Quantization
Quantization is a technique that can reduce the model size and accelerate its execution with little to no degradation in accuracy. CTranslate2 supports the most common types:
8bit integers (INT8)
16bit integers (INT16)
16bit floating points (FP16)
Tip
See the benchmark results in the main README to compare the performance and memory usage with and without quantization.
Quantize on model conversion
Enabling the quantization when converting the model is helpful to reduce its size on disk. The converters expose the option quantization
that accepts the following values:
int8
int8_float16
int16
float16
For example,
ct2opennmtpyconverter model_path model.pt quantization int8 output_dir ct2_model
Note
Whatever quantization type is selected here, the runtime ensures the model can be loaded and executed efficiently. This implies the model weights are possibly converted to another type when the model is loaded, see Implicit model conversion on load.
For reference, the table below compares the model size on disk for a base Transformer model without shared embeddings and a vocabulary of size 32k:
Quantization 
Model size 

None 
364MB 

187MB 

182MB 

100MB 

95MB 
Quantize on model loading
Quantization can also be enabled or changed when loading the model. The translator exposes the option compute_type
that accepts the following values:
auto
: selects the fastest computation type on this system and deviceint8
int8_float16
int16
float16
float
For example,
translator = ctranslate2.Translator(model_path, compute_type="int8")
Tip
Conversions between all types are supported. For example, you can convert a model with quantization="int8"
and then execute in full precision with compute_type="float"
.
Implicit model conversion on load
By default, the runtime tries to use the type that is saved in the converted model as the computation type. However, 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 in prebuilt binaries:
On CPU:
Architecture 
int8 
int8_float16 
int16 
float16 

x8664 (Intel) 
int8 
int8 
int16 
float 
x8664 (other) 
int8 
int8 
int8 
float 
AArch64/ARM64 (Apple) 
float 
float 
float 
float 
AArch64/ARM64 (other) 
int8 
int8 
int8 
float 
On GPU:
Compute Capability 
int8 
int8_float16 
int16 
float16 

>= 7.0 
int8 
int8_float16 
float16 
float16 
6.2 
float 
float 
float 
float 
6.1 
int8 
int8 
float 
float 
<= 6.0 
float 
float 
float 
float 
Tip
You can get more information about the detected capabilities of your system by setting the environment variable CT2_VERBOSE=1
. This information can also be queried at runtime with the Python function ctranslate2.get_supported_compute_types
.
Supported types
8bit integers (int8
)
Supported on:
NVIDIA GPU with Compute Capability >= 7.0 or Compute Capability 6.1
x8664 CPU with the Intel MKL or oneDNN backends
AArch64/ARM64 CPU with the Ruy backend
The implementation applies the equation from Wu et al. 2016 to quantize the weights of the embedding and linear layers:
scale[i] = 127 / max(abs(W[i,:]))
WQ[i,j] = round(scale[i] * W[i,j])
Note
This formula corresponds to a symmetric quantization (absolute maximum of the input range instead of separate min/max values).
16bit integers (int16
)
Supported on:
Intel CPU with the Intel MKL backend
The implementation follows the work by Devlin 2017. By default we use one quantization scale per layer. The scale is defined as:
scale = 2^10 / max(abs(W))
As suggested by the author, the idea is to use 10 bits for the input so that the multiplication is 20 bits which gives 12 bits left for accumulation.
Similar to the int8
quantization, only the weights of the embedding and linear layers are quantized to 16bit integers.
16bit floating points (float16
)
Supported on:
NVIDIA GPU with Compute Capability >= 7.0
In this mode, all model weights are stored in half precision and all layers are run in half precision.
Mixed 8bit integers and 16bit floating points (int8_float16
)
Supported on:
NVIDIA GPU with Compute Capability >= 7.0
This mode is the same as int8
, but all non quantized layers are run in FP16 instead of FP32.