Repository General Information

Inspired by and derived from the work of Helsinki-NLP, CTranslate2, and michaelfeil!

Link to Original Model (Helsinki-NLP): Model Link
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 SOTA and is one of the most performant ways of hosting translation models at scale. Current supported models include:

Encoder-decoder models: Transformer base/big, M2M-100, NLLB, BART, mBART, Pegasus, T5, Whisper
Decoder-only models: GPT-2, GPT-J, GPT-NeoX, OPT, BLOOM, MPT, Llama, Mistral, Gemma, CodeGen, GPTBigCode, Falcon
Encoder-only models: BERT, DistilBERT, XLM-RoBERTa

Speed up inference times by about 2x-8x using int8 inference in C++. CTranslate2 is SOTA for hosting translation models at scale.

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.

CTranslate2 Installation

pip install hf-hub-ctranslate2>=1.0.0 ctranslate2>=3.13.0

ct2-transformers-converter Command Used:

ct2-transformers-converter --model Helsinki-NLP/opus-mt-en-sn --output_dir ./ctranslate2/opus-mt-en-sn-ctranslate2 --force --copy_files README.md generation_config.json tokenizer_config.json vocab.json source.spm .gitattributes target.spm --quantization float16

CTranslate2 Converted Checkpoint Information:

Compatible With:

Compute Type:

compute_type=int8_float16 for device="cuda"
compute_type=int8 for device="cpu"

Sample Code - ctranslate2

from ctranslate2 import Translator
import transformers

model_name = "gaudi/opus-mt-en-sn-ctranslate2"
translator = Translator(
            model_path=model_name,
            device="cuda",
            inter_threads=1,
            intra_threads=4,
            compute_type="int8_float16",
)

tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)

source = tokenizer.convert_ids_to_tokens(tokenizer.encode("XXXXXX, XXX XX XXXXXX."))
results = translator.translate_batch([source])
target = results[0].hypotheses[0]

print(tokenizer.decode(tokenizer.convert_tokens_to_ids(target)))

Sample Code - hf-hub-ctranslate2

Derived From michaelfeil:

from hf_hub_ctranslate2 import TranslatorCT2fromHfHub, GeneratorCT2fromHfHub
from transformers import AutoTokenizer

model_name = "gaudi/opus-mt-en-sn-ctranslate2"
model = TranslatorCT2fromHfHub(
        model_name_or_path=model_name, 
        device="cuda",
        compute_type="int8_float16" # load in int8 on CUDA,
        tokenizer=AutoTokenizer.from_pretrained(model_name)
)
outputs = model.generate(
    text=["XXX XX XXX XXXXXXX XXXX?", "XX XX XXXX XX XXX!"],
)
print(outputs)

License and other remarks:

License conditions are intended to be idential to original huggingface repository by Helsinki-NLP.