TSFM-ai/chronos-bolt-tiny-onnx-browser

Chronos-Bolt Tiny for browser forecasting

Complete forecast graph exported from amazon/chronos-bolt-tiny revision a0e552de83495b5c28c14c71c374f3e33280b340. Weights and model architecture are Amazon's; TSFM supplies this browser export.

Contract

Input: context, float32 [1,512], one univariate series. Retain the latest 512 values; left-pad shorter histories with NaN. Missing observations may be NaN. Output: quantile_preds, float32 [1,9,64], quantiles 0.1 through 0.9 in ascending order. Outputs are already in the original series scale. Horizon is at most 64; median is index 4. No covariates or sampling controls in this artifact.

The complete encoder and decoder are included. This is not an embedding proxy or statistical replacement. Browser execution uses ONNX Runtime Web WASM on the visitor's CPU; no TSFM server inference is involved.

Validation

CPU ONNX output compared with untouched upstream PyTorch across seven synthetic cases: trend/seasonality, constant, zero, negative, random, short and missing histories. Maximum absolute error was 0.000031. One real Chrome WASM run differed by at most 0.000046. These are local numerical checks, not broad device compatibility or performance guarantees. See reference fixtures and export source.

Reproduction

Run export.py with Python 3.12, chronos-forecasting 2.3.0, torch 2.8.0, transformers 4.57.1, onnx and onnxruntime. Export adjustments replace unsupported nanmean with equivalent masked reductions, nonoverlapping unfold with reshape, and enforce integer embedding indices. The graph has a fixed 512-step context.

Upstream: https://github.com/amazon-science/chronos-forecasting Model: https://huggingface.co/amazon/chronos-bolt-tiny