Simon Boehm

I think an `march` parameter added to `Model.compile()` would make sense: - `march=native`: use host cpu architecture + all host cpu features. - `march=`: Use the specified architecture + all...

Note to self: I seems like the llvmlite pip packages only ship the backends required for the local arch, hence there's no way to set-up lleaves as a proper cross-compiler.

Yes, I tried building it for Windows a year ago, but I couldn't get it to run after trying for a few hours. There shouldn't be any major changes necessary,...

TLDR: It's possible that there's a bug that causes a segfault, though it's unlikely that this is happening in the parts of the code you're pointing to. For diagnosing the...

I'm having trouble understanding this issue. Could you write up a minimally reproducible example of the core dump / send me the `model.txt` that causes it?

I've "fixed" this here: https://github.com/siboehm/dgl-lifesci/tree/jtvae by setting the dtype of the whole graph to int64. Doubt that this is the root cause, but it's training fine now.

The pretraining script in this repo is still broken though. Shouldn't we keep this issue open?

Posting my HN comment here for reference: > I had a brief look at your package, and my impression was that it's only changing model training. If this is correct...

There's a `cache=` parameter for `lleaves.Model.compile()`. Does that do what you're looking for? See [docs](https://lleaves.readthedocs.io/en/latest/) for more info. I looked into supporting pickling a while ago, but the cache parameter...

Yeah, you're right, classmethod would be nicer. Currently, what's stored in the cache is an ELF file (on Linux), containing the compiled function. Recreating a `lleaves.Model` from the ELF file...