added ctranslate2 docs

docs/configs/configuration_files.md

@@ -147,6 +147,31 @@ In this case, the lang_tags mapping will be used in the prompt.
 
 Note: When using a Hugging Face model as a teacher, there is no scoring or cross-entropy filtering.
 
+#### CTranslate2
+
+The pipeline also supports CTranslate2 inference for HuggingFace models, which provides a considerable speedup.
+For that, simply add new boolean key:
+
+```yaml
+huggingface:
+    modelname: "facebook/nllb-200-distilled-1.3B"
+    lang_info: True
+    batch_size: 4096
+    lang_tags:
+      en: eng_Latn
+      ja: jpn_Jpan
+    ct2: True
+```
+
+We have done some benchmarking on 4 Nvidia Ampere A100 GPUs that shows CTranslate2 provides a 26x faster inference:
+
+```markdown
+| Model                               | Type         | Batch size | Return sequences | Sent/s  |
+|-------------------------------------|--------------|------------|------------------|---------|
+| facebook/nllb-200-distilled-1.3B   | ctranslate2   | 8192       | 8                | 406,316 |
+| facebook/nllb-200-distilled-1.3B   | huggingface  | 8          | 8                | 15,37   |
+```
+
 ## Backward models
 
 Currently, only OPUS-MT models are available as backward models for scoring translations.