A self-contained text-to-image pipeline at 3.2 GB: the Sana 1.6B transformer trunk runs ternary {-1, 0, +1} weights on real GemLite INT2 CUDA kernels, the Gemma-2 text encoder is bitsandbytes NF4 (4-bit), and the text encoder + VAE are otherwise the original Sana 1.6B components — all bundled so it loads in one call.

trust_remote_code=True is required: the transformer is a custom component (ClarkAirSanaGemliteTransformer) that builds the GemLite kernels on load. The 4-bit text encoder loads automatically from its embedded quantization_config.

The transformer's attention + GLU-FFN trunk is ternary, packed two bits per weight and re-quantized at each layer's largest power-of-two GemLite group (g64 attention / g32 FFN); the small "island" layers (embeddings, norms, modulation, depthwise conv) stay bf16. The transformer is 519 MB vs ~9 GB at fp16. The text encoder + VAE are the original Sana 1.6B weights — only the transformer is quantized (ternary) and the text encoder is NF4-compressed. The NF4 text encoder's 2.1 GB is dominated by the 256k-token embedding, which stays bf16 (quantizing it hurts caption quality).

The repo also carries clark_air_sana_gemlite_2bit.safetensors + load_clark_air_gemlite.py — a single-file packed transformer with a minimal loader, for when you already have a Sana pipeline and just want to swap in the gemlite trunk. The from_pretrained route above is the self-contained option.