WebGPU readSync is only available for CPU-resident tensors.

[lebron8dong]

I trained a model on the browser and used webgpu to load the model and reported an error：I trained a model on the browser and used webgpu to load the model and reported an error

[qjia7]

@lebron8dong Can you help to give an accessible case so that we can reproduce and debug this issue?

[lebron8dong]

@lebron8dong Can you help to give an accessible case so that we can reproduce and debug this issue?
@qjia7

https://codelabs.developers.google.com/codelabs/tfjs-training-classfication?hl=zh-cn#0 This is the official example, I reproduced the model, but reported the error when using webgpu as the backend. I used cpu and webgl to train the model respectively, but when I used webgpu to load the model, both reported the same error：WebGPU readSync is only available for CPU-resident tensors.

[qjia7]

Can you directly share the complete webgpu case code so that I can try it locally? Or a very small case to reproduce your issue?

[lebron8dong]

Can you directly share the complete webgpu case code so that I can try it locally? Or a very small case to reproduce your issue?
@qjia7
https://github.com/lebron8dong/tfjs This is the folder for training the model and loading the model with webgpu.

[lebron8dong]

Can you directly share the complete webgpu case code so that I can try it locally? Or a very small case to reproduce your issue?

@qjia7 source of error：
// TODO: Remove once this is fixed:
// #1595
override readSync(dataId: object): BackendValues {
const tensorData = this.tensorMap.get(dataId);
const {values} = tensorData;

if (values == null) {
  throw new Error(
      'WebGPU readSync is only available for CPU-resident tensors.');
}

return values;

}

[qjia7]

@lebron8dong Thanks for your information. I can reproduce it now. The main problem is that webgpu itself doesn't support synchronous read. However, there are still several places using dataSync in tfjs framework. The error in your case is triggered by https://github.com/tensorflow/tfjs/blob/master/tfjs-vis/src/util/math.ts#L290. I will check whether I can remove dataSync and use something else or workaround it in webgpu backend.

[lebron8dong]

@lebron8dong Thanks for your information. I can reproduce it now. The main problem is that webgpu itself doesn't support synchronous read. However, there are still several places using dataSync in tfjs framework. The error in your case is triggered by https://github.com/tensorflow/tfjs/blob/master/tfjs-vis/src/util/math.ts#L290. I will check whether I can remove dataSync and use something else or workaround it in webgpu backend.

@qjia7
Thank you very much, I will wait for your good news.

[lebron8dong]

@qjia7 There is still the same error in my demo. Does my demo code need to be changed?

[qjia7]

You need to wait for the next release of tfjs-vis. Or build it by yourself from the source code https://github.com/tensorflow/tfjs/tree/master/tfjs-vis. I tried it locally. The next issue you may meet is platform.isTypedArray is not a function. After some workarounds to void #7273, it works on webgpu.

[lebron8dong]

@qjia7 thank you very much,thank you very much.

[lebron8dong]

@qjia7
const res = model.predict(test);
res. print()；
reported the same error：WebGPU readSync is only available for CPU-resident tensors.
Under the backend of webgpu, I want to print the result on the console, is this possible?

[qjia7]

@lebron8dong WebGPU only supports async read. Maybe try const data = await res.data(); console.log(data)