ONNX exports of TabICLv2, KumoTabular (small, medium, large), KumoRelational, Chronos-2 and Kumo-Anomaly for Tabular Data Workbench
b043bdf verified | license: openmdw-1.1 | |
| library_name: onnx | |
| base_model: nvidia/Kumo-Relational | |
| base_model_relation: quantized | |
| pipeline_tag: tabular-classification | |
| tags: | |
| - onnx | |
| - onnxruntime-web | |
| - webgpu | |
| - relational-foundation-model | |
| - structured-data-models | |
| - tabular-classification | |
| - tabular-regression | |
| - in-context-learning | |
| # KumoRelational for onnxruntime-web | |
| An unofficial ONNX conversion of NVIDIA's [Kumo Relational](https://huggingface.co/nvidia/Kumo-Relational) | |
| (revision `v1.0.1`), a foundation model for classification and regression | |
| across related tables, made for Tabular Data Workbench, which runs it in the | |
| browser with onnxruntime-web on WebGPU or WebAssembly. The weights are | |
| NVIDIA's: unchanged in the `fp32` variant, and with the larger ones rounded to | |
| float16 in the `f16w` variant. NVIDIA has not reviewed or endorsed this | |
| conversion. | |
| ## Files | |
| Each head (`classifier`, `regressor`) is cut into five graphs: | |
| | file | inputs | output | | |
| | --- | --- | --- | | |
| | `<head>-col.<v>.onnx` | `X` (columns, rows, 3), `y` (train), `keys` (keys) | `E` (columns, rows, 128) | | |
| | `<head>-row.<v>.onnx` | `E` (columns, rows, 128) | `Z` (rows, 512) | | |
| | `<head>-hop.<v>.onnx` | `x` (nodes, 512), `src`, `dst`, `edge_type` (edges), `noise` (types, 512), `deg` (nodes, 1) | `h` (nodes, 512) | | |
| | `<head>-out.<v>.onnx` | `x` (readouts, 512) | `z` (readouts, 512) | | |
| | `<head>-icl.<v>.onnx` | `Z` (tokens, 512), `y` (train) | `out` (queries, 10 logits or 999 quantiles) | | |
| `col` and `row` embed each table's rows, `hop` is one round of relational | |
| message passing (run once per hop), `out` is the readout and `icl` is the | |
| in-context block over the task's context and query rows. The graph neural | |
| network's segment reductions are written as ScatterND, and the graphs contain | |
| no int64 shape arithmetic. | |
| Each graph comes in two variants (`<v>`): | |
| - `fp32`: the checkpoint's weights as they are (244 MB for both heads). | |
| - `f16w`: every float32 weight of 1,024 or more elements stored as float16 | |
| and cast back to float32 on load (123 MB). Compute stays float32. | |
| `manifest.json` lists every file with its size and sha256, the graphs' | |
| inputs and outputs, each head's empty-row embedding and the parity figures | |
| below. The time-aware neighbour sampler, the task tables, the encoders and | |
| the ensemble stay outside the graphs; the app implements them in TypeScript. | |
| onnxruntime 1.30's CPU ScatterND with a reduction is not deterministic with | |
| more than one intra-op thread when indices repeat, which they do in the hop | |
| graph. Run it single-threaded where results must be reproducible. | |
| ## How they were made | |
| Source: `nvidia/Kumo-Relational` at `v1.0.1` (`classifier.pt`, | |
| `regressor.pt`), loaded through NVIDIA's | |
| [structured-data-models](https://github.com/NVIDIA/structured-data-models) | |
| at commit `2b1696f4fdc22004cc757c44703a69cf326c0a11`. | |
| ```sh | |
| node models/rfm/dump_shop.mjs # once, from the repository root: the sample tables | |
| cd models/rfm | |
| uv run export.py # writes public/models/rfm/ | |
| ``` | |
| `export.py` (in `models/rfm/` of Tabular Data Workbench) exports the parts | |
| defined in `exportable.py` with `torch.onnx` (dynamo, opset 18, PyTorch 2.14 | |
| on CPU) and removes the exporter's node metadata and doc strings. It checks | |
| each graph against PyTorch, then checks the five graphs chained against | |
| structured-data-models' own forward pass on inputs captured from a real run | |
| over the app's sample shop tables. | |
| ## Parity | |
| Chained graphs (onnxruntime 1.30, CPU, one thread) against | |
| structured-data-models' forward pass, maximum absolute difference: | |
| | head | task | fp32 | f16w | largest output | | |
| | --- | --- | --- | --- | --- | | |
| | classifier | churn (logits) | 8.6e-3 | 1.2e-2 | 10.6 | | |
| | regressor | spend (quantiles) | 1.6e-3 | 6.5e-3 | 4.9 | | |
| Each graph alone, on random inputs, matches PyTorch to 1.1e-5 (fp32). | |
| ## Licence | |
| The weights are released by NVIDIA under the OpenMDW License Agreement 1.1 | |
| (`LICENSE`, Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES). It allows | |
| use, modification and redistribution provided that the agreement and the | |
| notices of origin travel with the weights, and it ends the rights of anyone | |
| who sues claiming the model infringes a patent or copyright. | |
| Kumo Relational fine-tunes the TabICLv2 weights (Soda team, Inria), | |
| licensed under BSD-3-Clause; `THIRD-PARTY-NOTICES.txt` is NVIDIA's notice | |
| with that licence's full text. The graphs were exported from | |
| structured-data-models' model code, licensed under Apache-2.0 | |
| (`LICENSE.structured-data-models`). | |
| Keep `LICENSE`, `THIRD-PARTY-NOTICES.txt` and | |
| `LICENSE.structured-data-models` with the graphs when redistributing them. | |