openmartin/CodeFormulaV2-ONNX-INT8

CodeFormulaV2 ONNX (INT8 decoder)

A mirror of the split CodeFormulaV2 ONNX assets published in the Docling Rust models-v1 release, derived from docling-project/CodeFormulaV2. The weights are mirrored unchanged. Model credit belongs to the upstream Docling authors; this repository provides downloadable ONNX files and does not claim to train or export them.

Only the autoregressive decoder is INT8. The vision encoder and token embeddings are FP32. The complete download, including the tokenizer, is 777.7 MB (decimal).

File	Purpose	Precision	Bytes
vision.onnx	Vision encoder	FP32	374,146,399
embed.onnx	Token embeddings	FP32	231,506,178
decoder_kv_int8.onnx	Autoregressive decoder with KV cache	INT8 weights	164,870,131
tokenizer.json	Tokenizer and special tokens	JSON	7,160,031

Usage

These are split graphs for a custom ONNX Runtime inference loop, not a drop-in Transformers pipeline() checkpoint. Encode image tiles with vision.onnx, embed the prompt tokens with embed.onnx, insert the image embeddings, and decode autoregressively with the decoder's KV cache. Preprocessing and prompts must match CodeFormulaV2/Idefics3.

The graphs have been exercised with ONNX Runtime Web 1.27.0 on WebGPU for formula recognition. This mirror is intended for direct browser model downloads; user images and PDFs are processed on the device and do not need to be uploaded to this repository.

Use immutable revision URLs and verify each file against manifest.json before execution. A successful decode does not guarantee an accurate or renderable formula; compare results with the source image. Code recognition is an upstream capability and was not evaluated here.

License and provenance

The upstream CodeFormulaV2 model is distributed under CDLA-Permissive-2.0. The complete license text is included as LICENSE. Original release asset URLs, byte counts, and SHA-256 hashes are recorded in manifest.json. This mirror does not change those files.