is wasi-nn.witx still relevant?

[yamt]

is it still possible to generate bindings (say, for C guest) from wasi-nn.witx?
if yes, how?
if no, is there an alternative which can be used for non component-model implementations like WAMR's wasi-nn?

[abrown]

At this point, the specification has moved on from what was specified in the old WITX document (here]. That document was retained to avoid issues with tools in various places that may rely on it. That was in August 2023, though, so it seems to me that now is a good time to remove it.

For a p1-only implementation, I'm not sure what path to recommend: I know that Wasmtime used p1-p2 adapter modules for some time and perhaps that approach could work here. But perhaps the state of the art in tooling is now in a place where something like that could be done more easily; I occasionally see people working in that space on Zulip. One challenge I can think of: WIT allows wasi-nn to pass around resources — not sure how that should work in a p1 version.

[yamt]

At this point, the specification has moved on from what was specified in the old WITX document (here]. That document was retained to avoid issues with tools in various places that may rely on it. That was in August 2023, though, so it seems to me that now is a good time to remove it.

For a p1-only implementation, I'm not sure what path to recommend: I know that Wasmtime used p1-p2 adapter modules for some time and perhaps that approach could work here. But perhaps the state of the art in tooling is now in a place where something like that could be done more easily; I occasionally see people working in that space on Zulip. One challenge I can think of: WIT allows wasi-nn to pass around resources — not sure how that should work in a p1 version.

doesn't wasmtime use two separate bindings?
crates/wasi-nn/src/wit.rs
crates/wasi-nn/src/witx.rs

in our case, currently both of apps and the engine are based on p1.
so we at this point merely want to have a way to generate necessary info for bindings (eg. C headers to describe the abi or equivalent) for p1.

[abrown]

In Wasmtime, the WITX functionality is only retained for legacy reasons. We were actually discussing removing it recently. It's not a great reason to keep this WITX around.

My opinion on what to do for WAMR is to explore a way to make the WIT work.

[yamt]

In Wasmtime, the WITX functionality is only retained for legacy reasons. We were actually discussing removing it recently. It's not a great reason to keep this WITX around.

did the legacy reason went away?
anyway, i guess some of other engines are more conservative.

My opinion on what to do for WAMR is to explore a way to make the WIT work.

maybe. but i have no idea how long it takes.

also, wamr is not only implementation using witx.

[abrown]

did the legacy reason went away?

I wouldn't say there are too many p1 users out there for wasi-nn, but it would appear you want to use it, so there's that! Again, I would caution against depending on this WITX too much, it truly is deprecated now that the spec has moved forward.