Lilt can translate multiple files in parallel. Grouping documents into batches for pretranslation significantly increases throughput compared to submitting files for pretranslation individually. You can submit batches of files through the UI or through the API.
Behind the scenes, Lilt uses GPUs to translate documents for maximum speed. On-premise customers benefit from the usage of all recognized GPUs in the system. A single batch request may be processed by 1 or more GPUs, depending on the number of documents and segments in it.
Performance Characteristics and Caveats
This article offers advice to optimize for the fastest time to completion for the entire set of files. This may result in a decreased per document translation speed, as there is some overhead to grouping jobs. Thus, if you desire the fastest possible translation time for a single document, submit it for translation individually.
The ideal batch size is between 10 and 100 files per batch. The greatest speed-ups will be seen in batches of hundreds of relatively small files, although mixed filesize workloads will also experience significant speed-ups.
Using the UI
Navigate to the project screen and click the “Select All” button. Click “Pretranslate” and select the last option, “Use MT for unmatched segments”.
Then, click “Pretranslate.” A confirmation appears indicating the batch job has been submitted for translation, and a spinner will appear on the right-hand side of each document indicating the progress of that file's batch translation.
Using the API
You can start machine translation of one or more files that have been previously uploaded by using the
/translate/file endpoint and specifying multiple file IDs as documented here. This option bypasses writing data to our database and is useful for pure machine translation workflows. Alternatively, you can use our
pretranslate endpoint here, which is similar to clicking the Pretranslate button on our UI.