kevin-mi/VDN-H3-overlay

VDN-H3 SGLang model overlay

Metadata-only overlay that lets SGLang Diffusion serve OpenVDN/vdn-minimax-h3 (Video DeltaNet MiniMax-H3: hybrid chunked window-softmax + frame-wise linear attention, 8-NFE DMD2 distill) as a VDNH3Pipeline.

There are no weights in this repository. SGLang downloads the VDN deltas (stage-dmd-step-250/, the linear branch and two LoRA adapters) plus the h3-base/ Diffusers export from the source repo, links the Qwen3-VL conditioner from MiniMaxAI/MiniMax-H3, and materializes a base-H3-layout directory under SGLANG_DIFFUSION_CACHE_ROOT/materialized_models/ with:

• both LoRA adapters prefused into the transformer weights (W += B @ A, fp32 accumulate, scale 1.0 = alpha / rank), written once as new shards;
• the linear branch attached as transformer/sglang_vdn_linear_branch.safetensors;
• transformer/config.json carrying hybrid_attention (the VDN transform config);
• the video VAE re-serialized in the fused source form SGLang's native decoder loads, exactly as the FastH3 overlay does.

Usage:

export SGLANG_DIFFUSION_CACHE_ROOT=/path/with/100GB/free
sglang serve --model-path OpenVDN/vdn-minimax-h3 --num-gpus 4 \
  --attention-backend hybrid_window_attn_h3

See _overlay/materialize.py for the exact steps and the provenance record (transformer/sglang_vdn_prefused.json) it writes.

License

VDN-H3 inherits the MiniMax H3 Community License, including its territorial exclusions; read it before downloading or running the weights. This overlay repository carries metadata and Apache-2.0 code only.