opticsWolf/texteller3-onnx

Texteller3 ONNX

Unified, self-contained ONNX serving repo for TexTeller formula OCR (ViT encoder + TrOCR decoder, VisionEncoderDecoderModel), built from the union of OleehyO/TexTeller and Ji-Ha/TexTeller3.

Why one repo

Archaeology (all verified, not assumed):

• Both repos carry identical weights: all 480 torch tensors equal, and the shipped ONNX weight hashes are equal. TexTeller3's ONNX trio is the same checkpoint through a newer optimum export vintage (node counts differ, weights don't).
• Tokenizers are semantically identical (same 11,500 vocab + 3,505 added tokens; only serialization formatting differs).
• Config diffs are version stamps only (transformers_version, _attn_implementation_autoset).
• TexTeller3 adds preprocessor_config.json (448px grayscale ViT); OleehyO alone ships decoder_model_merged.onnx — the file bobine's fp32 path loads.

So this repo takes TexTeller3's files as canonical and adds OleehyO's merged decoder verbatim (weights proven identical by hash, numerics proven identical by gate — see below). No re-export risk anywhere.

Files

File	Source	Notes
encoder_model.onnx	TexTeller3	fp32, opset 18, pixel_values → last_hidden_state [b,785,768]
encoder_model_dynamic-int8.onnx	Ji-Ha/TexTeller3-ONNX-dynamic (verbatim)	dynamic int8 encoder, pairs with the int8 merged decoder
decoder_model_merged.onnx	OleehyO/TexTeller (verbatim)	fp32, opset 11, 1-node If wrapper (use_cache_branch): no-past + with-past in one file
decoder_model_merged_dynamic-int8.onnx	Ji-Ha/TexTeller3-ONNX-dynamic (verbatim, onnx/ subfolder)	dynamic int8 (weights + activations), pairs with encoder_model_dynamic-int8.onnx
decoder_model.onnx	TexTeller3	fp32, no-past branch standalone
decoder_with_past_model.onnx	TexTeller3	fp32, KV-cache branch standalone
encoder_model_fp16.onnx	built here	pure-fp16 compute, fp32 I/O (entry+exit Casts), output names unchanged
decoder_model_merged_fp16.onnx	built here	same recipe, If branches included
model.safetensors	TexTeller3 (= OleehyO, proven equal)	torch provenance for re-export
tokenizer.json, vocab.json, merges.txt, …	TexTeller3	identical semantics to OleehyO's set
preprocessor_config.json	TexTeller3	448×448 grayscale, mean 0.9545 / std 0.1539
config.json, generation_config.json	TexTeller3	bos 0 / eos 2 / pad 1

Serving pair (bobine fp32 path): encoder_model.onnx + decoder_model_merged.onnx + tokenizer.json.

Parity gates (ORT 1.29, CUDA, 3 rendered formulas)

Check	Result
Merged (OleehyO) vs split decoder (T3), branch 0 + branch 1	worst|d| = 0.0, argmax agree
T3 shipped ONNX vs fresh optimum export from model.safetensors	worst|d| = 0.0
fp16 vs fp32, branch 0 / branch 1	worst|d| ≈ 0.01–0.02, argmax agree
fp16 greedy decode strings (32 steps, own cache loop)	exact on 3/3
q8 (MatMulNBits, incl. head-excluded variant) vs fp32	worst|d| 3.6–7.2, strings 1/3 — SHIPPED ON REQUEST, NOT RECOMMENDED (see below)
q8 encoder + fp32 decoder hybrid	diverges — rejected

The q8 files (encoder_model_q8.onnx, decoder_model_merged_q8.onnx) are published for experimentation only, not for serving: they fail the parity gate (see table) that the house rule requires for serving tiers. The KV-cache loop amplifies int8 noise past the point where greedy strings track, and excluding the LM head does not move the needle. If your decoding pipeline (beam search, repetition penalty, temperature) tracks fp32 output on your data, they cut the download substantially; until then, serve fp32/fp16. Recipe + numbers kept for reproducibility.

Use

import onnxruntime as ort
enc = ort.InferenceSession("encoder_model.onnx", providers=["CUDAExecutionProvider"])
dec = ort.InferenceSession("decoder_model_merged.onnx", providers=["CUDAExecutionProvider"])
hidden = enc.run(None, {"pixel_values": pixels})[0]   # [1,1,448,448] normalized per preprocessor_config
logits = dec.run(None, {"input_ids": [[0]], "encoder_hidden_states": hidden,
                         "past_key_values.0.decoder.key": empty, ..., "use_cache_branch": [False]})

Licence

Apache-2.0 (both upstreams). Formula OCR — image to LaTeX.