CTranslate2
CTranslate2 is a fast inference engine for OpenNMT-py and OpenNMT-tf models supporting both CPU and GPU execution. The goal is to provide comprehensive inference features and be the most efficient and cost-effective solution to deploy standard neural machine translation systems such as Transformer models.
The project is production-oriented and comes with backward compatibility guarantees, but it also includes experimental features related to model compression and inference acceleration.
Table of contents
- Key features
- Quickstart
- Installation
- Converting models
- Translating
- Environment variables
- Building
- Testing
- Benchmarks
- Frequently asked questions
Key features
- Fast and efficient execution on CPU and GPU
The execution is significantly faster and requires less resources than general-purpose deep learning frameworks on supported models and tasks. - Quantization and reduced precision
The model serialization and computation support weights with reduced precision: 16-bit floating points (FP16), 16-bit integers, and 8-bit integers. - Multiple CPU architectures support
The project supports x86-64 and ARM64 processors and integrates multiple backends that are optimized for these platforms: Intel MKL, oneDNN, OpenBLAS, and Apple Accelerate. - Automatic CPU detection and code dispatch
One binary can include multiple backends (e.g. Intel MKL and oneDNN) and instruction set architectures (e.g. AVX, AVX2) that are automatically selected at runtime based on the CPU information. - Parallel and asynchronous translations
Translations can be run efficiently in parallel and asynchronously using multiple GPUs or CPU cores. - Dynamic memory usage
The memory usage changes dynamically depending on the request size while still meeting performance requirements thanks to caching allocators on both CPU and GPU. - Lightweight on disk
Models can be quantized below 100MB with minimal accuracy loss. A full featured Docker image supporting GPU and CPU requires less than 400MB. - Simple integration
The project has few dependencies and exposes translation APIs in Python and C++ to cover most integration needs. - Interactive decoding
Advanced decoding features allow autocompleting a partial translation and returning alternatives at a specific location in the translation.
Some of these features are difficult to achieve with standard deep learning frameworks and are the motivation for this project.
Supported decoding options
The translation API supports several decoding options:
- decoding with greedy or beam search
- random sampling from the output distribution
- translating with a known target prefix
- returning alternatives at a specific location in the target
- constraining the decoding length
- returning multiple translation hypotheses
- returning attention vectors
- approximating the generation using a pre-compiled vocabulary map
- replacing unknown target tokens by source tokens with the highest attention
See the Decoding documentation for examples.
Quickstart
The steps below assume a Linux OS and a Python installation (3.6 or above).
1. Install the Python package:
pip install --upgrade pip
pip install ctranslate2
2. Convert a model trained with OpenNMT-py or OpenNMT-tf, for example the pretrained Transformer model (choose one of the two models):
a. OpenNMT-py
pip install OpenNMT-py
wget https://s3.amazonaws.com/opennmt-models/transformer-ende-wmt-pyOnmt.tar.gz
tar xf transformer-ende-wmt-pyOnmt.tar.gz
ct2-opennmt-py-converter --model_path averaged-10-epoch.pt --model_spec TransformerBase \
--output_dir ende_ctranslate2
b. OpenNMT-tf
pip install OpenNMT-tf
wget https://s3.amazonaws.com/opennmt-models/averaged-ende-export500k-v2.tar.gz
tar xf averaged-ende-export500k-v2.tar.gz
ct2-opennmt-tf-converter --model_path averaged-ende-export500k-v2 --model_spec TransformerBase \
--output_dir ende_ctranslate2
3. Translate tokenized inputs, for example with the Python API:
import ctranslate2
translator = ctranslate2.Translator("ende_ctranslate2/")
translator.translate_batch([["▁H", "ello", "▁world", "!"]])
Installation
Python package
Python packages are published on PyPI for Linux and macOS:
pip install ctranslate2
All software dependencies are included in the package, including CUDA libraries for GPU support on Linux. The macOS version only supports CPU execution.
Requirements:
- OS: Linux, macOS
- Python version: >= 3.6
- pip version: >= 19.3
- GPU driver version: >= 418.39
Docker images
The opennmt/ctranslate2 repository contains images for multiple Linux distributions, with or without GPU support:
docker pull opennmt/ctranslate2:latest-ubuntu18-cuda11.0
The images include:
- a translation client to directly translate files
- Python 3 packages
libctranslate2.solibrary development files
Manual compilation
See Building.
Converting models
The core CTranslate2 implementation is framework agnostic. The framework specific logic is moved to a conversion step that serializes trained models into a simple binary format.
The following frameworks and models are currently supported:
| OpenNMT-tf | OpenNMT-py | |
|---|---|---|
| Transformer (Vaswani et al. 2017) | ✓ | ✓ |
| + relative position representations (Shaw et al. 2018) | ✓ | ✓ |
If you are using a model that is not listed above, consider opening an issue to discuss future integration.
Conversion scripts are parts of the Python package and should be run in the same environment as the selected training framework:
ct2-opennmt-py-converterct2-opennmt-tf-converter
The converter Python API can also be used to convert Transformer models with any number of layers, hidden dimensions, and attention heads.
Integrated model conversion
Models can also be converted directly from the supported training frameworks. See their documentation:
Quantization and reduced precision
The converters support reducing the weights precision to save on space and possibly accelerate the model execution. The --quantization option accepts the following values:
int8int16float16
When loading a quantized model, the library tries to use the same type for computation. If the current platform or backend do not support optimized execution for this computation type (e.g. int16 is not optimized on GPU), then the library converts the model weights to another optimized type. The tables below document the fallback types:
On CPU:
| Model | int8 | int16 | float16 |
|---|---|---|---|
| Intel | int8 | int16 | float |
| other | int8 | int8 | float |
(This table only applies for prebuilt binaries or when compiling with both Intel MKL and oneDNN backends.)
On GPU:
| Compute Capability | int8 | int16 | float16 |
|---|---|---|---|
| >= 7.0 | int8 | float16 | float16 |
| 6.1 | int8 | float | float |
| <= 6.0 | float | float | float |
Notes:
- The computation type can also be changed when creating a translation instance by setting the
--compute_typeargument. - Integer quantization is only applied for GEMM-based layers and embeddings.
Adding converters
Each converter should populate a model specification with trained weights coming from an existing model. The model specification declares the variable names and layout expected by the CTranslate2 core engine.
See the existing converters implementation which could be used as a template.
Translating
The examples use the English-German model converted in the Quickstart. This model requires a SentencePiece tokenization.
With the translation client
echo "▁H ello ▁world !" | docker run --gpus=all -i --rm -v $PWD:/data \
opennmt/ctranslate2:latest-ubuntu18-cuda11.0 --model /data/ende_ctranslate2 --device cuda
See docker run --rm opennmt/ctranslate2:latest-ubuntu18-cuda11.0 --help for additional options.
With the Python API
import ctranslate2
translator = ctranslate2.Translator("ende_ctranslate2/")
translator.translate_batch([["▁H", "ello", "▁world", "!"]])
See the Python reference for more advanced usages.
With the C++ API
#include <iostream>
#include <ctranslate2/translator.h>
int main() {
ctranslate2::Translator translator("ende_ctranslate2/");
ctranslate2::TranslationResult result = translator.translate({"▁H", "ello", "▁world", "!"});
for (const auto& token : result.output())
std::cout << token << ' ';
std::cout << std::endl;
return 0;
}
See the Translator class for more advanced usages, and the TranslatorPool class for running translations in parallel.
Environment variables
Some environment variables can be configured to customize the execution:
CT2_CUDA_ALLOCATOR: Select the CUDA memory allocator. Possible values are:cub_caching(default),cuda_malloc_async(requires CUDA >= 11.2).CT2_CUDA_ALLOW_FP16: Allow using FP16 computation on GPU even if the device does not have efficient FP16 support.CT2_CUDA_CACHING_ALLOCATOR_CONFIG: Tune the CUDA caching allocator (see Performance).CT2_FORCE_CPU_ISA: Force CTranslate2 to select a specific instruction set architecture (ISA). Possible values are:GENERIC,AVX,AVX2. Note: this does not impact backend libraries (such as Intel MKL) which usually have their own environment variables to configure ISA dispatching.CT2_TRANSLATORS_CORE_OFFSET: If set to a non negative value, parallel translators are pinned to cores in the range[offset, offset + inter_threads]. Requires compilation with-DOPENMP_RUNTIME=NONE.CT2_USE_EXPERIMENTAL_PACKED_GEMM: Enable the packed GEMM API for Intel MKL (see Performance).CT2_USE_MKL: Force CTranslate2 to use (or not) Intel MKL. By default, the runtime automatically decides whether to use Intel MKL or not based on the CPU vendor.CT2_VERBOSE: Enable some verbose logs to help debugging the run configuration.
Building
Docker images
The Docker images build all translation clients presented in Translating. The build command should be run from the project root directory, e.g.:
docker build -t opennmt/ctranslate2:latest-ubuntu18 -f docker/Dockerfile.ubuntu .
When building GPU images, the CUDA version can be selected with --build-arg CUDA_VERSION=11.0.
See the docker/ directory for available images.
Build options
The project uses CMake for compilation. The following options can be set with -DOPTION=VALUE:
| CMake option | Accepted values (default in bold) | Description |
|---|---|---|
| BUILD_TESTS | OFF, ON | Compiles the tests |
| CMAKE_CXX_FLAGS | compiler flags | Defines additional compiler flags |
| ENABLE_CPU_DISPATCH | OFF, ON | Compiles CPU kernels for multiple ISA and dispatches at runtime (should be disabled when explicitly targeting an architecture with the -march compilation flag) |
| ENABLE_PROFILING | OFF, ON | Enables the integrated profiler (usually disabled in production builds) |
| LIB_ONLY | OFF, ON | Disables the translation client |
| OPENMP_RUNTIME | INTEL, COMP, NONE | Selects or disables the OpenMP runtime (INTEL: Intel OpenMP; COMP: OpenMP runtime provided by the compiler; NONE: no OpenMP runtime) |
| WITH_CUDA | OFF, ON | Compiles with the CUDA backend |
| WITH_DNNL | OFF, ON | Compiles with the oneDNN backend (a.k.a. DNNL) |
| WITH_MKL | OFF, ON | Compiles with the Intel MKL backend |
| WITH_ACCELERATE | OFF, ON | Compiles with the Apple Accelerate backend |
| WITH_OPENBLAS | OFF, ON | Compiles with the OpenBLAS backend |
Some build options require external dependencies:
-DWITH_MKL=ONrequires:- Intel MKL (>=2019.5)
-DWITH_DNNL=ONrequires:- oneDNN (>=1.5)
-DWITH_ACCELERATE=ONrequires:- Accelerate (only available on macOS)
-DWITH_OPENBLAS=ONrequires:-DWITH_CUDA=ONrequires:- cuBLAS (>=10.0)
Multiple backends can be enabled for a single build. When building with both Intel MKL and oneDNN, the backend will be selected at runtime based on the CPU information.
Example (Ubuntu)
Install Intel MKL (optional for GPU only builds)
Use the following instructions to install Intel MKL:
wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
sudo sh -c 'echo "deb https://apt.repos.intel.com/oneapi all main" > /etc/apt/sources.list.d/oneAPI.list'
sudo apt-get update
sudo apt-get install intel-oneapi-mkl-devel
See the Intel MKL documentation for other installation methods.
Install CUDA (optional for CPU only builds)
See the NVIDIA documentation for information on how to download and install CUDA.
Compile
Under the project root, run the following commands:
git submodule update --init
mkdir build && cd build
cmake -DWITH_MKL=ON -DWITH_CUDA=ON ..
make -j4
(If you did not install one of Intel MKL or CUDA, set its corresponding flag to OFF in the CMake command line.)
These steps should produce the cli/translate binary. You can try it with the model converted in the Quickstart section:
$ echo "▁H ello ▁world !" | ./cli/translate --model ende_ctranslate2/ --device auto
▁Hallo ▁Welt !
Testing
C++
To enable the tests, you should configure the project with cmake -DBUILD_TESTS=ON. The binary tests/ctranslate2_test runs all tests using Google Test. It expects the path to the test data as argument:
./tests/ctranslate2_test ../tests/data
Python
# Install the CTranslate2 library.
cd build && make install && cd ..
# Build and install the Python wheel.
cd python
pip install -r install_requirements.txt
python setup.py bdist_wheel
pip install dist/*.whl
# Run the tests with pytest.
pip install -r tests/requirements.txt
pytest tests/test.py
Depending on your build configuration, you might need to set LD_LIBRARY_PATH if missing libraries are reported when running tests/test.py.
Benchmarks
We compare CTranslate2 with OpenNMT-py and OpenNMT-tf on their pretrained English-German Transformer models (available on the website). For this benchmark, CTranslate2 models are using the weights of the OpenNMT-py model.
Model size
| Model size | |
|---|---|
| OpenNMT-py | 542MB |
| OpenNMT-tf | 367MB |
| CTranslate2 | 364MB |
| - int16 | 187MB |
| - float16 | 182MB |
| - int8 | 100MB |
CTranslate2 models are generally lighter and can go as low as 100MB when quantized to int8. This also results in a fast loading time and noticeable lower memory usage during runtime.
Results
We translate the test set newstest2014 and report:
- the number of target tokens generated per second (higher is better)
- the maximum memory usage (lower is better)
- the BLEU score of the detokenized output (higher is better)
Translations are running beam search with a size of 4 and a maximum batch size of 32.
See the directory tools/benchmark for more details about the benchmark procedure and how to run it. Also see the Performance document to further improve CTranslate2 performance.
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
| Tokens per second | Max. memory | BLEU | |
|---|---|---|---|
| OpenNMT-tf 2.14.0 (with TensorFlow 2.4.0) | 279.3 | 2308MB | 26.93 |
| OpenNMT-py 2.0.0 (with PyTorch 1.7.0) | 292.9 | 1840MB | 26.77 |
| - int8 | 383.3 | 1784MB | 26.86 |
| CTranslate2 1.17.0 | 593.2 | 970MB | 26.77 |
| - int16 | 777.2 | 718MB | 26.84 |
| - int8 | 921.5 | 635MB | 26.92 |
| - int8 + vmap | 1143.4 | 621MB | 26.75 |
Executed with 4 threads on a c5.metal Amazon EC2 instance equipped with an Intel(R) Xeon(R) Platinum 8275CL CPU.
GPU
| Tokens per second | Max. GPU memory | Max. CPU memory | BLEU | |
|---|---|---|---|---|
| OpenNMT-tf 2.14.0 (with TensorFlow 2.4.0) | 1753.4 | 4958MB | 2525MB | 26.93 |
| OpenNMT-py 2.0.0 (with PyTorch 1.7.0) | 1189.4 | 2838MB | 2666MB | 26.77 |
| CTranslate2 1.17.0 | 2721.1 | 1164MB | 954MB | 26.77 |
| - int8 | 3710.0 | 882MB | 541MB | 26.86 |
| - float16 | 3965.8 | 924MB | 590MB | 26.75 |
| - float16 + local sorting | 4869.4 | 1148MB | 591MB | 26.75 |
Executed with CUDA 11.0 on a g4dn.xlarge Amazon EC2 instance equipped with a NVIDIA T4 GPU (driver version: 450.80.02).
Frequently asked questions
- How does it relate to the original CTranslate project?
- What is the state of this project?
- Why and when should I use this implementation instead of PyTorch or TensorFlow?
- What hardware is supported?
- What are the known limitations?
- What are the future plans?
- What is the difference between
intra_threadsandinter_threads? - Do you provide a translation server?
- How do I generate a vocabulary mapping file?
How does it relate to the original CTranslate project?
The original CTranslate project shares a similar goal which is to provide a custom execution engine for OpenNMT models that is lightweight and fast. However, it has some limitations that were hard to overcome:
- a strong dependency on LuaTorch and OpenNMT-lua, which are now both deprecated in favor of other toolkits;
- a direct reliance on Eigen, which introduces heavy templating and a limited GPU support.
CTranslate2 addresses these issues in several ways:
- the core implementation is framework agnostic, moving the framework specific logic to a model conversion step;
- the internal operators follow the ONNX specifications as much as possible for better future-proofing;
- the call to external libraries (Intel MKL, cuBLAS, etc.) occurs as late as possible in the execution to not rely on a library specific logic.
What is the state of this project?
The implementation has been generously tested in production environment so people can rely on it in their application. The project versioning follows Semantic Versioning 2.0.0. The following APIs are covered by backward compatibility guarantees:
- Converted models
- Python converters options
- Python symbols:
ctranslate2.Translatorctranslate2.converters.OpenNMTPyConverterctranslate2.converters.OpenNMTTFConverter
- C++ symbols:
ctranslate2::models::Modelctranslate2::TranslationOptionsctranslate2::TranslationResultctranslate2::Translatorctranslate2::TranslatorPool
- C++ translation client options
Other APIs are expected to evolve to increase efficiency, genericity, and model support.
Why and when should I use this implementation instead of PyTorch or TensorFlow?
Here are some scenarios where this project could be used:
- You want to accelarate standard translation models for production usage, especially on CPUs.
- You need to embed translation models in an existing C++ application without adding large dependencies.
- Your application requires custom threading and memory usage control.
- You want to reduce the model size on disk and/or memory.
However, you should probably not use this project when:
- You want to train custom architectures not covered by this project.
- You see no value in the key features listed at the top of this document.
What hardware is supported?
CPU
CTranslate2 supports x86-64 and ARM64 processors. It includes optimizations for AVX, AVX2, and NEON and supports multiple BLAS backends that should be selected based on the target platform (see Building).
Prebuilt binaries are designed to run on any x86-64 processors supporting at least SSE 4.2. The binaries implement runtime dispatch to select the best backend and instruction set architecture (ISA) for the platform. In particular, they are compiled with both Intel MKL and oneDNN so that Intel MKL is only used on Intel processors where it performs best, whereas oneDNN is used on other x86-64 processors such as AMD.
GPU
CTranslate2 supports NVIDIA GPUs with a Compute Capability greater or equal to 3.0 (Kepler). FP16 execution requires a Compute Capability greater or equal to 7.0.
The driver requirement depends on the CUDA version. See the CUDA Compatibility guide for more information.
What are the known limitations?
The current approach only exports the weights from existing models and redefines the computation graph via the code. This implies a strong assumption of the graph architecture executed by the original framework.
We are actively looking to ease this assumption by supporting ONNX as model parts.
What are the future plans?
There are many ways to make this project better and even faster. See the open issues for an overview of current and planned features. Here are some things we would like to get to:
- Support of running ONNX graphs
What is the difference between intra_threads and inter_threads?
intra_threadsis the number of OpenMP threads that is used per translation: increase this value to decrease the latency.inter_threadsis the maximum number of CPU translations executed in parallel: increase this value to increase the throughput. Even though the model data are shared, this execution mode will increase the memory usage as some internal buffers are duplicated for thread safety.
The total number of computing threads launched by the process is summarized by this formula:
num_threads = inter_threads * intra_threads
Note that these options are only defined for CPU translation and are forced to 1 when executing on GPU. Parallel translations on GPU require multiple GPUs. See the option device_index that accepts multiple device IDs.
Do you provide a translation server?
The OpenNMT-py REST server is able to serve CTranslate2 models. See the code integration to learn more.
How do I generate a vocabulary mapping file?
See here.
62 Dec 14, 2022
58 Jan 3, 2023
11 Dec 15, 2022
6.9k Dec 31, 2022
980 Dec 29, 2022
2 Nov 29, 2021
931 Dec 29, 2022
2 Jan 10, 2022
442 Dec 25, 2022
939 Dec 29, 2022
4.4k Dec 30, 2022
10 Jun 30, 2022
2.4k Dec 24, 2022
20.8k Jan 9, 2023
20 Dec 6, 2022
23 Oct 29, 2022
13 Nov 4, 2022
15 Nov 29, 2022