LiteRT Next GPU delegate fails for Whisper encoder

[Jonny-GM]

Following these instructions, I converted the encoder of https://huggingface.co/openai/whisper-large-v3-turbo to the LiteRT format and validated it (it takes an input of [1,128,3000]). I can upload the resulting .tflite file anywhere convenient if necessary.

When running with Kotlin on my Samsung S23U phone using these instructions, it works fine on the CPU delegate, but the GPU delegate throws an error and crashes (see log attached).

I'd be happy to try to do surgery on the model if I could understand the issue better.

log.txt

[Jonny-GM]

Any updates?