer2DModel, so the integration surface is well understood.

Open source status

• The model implementation is available.
• The model weights are available (Only relevant if addition is not a scheduler).

The reference implementation is open source (MIT), but no DiG checkpoint is currently exposed on the Hub in diffusers format, and the repo has no path to obtain one. DiG's value is as a runnable, pretrained backbone; a model class with no weights and no pipeline that instantiates it is a scaffold, not a usable capability. Whether to ship architecture-only is one of the open decisions below.

Provide useful links for the implementation

• Paper: DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention (arXiv:2405.18428)
• Reference implementation: https://github.com/hustvl/DiG
• License: MIT (permissive — safe to adopt)
• Authors: the hustvl (HUST Visual Learning) group

Why this was opened as an Issue rather than a PR

• No scripts/ call site. Every file under scripts/ (93 of them) is a checkpoint-conversion utility (convert_*_to_diffusers.py, e.g. convert_dit_to_diffusers.py). There is no inference, eval, or scoring entry point there that would invoke a new attention backbone, so there is no natural home for a small, self-contained edit.
• The real integration lives in src/diffusers/models/, and is multi-file. A faithful, registered DiG would mean (1) a new DiGTransformer2DModel mirroring DiTTransformer2DModel in src/diffusers/models/transformers/, (2) a new GLA attention processor alongside SanaLinearAttnProcessor2_0 in src/diffusers/models/attention_processor.py, (3) registration in models/__init__.py + models/transformers/__init__.py (+ a dummy_* object), and (4) a model test under tests/models/transformers/. Without (3) the model is unreachable from any pipeline and is effectively freestanding — deleting it would change nothing in the repo. This is several files across src/, each needing registration rather than a 3-line hook, which is beyond the small-slot-in shape this pass targets.
• No checkpoint to load (see open-source status above).

Open questions to decide first

1. Is a DiG checkpoint forthcoming on the Hub? If yes, the integration becomes a standard "add model + converter + pipeline" task (sibling to the recent Anima/AnyFlow/Cosmos 3 additions). If no, we'd be shipping architecture without weights — worth deciding whether that's wanted on its own.
2. GLA as a new model vs. an optional attention path on the existing DiT block. Do we want a standalone DiGTransformer2DModel, or would extending DiTTransformer2DModel with a linear_attn/GLA variant in its attention processor set be preferable (closer to how Sana exposes SanaLinearAttnProcessor2_0)? This decides whether it's a new file or a flag on an existing one.
3. Processor placement. Confirm the GLA processor should register next to SanaLinearAttnProcessor2_0 in attention_processor.py and follow that processor's dataclass/__call__ contract — the cleanest existing precedent to mirror.

Drafted by Outrider — paper: arXiv:2405.18428.