AI and machine learning on the IC?

I didn’t code much in Rust on IC. If the canister is able to load the models from Rust plz give let me know lol :sweat_smile:

Sure, I will post the result here


Hi everyone, I have tested the pre-trained deep learning model named “albert2” on IC.
I add an onnx model into the contact folder and use Path lib to read it.
Then I add the tokenizer,json file for performing natural language tasks.
Here is my Cargo.toml:
candid = “0.7.14”
ic-cdk = “0.5.2”
ic-cdk-macros = “0.5.2”
tract-onnx = “0.19.2”
ndarray = “0.15.6”
tokenizers = “0.13.2”
tract_onnx: a library that can load the model built by pytorch, tensorflow, or jax.
ndarray; a library that can perform multi-dimensional tensor computation.

[target.‘cfg(target = “wasm32-unknown-unknown”)’.dependencies]
getrandom = { version = “0.2.6”, features = [“js”]}
The errors come out for the following aspects:

  1. random dependencies:
    When you deploy the canisters that needed the random something, you must add:
    getrandom = { version = “0.2.6”, features = [“js”]}
  2. head file error:
    This was what I meet here:

In one word, the first experiment failed.
But I already have some road maps for the next experiment, so I will continue to post my experiment result here.
So if anyone has some ideas, you can post them here and let me know, thx :grinning:
BTW, all code was written by rust.


Following your experiment. I think this is a domain the foundation should definitely invest in.


do u have a GitHub repo link?

As far as we practice, IC currently has a limitation for updating the var state because each round of consensus will end up creating a greater memory heap. Meanwhile, I am not so sure how canister handles the time-out problem for the execution. If training takes a while and computation is going on, how does the canister manage the call?

We have DTS now which helps a bit but is still limited. What you can do is repeated self-await because every await is a commit point and starts a new message (and therefore the execution limit). If you could e.g. perform one epoch of training in a single message then you can just await every epoch and you should be fine.


No, not training, just upload your model to IC for inference.

sorry, I delete it, but I will update new repo later.

I am also very interested in exploring the inference option using C++ compiled to wasm .

Hi all, so there is a question: for my experiment, there must be a basic random lib based on rust that can generate different kinds of distribution , but now the most rust random lib can not be compiled on IC. So I want to ask you guys who can recommend a simple random lib? We can build a basic IC random lib.
Here is the code from the open storage:
pub struct CanisterEnv {
rng: StdRng,

impl CanisterEnv {
pub fn new() → Self {
CanisterEnv {
// Seed the PRNG with the current time.
// This is safe since all replicas are guaranteed to see the same result of
// timestamp::now() and it isn’t easily predictable from the outside.
rng: {
let now_millis = time::now_nanos();
let mut seed = [0u8; 32];

impl Environment for CanisterEnv {
fn now(&self) → TimestampMillis {

fn caller(&self) -> Principal {

fn canister_id(&self) -> CanisterId {

fn random_u32(&mut self) -> u32 {

fn cycles_balance(&self) -> Cycles {


1 Like

Hi @cymqqqq, to get random value try calling raw_rand in ic_cdk::api::management_canister::main - Rust

Hi there, I’m glad to tell you guys that I have successfully deployed a word2vec canister model on IC, it’s my second machine learning experiment. Later on, I will post the github repo link.
Though it’s a toy version, I will continue to update it.


congrats, fantastic work!

1 Like

It’s a toy version machine learning canister, so everyone can deploy and test it.
BTW, it’s a shit version :rofl:

But, I will continually update a tensor library that can perform machine learning algorithms written in Rust like KNN, simple neural networks, etc… And I will post the link later. :laughing:
So everything started from the first code.
I hope the ecosystem of machine learning in IC will be more advanced in 2023.


Jesus Chris… that can’t be real. I am currently very packed, but I will try your code ASAP.

1 Like

If inference is possible (no training) on ic, that would be very mind blowing, because it means we can start to gently say good-bye to AWS’ lambda and Sagemaker​:wave::wave: bth I don’t believe it’s happening now.

1 Like

Hi, I understand your points, now I just write simple machine learning code to run.
If we want to perform more machine learning computation on IC, there is a lot of work to do now, the first is to write a basic tensor library, that’s what I am doing now.
Now it’s impossible to run large scale machine learning model on IC. :grinning:


Yes, it’s possible in theory, as follows:

  1. to transform pre-trained deep learning models into .wasm format(like onnx for .onnx format, keras for ,h5 format and the others),
  2. to put the wasm file inside a canister and deploy it on IC.

Now that someone got LLaMA and Whisper inferencing running on a Macbook M1, I’m hoping that these models can somehow be ported to the IC more easily.

1 Like