WebGL backend crash on Samsung Galaxy devices after upgrading to Android 15 / One UI 7.0: "Failed to link vertex and fragment shaders"

[risteristevski]

System information

• I have written some custom code, but issue is somewhat reproducible on the mobilenet demo as well.
• Samsung Galaxy S23 & Samsung Galaxy S23 Ultra on Android 15 (One UI 7.0) - on both phones the issue started appearing after the One UI 7.0 upgrade.
• TensorFlow.js installed from npm ("@tensorflow/tfjs")
• TensorFlow.js version: 4.22.0
• Browser version: Chrome 136.0.7103.87
• Tensorflow.js Converter Version: 3.14.0 for our custom model, but also reproducible with the official mobilenet model which is converted with 1.2.10.1

Describe the current behavior
After upgrading both phones to One UI 7.0, our PWA started to crash consistently during the warm-up pass with the Failed to link vertex and fragment shaders. error message. Other devices, like iPhones or Google Pixel phones on Android 15 are not affected by this issue.

Describe the expected behavior
Model loading and prediction should succeed consistently without crashing the browser or producing shader linking errors.

Standalone code to reproduce the issue
Open https://codepen.io/Riste-Ristevski/pen/vEEqXBM on a Samsung Galaxy S23 with One UI 7.0

Other info / logs
While experimenting with the flags, I figured out that setting the WEBGL_PACK flag to false helps with the initial Failed to link vertex and fragment shaders. error. The warm-up pass is successful, and even most of the predictions seem to work (albeit a bit slower). However, sometimes, like in 1 out of 20 predictions, the browser crashes and shows a Unable to complete previous operation due to low memory. message upon restart. This happens also when testing with this demo.

Let me know if you need more information - I would be happy to help out however I can.

[shmishra99]

Hi @risteristevski ,

When attempting to run your example, I received the error: Failed to execute 'drawImage' on 'CanvasRenderingContext2D': The HTMLImageElement provided is in the 'broken' state.

From my debugging, it seems this is not a memory leak issue within the code. Instead, it seems the total memory required for the input tensor, model weights, and output tensor is exceeding the system's available memory.

Thank You!!!

[risteristevski]

Hi @shmishra99 ,

Thanks for looking into this issue. I assume that the problem with the example was caused by loading the photo from the link - I have now updated my example and embedded it as a base64 string directly - now you should be able to run the example.

When running the example on a Samsung Galaxy S23 with One UI 7.0, you should get a Failed to link vertex and fragment shaders. error. On devices/operating systems that are not affected, you should get a Prediction succeeded. message below the image.

[ttux]

Having the same problem when using face-api.js https://github.com/vladmandic/face-api

The same issue also happens with the samsung z flip 5.

Both the galaxy S23 have an adreno 740 GPU while the S24 doesn't seem to have the problem and it has an adreno 750 GPU.

[shmishra99]

Having the same problem when using face-api.js https://github.com/vladmandic/face-api

The same issue also happens with the samsung z flip 5.

Both the galaxy S23 have an adreno 740 GPU while the S24 doesn't seem to have the problem and it has an adreno 750 GPU.

Hi @ttux ,

Thank you for identifying the issue with the 740 GPU. Could you please verify the memory availability while running the face-api? Does it have sufficient memory while running the model?

[risteristevski]

Hi @shmishra99 ,

Thanks for looking into this.

I have tested the output of tf.memory() on a few devices and have observed the following:

Device	GPU	WEBGL_PACK	Total Allocated	Free Memory	Used Memory (Approx.)
Samsung Galaxy S23 Ultra	Adreno 740	true	7.33 MB	6.73 MB	621.91 KB (8.3%)
Samsung Galaxy S23 Ultra	Adreno 740	false	72.32 MB	58.96 MB	13.36 MB (18.5%)
Samsung Galaxy S22	Adreno 730	true	67.68 MB	54.03 MB	13.63 MB (20.1%)
Google Pixel 7	Mali-G710 MP7	true	67.68 MB	54.03 MB	13.63 MB (20.1%)

So the low allocated memory on the Adreno 740 seems to be a consequence of the shader link failure, rather than a cause. Disabling WEBGL_PACK works around the issue and allows the model to run as expected, but is is not a workaround, because it still may cause Chrome to crash in like, 1 out of 20 predictions.

Let me know if you need any other info to help reproduce this further.

Best regards,
Riste