Janus-Pro-7B-ONNX-WebGPU

The text-to-image part of deepseek-ai/Janus-Pro-7B (converted from deepseek-community/Janus-Pro-7B, the same weights in transformers' format), converted to ONNX for running in a web browser with ONNX Runtime Web's WebGPU backend, as used by Quick Inference.

Changes from the original

• Split into five ONNX graphs: language_model (input embeddings → hidden states, with a KV cache), embed_tokens, gen_head, gen_img_embeds and image_decode (the VQ decoder). The vision encoder used for image understanding is left out.
• The language model's linear layers are quantised to 4 bits (round to nearest, block size 32); its activations and KV cache are float16. embed_tokens, gen_head and gen_img_embeds are float16; image_decode is float32.

These are modified files: their outputs are close to, but not identical to, the original model's.

Licence and use restrictions

The code of Janus is MIT-licensed (LICENSE-CODE); the model weights are licensed under the DeepSeek License (LICENSE-MODEL.md), as the original. Its use-based restrictions (section 5 and Attachment A) apply to these files and to anything made from them: you may not use them for the uses listed in Attachment A, and if you distribute them or a derivative you must include those restrictions as an enforceable provision and give recipients a copy of the licence.