Community port: native .NET inference (Tsfm.Forecasting on NuGet)

Not a bug report — just letting you know, in case it's useful to others...

Firstly, thank you for publishing the open source model and reference implementation.

I have published a native .NET port of the Kronos tokenizer and predictor, built on TorchSharp:

• Source: https://github.com/markhammond/tsfm-forecasting
• NuGet: Tsfm.Forecasting

It runs the BSQ tokenizer and the autoregressive decoder in-process (on CPU or GPU depending on TorchSharp backend), with no Python at build or run time. Parity verified (subject to floating point epsilon) against prediction_example.py.

Two deliberate departures, both documented:

• Sampling draws per-bar uniforms and selects by inverse CDF, so a draw does not depend on batch grouping. Distributionally identical to the reference; the stream is not.
• No key-value cache. Attention is recomputed per decode step, matching the reference arithmetic so parity stays checkable.

No action needed. Happy to open a PR if it would be worthwhile mentioning dotnet packages in your README, etc.

On licensing: everything stays MIT. Your architecture, BSQ tokenizer and reference implementation are duly credited, and the NeoQuasar checkpoints are redistributed unmodified. ‍♂️

[edited: package and repo link updated]