How to incrementally predict for incremental data?

[superlomo]

When new Kline data becomes available, I push it to the local dataset file.
However, how can I incrementally predict the new coming data instead of recreating the whole relevant qlib instances?
I know qlib has some caching mechanisms, such as D.calendar or D.feature.
Can I do just like this? I'm unsure if qlib requires any warmup bars.

ds = DatasetH(handler, segments={"test": (new_start_date, new_end_date)})
pred = self.model.predict(ds)

[superlomo]

I know there is an online mode of qlib, but i think it a little too heavy for me. Because i just download data by myself and just needs the new scores

[superlomo]

I checked this: #951
But is it possible to incrementally predict for incremental data without to clear the mem_cache and reload the whole dataset