Kenneth Heafield

Sure, just tell us your list of ~sales prospects~ clients.

Can we add the word itself as a second sort criterion (or even just make the sort stable) to make it deterministic?

Seen on the brand-new 3090s. ``` [2021-04-27 18:45:38] Error: Cublas Error: 13 - /home/heafield/marian-dev/src/tensors/gpu/prod.cpp:118: cublasGemmEx(handle, transa, transb, m, n, k, alpha, A, CUDA_R_32F, lda, B, CUDA_R_32F, ldb, beta, C, CUDA_R_32F,...

That makes a FP32 model that's ready to be 8-bit quantized. Next step is to binarize it. https://github.com/browsermt/students/tree/master/train-student Note, due to stubbornness in marian-nmt/marian-dev#762 you won't get the best 8-bit...

There is documentation at https://github.com/browsermt/students/tree/master/train-student ; if it's unclear feel free to file an issue against that repo.

The short answer is `quantize-bits` doesn't work when you train a model from scratch. I think it's an interesting research question to see if one could fully train a model...

> I agree, but at the same time I think any statistical testing is probably better than none. I think a statistical test that always claims significance (and bootstrap does...

I've only bothered to measure AVX512 but we should check. Paging @sidkashyap.

MKL is blocking Wikipedia from deploying Marian because it is closed source.

The trick is all that registration just goes to a constant URL with no authentication: http://registrationcenter-download.intel.com/akdlm/irc_nas/tec/16849/l_mkl_2020.2.254.tgz And the "customizable" package is http://registrationcenter-download.intel.com/akdlm/irc_nas/tec/16849/l_mkl_2020.2.254_online.tgz