NotFoundError could not find registered transfer manager for platform Host -- check target linkage [Op:__inference__jit_compiled_convolution_op_26169] about tf-keras HOT 3 CLOSED

@innat TPU nodes (Colab) and TPU VMs are quite different in my experience.

The TPU node is basically a Tensorflow server you send tf code to execute on, vs. the TPU VM has the TPU directly connected to the VM so instructions run "locally".

Because of this TPU nodes can be talked to with "stock" tensorflow, but TPU VMs use a specialized tensorflow library to talk directly to the TPU.

Its probably best to report these issues to the Tensorflow team directly https://github.com/tensorflow/tensorflow I don't see if they have a TPU VM specific contact anywhere but I'd call out that it happens on TPU VM.

from tf-keras.

A bit confusing, if I run the above code block on colab-tpu pod, it works but it doesn't on the kaggle tpu-vm.

gist: https://www.kaggle.com/code/ipythonx/github-issue-655-keras-tpu-vm/notebook

Hello @djherbis, could you please add something here? I think it's somewhat related to hardware issue, not sure.

from tf-keras.

Thanks @djherbis for the explanation. As you suggested, I create a separate ticket in tensorflow. tensorflow/tensorflow#61976

from tf-keras.