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:
- 8-bit integers (INT8)
- 16-bit integers (INT16)
- 16-bit floating points (FP16)
See the benchmark results in the main README for a performance and memory usage comparison.
Quantization can be enabled when converting the model or when loading the model.
Enabling the quantization during conversion is helpful to reduce the model size on disk. The converters expose the option quantization
that accepts the following values:
int8
int8_float16
int16
float16
For example,
ct2-opennmt-py-converter --model_path model.pt --quantization int8 --output_dir ct2_model
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 next section).
Quantization can also be enabled or changed when loading the model. The translator exposes the option compute_type
that accepts the following values:
default
: see description belowauto
: selects the fastest computation typeint8
int8_float16
int16
float16
float
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"
.
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:
CPU vendor | int8 | int8_float16 | int16 | float16 |
---|---|---|---|---|
Intel | int8 | int8 | int16 | float |
other | int8 | int8 | int8 | float |
On GPU:
GPU Compute Capability | int8 | int8_float16 | int16 | float16 |
---|---|---|---|---|
>= 7.0 | int8 | int8_float16 | float16 | float16 |
6.1 | int8 | int8 | float | float |
<= 6.0 | float | float | float | float |
You can get more information about the detected capabilities of your system by setting the environment variable CT2_VERBOSE=1
.
Supported on:
- NVIDIA GPU with Compute Capability >= 6.1
- x86-64 CPU with the Intel MKL or oneDNN backends
The implementation applies the equation from Wu et al. 2016 to compute the quantized weights:
Note that this corresponds to a symmetric quantization (absolute maximum of the input range instead of separate min/max values). We only quantize the weights of the embedding and linear layers.
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. We only quantize the weights of the embedding and linear layers.
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.
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.