Brian Jones (Dec 02 2021 at 17:45): Brian Jones (Dec 02 2021 at 17:45):I'm adding another backend (Tensorflow) to the wasi-nn crate in Wasmtime. One of the calls it makes requires a path to some files the .wasm app is using. I can open the files fine on the client side using fs:read("mapdir_name"). But when I want to read those files on the Host side, I'm not clear on how to do this. More plainly, I have a web assembly example I'm running using --mapdir to specify where my files are located. I want the app to pass that mapped folder name to wasi-nn and be able to open the files on the wasi-nn side. Is this currently doable, and if so what is the best way to accomplish this? Thanks.
Till Schneidereit (Dec 02 2021 at 19:07):CC @Dan Gohman
Dan Gohman (Dec 02 2021 at 19:10):Is the problem that the path requires the --mapdir mapping to resolve, and you need that on the host side?
Brian Jones (Dec 03 2021 at 00:51):Correct. I need the mapdir mapping to resolve on the host side.
Dan Gohman (Dec 03 2021 at 00:52):Does wasi-nn have an API where it takes a filename?
Dan Gohman (Dec 03 2021 at 00:53):Or are you opening the file and feeding the data into wasi-nn manually?
Brian Jones (Dec 03 2021 at 00:54):This is for a new wasi-nn backend, so new code. Specifically I'm trying to use this function 'load' https://tensorflow.github.io/rust/tensorflow/struct.SavedModelBundle.html#method.load
Brian Jones (Dec 03 2021 at 00:56):that function requires a path, and I want to know how to properly (if possible) get that from a wasi-nn call
Dan Gohman (Dec 03 2021 at 00:57):ah, I see.
Brian Jones (Dec 03 2021 at 00:57):Yeah if load could just take a byte array I would do it the same way we do for OpenVino
Brian Jones (Dec 03 2021 at 00:58):Open on the client side and pass bytes to wasi-nn
Dan Gohman (Dec 03 2021 at 00:58):Yeah. And in the future, one could imagine passing in a bytestream to wasi-nn so that it could stream the data in.
Dan Gohman (Dec 03 2021 at 00:59):Stream it in directly from a network source, or something else. Filesystem paths tie things to the filesystem, which is so limiting ;-)
Brian Jones (Dec 03 2021 at 00:59):absolutely :)
Dan Gohman (Dec 03 2021 at 00:59):I don't think we currently have a way to do a mapdir mapping in host code.
Dan Gohman (Dec 03 2021 at 01:00):We could add one, but it's very tempting to see if we can find a way to do this without having to.
Brian Jones (Dec 03 2021 at 01:01):I certainly understand it could potentially break some sandboxing rules
Dan Gohman (Dec 03 2021 at 01:01):Do you think tensorflow would be open to a PR adding an alternate form of load?
Brian Jones (Dec 03 2021 at 01:02):I don't know, that was the alternative I discussed with Andrew. Pushing a PR that added a load that would accept a byte array
Brian Jones (Dec 03 2021 at 01:02):its not a complex function - so might not be hard to do (knock on wood)
Dan Gohman (Dec 03 2021 at 01:05):Yeah. It would make the sandboxing situation much simpler, as well as the deployment situation. Some WASI environments don't have filesystems, such as Compute@Edge
Brian Jones (Dec 03 2021 at 01:06):Oh interesting, yeah that certainly wouldn't work otherwise
Dan Gohman (Dec 03 2021 at 01:07):Looks like load is a wrapper around a tensorflow_sys function
Dan Gohman (Dec 03 2021 at 01:09):Hrm. It calls TF_LoadSessionFromSavedModel
Dan Gohman (Dec 03 2021 at 01:13):This is not going to be trivial. The path is a directory path, and tensorflow expects to find multiple files under that path.
Dan Gohman (Dec 03 2021 at 01:16):It's not just a byte array, it's a whole tree of different things.
Dan Gohman (Dec 03 2021 at 01:16):Ok. I have to go offline now, but I'll be around tomorrow; if possible, let's sync up and talk about possible options here.
Brian Jones (Dec 03 2021 at 01:17):Ah yes, its a path to a dir
Brian Jones (Dec 03 2021 at 01:17):Sure thing thanks Dan
Dan Gohman (Dec 03 2021 at 20:44):@Brian Jones I think what makes sense for the present is to add a way to get at the mapdir mappings in the host, for now.
Dan Gohman (Dec 03 2021 at 20:45):I expect we'll eventually have better features for dealing with kind of thing, in wit-bindgen, but to get something up and running right now, just using the existing tooling, and the existing Tensorflow API, makes sense.
Brian Jones (Dec 03 2021 at 21:03):Ok sounds good. Is that something I can do and submit a PR? Or should I submit a bug/feature request and let the appropriate person handle it?
Brian Jones (Dec 03 2021 at 21:19):@Dan Gohman Oh sorry I read that a a little more carefully and sounds like its ok If work on this. Let me know if that's not the case, not trying to break things :)
Dan Gohman (Dec 03 2021 at 21:21):Yes, to clarify, it's ok to work on this :-)
Brian Jones (Dec 03 2021 at 21:23):Excelent thanks agian
Last updated: Nov 22 2024 at 16:03 UTC