Andrew Brown

This change is probably not desirable if we adopt #74: if a user can't change where a _model_ resides any more, they shouldn't be allowed to change where a _tensor_...

@chongshenng, sounds like there are several options here: - if you're interested in getting something working soon-ish, you could go down the model conversion route and try to run your...

I think what tract is trying to do is compile the ML backend itself — the implementations of all the operators — into a WebAssembly module. (That is if I...

I had been thinking in terms of option 3 since I assumed the implementations of wasi-nn could use quite different formats, not just for encoding 16-bit floats but even for...

To support both formats and avoid implicit formatting (e.g. dependent on execution target), we could expand the tensor types to include both `$f16` and `$bf16`; currently [that enum](https://github.com/WebAssembly/wasi-nn/blob/master/phases/ephemeral/witx/wasi_ephemeral_nn.witx#L22) only contains...

> does wit/interface-types/etc provide a large enough alignment? I guess my question would "large enough" for what? There's a couple of directions this could go but I'm not exactly sure...

No, I don't think so. When I switched this repository over to WIT in #17, I kept the API the same, so unless I unwittingly made a mistake, there should...

> wit version of get-output returns the tensor record while witx version of it only returns tensor-data Oh, sorry, I will fix that. > i'm more interested in C implementation...

Looking up at the previous comments, I don't think I knew how to respond to this (and still am not clear). Let me take a stab at explaining how I...

cc: @brianjjones