OvisOCR2 (ONNX, WebGPU)
OvisOCR2 converted to ONNX for page OCR in the browser (WebGPU), with an exact speculative-decoding draft model. See NOTICE for the source models, the changes and the licences.
| File | What |
|---|---|
onnx/vision_encoder_fp16.onnx (+ data) |
Image reader, fp16 |
onnx/embed_tokens_q4.onnx (+ data) |
Token embedding, 4-bit |
onnx/decoder_q4_mtp3.onnx + onnx/decoder_model_merged_q4.onnx_data |
Decoder, 4-bit, with state window 3 and hidden-state output |
onnx/drafter_k2_s16384.onnx (+ data) |
2-token draft chain (retrained MTP layer, 16,384-token head) |
Prompt, decoding and page size follow the OvisOCR2 model card (greedy; about 1 megapixel per page is used by the app). The graphs need a patched onnxruntime-web build: LinearAttention state_window and chunked prefill on WebGPU.
- Downloads last month
- 27