sana-sprint-text

The Gemma-2-2B-IT prompt encoder of SANA-Sprint 0.6B, prepared for on-device use in Latent Studio:

text/prefix_kv.bin, text/prefix_hidden.bin: Gemma keys/values and hidden states of SANA's fixed ~200-token instruction prefix, precomputed (fp16), so the phone only runs Gemma over the user's prompt.
text/gemma_tail.onnx (+ .data): Gemma-2 decoder over the prompt tokens attending to that prefix; weight-only int8 MatMuls (block 32). Inputs inputs_embeds [1,T,2304], past_k/past_v [26,1,4,P,256], start; output final hidden states.
text/embed_q8.bin + embed_scale.bin: the token embedding table, int8 per row.
text/gemma_tokenizer.bin: Gemma's BPE tokenizer compiled to a compact binary.

Verified against the diffusers SanaSprintPipeline: embeddings cosine 0.9999 (int8), and the same images seed for seed.

License

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms, including the Gemma Prohibited Use Policy. By using these files you agree to them.

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Base model

Finetuned

Quantized

(206)

this model