[Web] Running ORT model results in NaN values output

[nezaBacar]

Describe the issue

Hi, when I attempt to run inference of .ort model I turn on flags recommended in this issue: #13445 (comment).

const sessionOption = {
    executionProviders: ["wasm"],
    enableMemPattern: false,
    enableCpuMemArena: false,
    extra: {
      session: {
        disable_prepacking: "1",
        use_device_allocator_for_initializers: "0",
        use_ort_model_bytes_directly: "1",
        use_ort_model_bytes_for_initializers: "1"
      }
    }

Not turning them on results in an error, but turning them on results in an output of NaN values.

To reproduce

Run inference session with this model: https://drive.google.com/drive/folders/12tOtPWpANIlCPrsDMHSqkgxdhM0gwW88?usp=sharing and these flags:

const sessionOption = {
    executionProviders: ["wasm"],
    enableMemPattern: false,
    enableCpuMemArena: false,
    extra: {
      session: {
        disable_prepacking: "1",
        use_device_allocator_for_initializers: "0",
        use_ort_model_bytes_directly: "1",
        use_ort_model_bytes_for_initializers: "1"
      }
    }

Urgency

It is somewhat urgent

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.16

Execution Provider

'wasm'/'cpu' (WebAssembly CPU)

[fs-eire]

could you share what exact input data that you feed into the model, so that I can try to reproduce the problem

[nezaBacar]

Thanks for the fast response! I've added you to the demo repository. Just pop the model into the /demo/models folder and make sure that the line 20 in index.html points to model.ort. I noticed that if I try to run the ort model with these flags:
let sessionOptions = { executionProviders: [ 'wasm' ], graphOptimizationLevel: 'all', };
the outputs are ok. Should I maybe do it this way?
(thought this is not the right way because of this comment: github.com//issues/13445#issuecomment-1430153341)

[fs-eire]

I can reproduce the problem. And I checked the session options. If I remove use_ort_model_bytes_for_initializers: "1", the output value will be correct.

It takes a while for me to figure out why. From the comments in the source code, when specifying use_ort_model_bytes_for_initializers, it requires the model buffer keep valid during the whole life cycle of the inference session. Unfortunately this is not how onnxruntime-web manages the model data. Now, onnxruntime-web frees the model data after inference session is initialized. So using onnxruntime-web with config use_ort_model_bytes_for_initializers is not working.

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.