This respository was based on the work of CTranslate2.
This repository was based on the work of michaelfeil.
What is CTranslate2?
CTranslate2 is a C++ and Python library for efficient inference with Transformer models.
CTranslate2 implements a custom runtime that applies many performance optimization techniques such as weights quantization, layers fusion, batch reordering, etc., to accelerate and reduce the memory usage of Transformer models on CPU and GPU.
CTranslate2 is one of the most performant ways of hosting translation models at scale. Current supported models include:
The project is production-oriented and comes with backward compatibility guarantees, but it also includes experimental features related to model compression and inference acceleration.
CTranslate2 Benchmarks
Please note that the results presented below are only valid for the configuration used during this benchmark: absolute and relative performance may change with different settings. Tested against newstest2014 (En -> De) dataset.
The benchmark reports the number of target tokens generated per second (higher is better). The results are aggregated over multiple runs. See the benchmark scripts for more details and reproduce these numbers.
Please note that the results presented below are only valid for the configuration used during this benchmark: absolute and relative performance may change with different settings.
CPU Benchmarks for Generic Opus-MT Models
Library
Tokens per Second
Max Memory Usage
BLEU
Transformers 4.26.1 (with PyTorch 1.13.1)
147.3
2332MB
27.90
Marian 1.11.0 (int16)
330.2
5901MB
27.65
Marian 1.11.0 (int8)
355.8
4763MB
27.27
CTranslate2 3.6.0 (int16)
596.1
660MB
27.53
CTranslate2 3.6.0 (int8)
696.1
516MB
27.65
GPU Benchmarks for Generic Opus-MT Models
Library
Tokens per Second
Max GPU Memory Usage
Max Memory Usage
BLEU
Transformers 4.26.1 (with PyTorch 1.13.1)
1022.9
4097MB
2109MB
27.90
Marian 1.11.0 (float16)
3962.4
3239MB
1976MB
27.94
CTranslate2 3.6.0 (float16)
9296.7
909MB
814MB
27.9
CTranslate2 3.6.0 (int8 + float16)
8362.7
813MB
766MB
27.9
Executed with 4 threads on a c5.2xlarge Amazon EC2 instance equipped with an Intel(R) Xeon(R) Platinum 8275CL CPU.
Source to benchmark information can be found here. Original model BLEU scores can be found here.
Internal Benchmarks
Internal testing on our end showed inference times reduced by 6x-10x on average compared the vanilla checkpoints using the transformers library. A slight reduction on BLEU scores (~5%) was also identified in comparison to the vanilla checkpoints with a few exceptions. This is likely due to several factors, one being the quantization applied. Further testing is needed from our end to better assess the reduction in translation quality. The command used to compile the vanilla checkpoint into a CTranslate2 model can be found below. Modifying this command can yield differing balances between inferencing performance and translation quality.
Take the python code below and update the 'model_dir' variable to the location of the cloned repository.
python
1from ctranslate2 import Translator
2import transformers
34model_dir ="./opus-mt-fi-tvl-ctranslate2"# Path to model directory.5translator = Translator(6 model_path=model_dir,7 device="cuda",# cpu, cuda, or auto.8 inter_threads=1,# Maximum number of parallel translations.9 intra_threads=4,# Number of OpenMP threads per translator.10 compute_type="int8_float16",# int8 for cpu or int8_float16 for cuda.11)1213tokenizer = transformers.AutoTokenizer.from_pretrained(model_dir)1415source = tokenizer.convert_ids_to_tokens(tokenizer.encode("XXXXXX, XXX XX XXXXXX."))16results = translator.translate_batch([source])17target = results[0].hypotheses[0]1819print(tokenizer.decode(tokenizer.convert_tokens_to_ids(target)))
