[BUG] NNDescent：unused local_join function and reports an unaligned error

[zben777]

Describe the bug
When I use wmma with a uint8_t parameter, I find that the program does not make the call to the local_join function and reports an unaligned error：code=716(cudaErrorMisalignedAddress) "cudaDeviceSynchronize()"
I would like to ask if there is support for uint8_t or int8_t matrix calculations in local_join?

I get an error when using the following uint8_t. And equipment ：
NVIDIA GeForce RTX 4090 Driver Version: 560.28.03 CUDA Version: 12.6

And I'd like to ask if the current NNDescent supports local_join calculations other than __half

[cjnolet]

Hi @zben777 thanks for opening an issue. Can you please provide a minimal reproducible code snippet / example that we can use to reproduce this behavior? nn-descent is only instantiated for specific supported types and it's not header-only. Are you calling public APIs or internal APIs?