Kenneth Heafield
Kenneth Heafield
There's getting CPU features which is relatively easy: ``` template T ChooseCPU(T avx512vnni, T avx512, T avx2, T ssse3, T sse2, T unsupported) { // TODO: don't catch Knights processors...
Regarding MKL, paging @sidkashyap-at-Intel
> So, the takeaway here is, we should make `--maxi-batch-sort src` default for translation. It is `trg` by-default for training. Should we go further and make batched translation the default,...
Aren't beam search based approaches deprecated in favor of model based approaches? See https://aclanthology.org/P19-1294/ Here's Marian's implementation of the above paper: https://github.com/marian-nmt/marian-examples/tree/master/forced-translation Regarding "disjunctive" constraints, it would seem the natural...
I know, will fix it this week.
Ironically KNL was the original CPU port of Marian. Then we lost interest in it but there are still 2 KNLs and 8 KNMs floating around. How is the updated...
IIRC GPUs don't have a native 64-bit int type which is why you would see a penalty.
"(and fail if out of memory)"
The dominant use case is translating things. Where you want one line in and one line out. You also want that for backtranslation. https://github.com/kpu/preprocess/blob/master/preprocess/remove_long_lines_main.cc
Marian should preserve the principle of one line in, one line out. If you're backtranslating web data, the long line should have been removed from the input; there's nothing consistent...