Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hey. Do you need to use the saved models in a distributed setting as well? We're currently working on adding a to_local method to the distributed mlforecast that would return a regular MLForecast instance. Would that work for your use case?

In the meantime you should be able to serialize it by removing the engine, which for dask is a dask client, e.g.
```fcst = DistributedMLForecast(...)
fcst.fit(...)
del fcst.engine
with open('fcst.pkl', 'wb') as f:
    cloudpickle.dump(fcst, f)```
You'd need to attach a client when loading it back, e.g.
```with open('fcst.pkl', 'rb') as f:
    fcst = cloudpickle.load(f)
fcst.engine = Client(...)```

<@U03UX9KR75W> this solution works very well for us, thanks for a lot.

<@U069Z6YJ4TV> Just wanted why are you deprecating darts again? Are the darts models not scalable enough for you or they don't perform well? Or both?