Toto-2.0-22m (int8 ONNX)
int8 ONNX export of Datadog/Toto-2.0-22m (Apache-2.0), quantized as: blocked weight-only QInt8 (opset-21 blocked DequantizeLinear, 16-row blocks); output head kept fp32. Requires onnxruntime >= 1.20; untested under onnxruntime-web.
Unofficial export, not affiliated with or endorsed by the model authors.
Contract
| input | dtype | shape | notes |
|---|---|---|---|
context |
float32 | ['variates', 2048] |
NaN |
series_ids |
int64 | ['variates'] |
variates sharing an id are forecast jointly through variate-axis attention; distinct ids = independent |
Output quantiles: float32 ['variates', 96, 9]. Quantile levels: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9].
One parallel pass to 96 steps. Left-pad short series with NaN (missing-aware). Median is level index 4.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for sktime/toto2-22m-onnx-int8
Base model
Datadog/Toto-2.0-22m