tabular.cpp

C++ inference for tabular ML models — GGUF-native, zero Python dependency.

Embed XGBoost, LightGBM, TabPFN, and FT-Transformer into any C++ application — IoT firmware, PLC controllers, mobile apps, trading systems — without pulling in a Python runtime or separate ML library.

Model	Type	Edge?	tabular.cpp	sklearn/xgb
XGBoost	Gradient boosted trees	✅	✅	Heavy dep
LightGBM	Gradient boosted trees	✅	✅	Heavy dep
TabPFN v2	In-context transformer	—	✅	❌
FT-Transformer	Feature tokenizer + attn	✅	✅	❌

Built on ggml — same backend as llama.cpp and whisper.cpp. One GGUF format for all model types.

---

Quick start

git clone https://github.com/liodon-ai/tabular.cpp
cd tabular.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

---

Models

XGBoost / LightGBM — Tree Ensembles

Convert your existing models in seconds. No accuracy loss — exact same tree traversal.

# Convert
python convert/convert_xgboost.py  --model model.json   --output model.gguf
python convert/convert_lightgbm.py --model model.txt    --output model.gguf

# Predict
./build/tb-predict -m model.gguf -i features.csv -o predictions.csv

# From C
struct tb_context * ctx = tb_init("model.gguf", tb_default_params());
int n_out;
float * pred = tb_predict(ctx, X, n_rows, &n_out);  // [n_rows, n_out]

Why not just use xgboost's C++ library?

• xgboost.so is 30 MB; tabularcpp + ggml is smaller and already linked if you use llama.cpp
• GGUF format is consistent with your other models
• NaN/missing value handling baked in via default_left per split node
• Multiclass softmax, binary sigmoid, regression — all handled

Flat-array tree representation (cache-friendly, no pointer chasing):

trees.feature      int32[n_nodes]   -1 = leaf
trees.threshold    f32[n_nodes]
trees.left_child   int32[n_nodes]   node index
trees.right_child  int32[n_nodes]
trees.default_left int8[n_nodes]    1 = NaN goes left
trees.leaf_value   f32[n_nodes]
trees.roots        int32[n_trees]

---

TabPFN v2 — In-context Learning

Zero training. Pass your labeled examples as context; get predictions immediately. Beats XGBoost on datasets with < 10k rows on most benchmarks.

python convert/convert_tabpfn.py --model priorlabs/tabpfn-v2 --output tabpfn-v2.gguf

./build/tb-predict \
    -m tabpfn-v2.gguf \
    --train train.csv --label-col 0 \
    -i test.csv -o predictions.csv

struct tb_context * ctx = tb_init("tabpfn-v2.gguf", tb_default_params());
float * proba = tb_predict_pfn(ctx,
    X_train, y_train, n_train,
    X_test,  n_test,
    n_features, n_classes);
// proba: [n_test, n_classes] softmax probabilities

Architecture: Feature tokenizer → 12-layer transformer encoder → classification head. In-context: training samples and test samples attend to each other in one forward pass.

---

FT-Transformer — Neural Tabular

Feature Tokenizer + Transformer. Handles mixed numeric + categorical features natively.

python convert/convert_fttransformer.py \
    --checkpoint ft_transformer.pt \
    --output ft_transformer.gguf

./build/tb-predict -m ft_transformer.gguf -i features.csv -o preds.csv

Each feature gets its own linear projection to d-dimensional token space. The transformer then reads all feature tokens, and a CLS token's output drives the prediction head.

---

Edge deployment use cases

Predictive maintenance — XGBoost/LightGBM models running on ARM PLCs in factories with no internet connectivity. No Python, no xgboost library, just a static binary.

Medical wearables — Risk scoring on heart rate + accelerometer data. FDA class II devices require on-device inference. tabular.cpp compiles to bare-metal ARM Cortex-M.

POS fraud detection — Transaction feature scoring in < 5ms, fully offline. Card data never leaves the terminal.

Automotive ECUs — OBD-II sensor anomaly detection in the vehicle. C++ only, strict memory budgets.

---

GGUF metadata keys (tb.* namespace)

Key	Type	Description
tb.arch	str	xgboost / lightgbm / tabpfn / ft_transformer
tb.model_type	i32	Enum (1=XGBoost, 2=LGBM, 3=TabPFN, 4=FTT)
tb.objective	str	regression / binary / multiclass
tb.n_trees	i32	Total tree count (trees)
tb.n_features	i32	Input feature count
tb.n_classes	i32	Output class count
tb.n_nodes	i32	Total node count across all trees
tb.base_score	f32	Initial prediction (XGBoost)
tb.d_model	i32	Transformer hidden dim (TabPFN/FTT)
tb.n_layers	i32	Transformer layer count
tb.max_features	i32	Max features supported (TabPFN)
tb.max_context	i32	Max training samples in context (TabPFN)

---

Roadmap

• CatBoost converter (native categorical support)
• NODE (Neural Oblivious Decision Ensembles)
• ONNX import for any sklearn-compatible model
• AVX-512 tree traversal for x86 edge servers
• Quantized trees (int8 thresholds, int16 leaf values)
• Python bindings (ctypes wrapper)

---

License

MIT. Built on ggml (MIT).