t0-web
A from-scratch port of a probabilistic time-series forecasting transformer to the browser.
Own WGSL implementation of The Forecasting Company’s t0 time-series forecasters, ported to Burn and wgpu for native and browser use. Weights quantize to GGUF Q8_0 and Q4_0 (t0-alpha-q4_0-webgpu, t0-alpha-q8_0-webgpu, t0-beta-q4_0-webgpu, t0-beta-q8_0-webgpu), match the PyTorch reference to within 1.2e-6, and run in Chrome via WebGPU. A compare page runs this engine next to the official ONNX INT8 export side by side in one tab.
Includes
theforecastingcompany/t0-alpha, by The Forecasting Company
Q4_0 quantization for WebGPU by Idle Intelligence.
Apache-2.0
modeltheforecastingcompany/t0-alpha, by The Forecasting Company
Q8_0 quantization for WebGPU by Idle Intelligence.
Apache-2.0
modeltheforecastingcompany/t0-beta, by The Forecasting Company
Q4_0 quantization for WebGPU by Idle Intelligence.
Apache-2.0
modeltheforecastingcompany/t0-beta, by The Forecasting Company
Q8_0 quantization for WebGPU by Idle Intelligence.
Apache-2.0
- Runs
- browser, WASM, WebGPU, native
- Demo (GitHub Pages)
- https://idle-intelligence.github.io/t0-web/web/
- Source
- idle-intelligence/t0-web
- HuggingFace
- idle-intelligence/t0-alpha-q4_0-webgpu
- Perf
- 34 ms per forecast (Q8_0, single signal, 512 context) in Chrome on an Apple M2; 30 ms with Q4_0
- Status
- maintained
Learn more
Used on: forecasting