Alexander Grund

Sorry but your gist doesn't show the test passing. Last lines I see are ``` [5,459 / 9,765] 2 actions running Compiling .../mlir/expansions/segmentation_spmd_expander.cc; 13s local Compiling llvm/lib/Target/X86/X86FastTileConfig.cpp; 1s local ```...

The issue still persists: Only a CPU implementation is registered: https://github.com/tensorflow/tensorflow/blob/v2.17.0/tensorflow/core/kernels/scatter_nd_op.cc#L655-L657

Sure, although the full command line is massive and I expect that to be a general error in the code or test, as the way TF was build shouldn't affect...

I think I'm doing that already beeing even more conservative. The relevant parts from my above command also reformatted: ``` bazel \ --compilation_mode=opt --config=opt --config=noaws --jobs=64 \ --distinct_host_configuration=false \ --test_env=TF_GPU_COUNT=6...

> Other tests are not guaranteed to pass on GPUs. This makes me very surprised. My understanding of unit tests is, that they should test code for correctness. The availability...

> So I recommend you do two runs, one without GPUs and one with GPUs Although inconvenient (and I wonder what the `nogpu` and `no_gpu` tags are for then) I...

Ok this seems to be another issue which can be workaround by excluding `no_cuda_on_cpu_tap`, see https://github.com/tensorflow/tensorflow/issues/46425 for an argument against that tag. In a similar case the presence of the...

> So I would recommend reconfiguring TF before running the non-GPU test suite. This is highly inconvenient. Doing it this way I can't be sure that the TF I built...

@mihaimaruseac Has there been any progress in the refactoring you mentioned?

The issue persists and can be triggered using the same reproducer