TinyCast (int8 ONNX)

int8 ONNX export of raws-labs/tinycast (Apache-2.0), quantized as: dynamic per-channel QInt8, MatMul-only; dilated convs and interface projections kept fp32 (measured recipe).

int8 verified within 4.1% of forecast spread of the official TinyCastPredictor.predict(), 15-block rollouts included (fp32 export within 0.0003%). Unofficial export, not affiliated with or endorsed by the model authors.

Contract

input	dtype	shape	notes
`context`	float32	`['batch', 2048]`	not supported - impute first (np.interp), left-pad with the FIRST value

Output quantiles: float32 ['batch', 48, 9]. Quantile levels: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9].

One 48-step AR block per call. For longer horizons feed the raw median (index 4) back into the context and rerun; sort each step's 9 deciles ascending before use (official TinyCastPredictor semantics).

See manifest.json for the machine-readable contract and the artifact's sha256.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sktime/tinycast-onnx-int8

Base model

raws-labs/tinycast

Quantized

(2)

this model