diff --git a/pipeline/translate/decoder.yml b/pipeline/translate/decoder.yml index 4b2b2d3df..678a8ef00 100644 --- a/pipeline/translate/decoder.yml +++ b/pipeline/translate/decoder.yml @@ -1,10 +1,16 @@ normalize: 1.0 word-penalty: 0 -mini-batch: 16 -mini-batch-words: 500 -maxi-batch: 1000 + +# TODO #936 - These values can be unsafe for certain machines. + +# Batch size tuned for four Tesla V100-SXM2-16GB +# See: https://github.com/mozilla/translations/issues/931 +mini-batch-words: 5000 +maxi-batch: 10000 maxi-batch-sort: src +fp16: true + max-length: 200 max-length-crop: true beam-size: 8 -quiet-translation: True \ No newline at end of file +quiet-translation: True diff --git a/taskcluster/configs/config.prod.yml b/taskcluster/configs/config.prod.yml index 9b8d8bb04..09eae49ea 100644 --- a/taskcluster/configs/config.prod.yml +++ b/taskcluster/configs/config.prod.yml @@ -216,9 +216,12 @@ marian-args: beam-size: '12' mini-batch-words: '2000' decoding-teacher: - mini-batch-words: '4000' precision: float16 - + # Batch size tuned for four Tesla V100-SXM2-16GB + # See: https://github.com/mozilla/translations/issues/931 + mini-batch-words: '5000' + maxi-batch: '10000' + fp16: true # Early stopping can be adjusted to ensure models converge. See: # https://mozilla.github.io/firefox-translations-training/training-guide.html#model-training training-backward: