WASM: External initializers are not enabled in this build

[decahedron1]

          Hi

@decahedron1

Build and load of wasm module in the browser is successful.

But when inferencing the model, I face following error:

panicked at libtashkeel-wasm\src\lib.rs:23:34:
called `Result::unwrap()` on an `Err` value: InferenceError("Failed to run model using onnxruntime via ort. Caused by Failed to create ONNX Runtime session: Deserialize tensor attn_layers.0.ccm.layernorm.bias failed.External initializers are not enabled in this build

I'm using an ort model.

Originally posted by @mush42 in #75 (comment)

[decahedron1]

@mush42 Are you able to upload the converted .ort model?

[decahedron1]

Should be fixed by 80be206

[mush42]

@decahedron1
Here's the model. It is a quantized model using neural compressor.
model.zip

[decahedron1]

Can you confirm that 80be206 fixed this issue by setting ort as a git dependency:

[dependencies]
ort = { git = "https://github.com/pykeio/ort.git", rev = "80be2068296b04e8ae0a442d4b9c2d635362d07e" }

[mush42]

@decahedron1 I confirm that the commit fixes the issue.

[mush42]

@decahedron1
If possible, make a new rc release.

[mush42]

My model is now live at:
https://mush42.github.io/libtashkeel/

[mush42]

Do you have any idea of what has been added to nightly to make ort work in wasm-unknown-unknown?

Just came by this:
https://github.com/Liamolucko/find_wasm_import

[decahedron1]

No idea. My best guesses are rust-lang/rust#124036 or the LLVM 18 upgrade (which should land in v1.78 stable, releasing tomorrow).

[mush42]

OK.

Whatever it is, it seams to add around 5-6 MB to the .wasm module bundle.

Using stable, my build is around 6 MB when compiled with stable, jumped to 12 MB when compiled with nightly.