docs/articles_en/openvino-workflow/running-inference/optimize-inference/precision-control.rst
The choice of data types is essential to the inference runtime, which can have a huge impact on the performance and other metrics. Usually 2 types of precision are identified:
Inference precision no longer depends on the precision of IR, which means that users have several options to find the balance between model performance and accuracy.
Essentially, the IR precision becomes a way of compressing the model by reducing the precision of the weights, and it does not affect how the devices execute the model. This change clears up a lot of confusion where, for example, you couldn't execute a high-performance model on the GPU by default, and the behavior between devices was different.
This guide will focus on how to control inference precision. And using lower precision is
important for performance because compute bandwidth tends to be higher for smaller data
types, and hardware often has special blocks for efficient multiply-accumulate operations
with smaller data types only (e.g. Intel Xᵉ Matrix Extensions (XMX) on GPU and Intel
Advanced Matrix Extensions (AMX) on CPU do not support f32). Also, I/O operations
requires less memory due to the smaller tensor byte size. This guide will focus on how
to control inference precision.
.. _execution-mode:
Execution Mode ##############
ov::hint::execution_mode is a high-level hint to control whether the user wants to keep
the best accuracy (ACCURACY mode) or if the device can do some optimizations that
may lower the accuracy for performance reasons (PERFORMANCE mode)
f16 and f32, and the model is created for
f16, the network will execute in f16. Similarly, if the model is created for f32,
the network will execute in f32.
Additionally, :ref:Dynamic Quantization <enabling-runtime-optimizations> is disabled in this mode.If the model has been quantized using
:doc:OpenVINO optimization tools <../../model-optimization-guide/quantizing-models-post-training>
or any other method, the quantized operators will be executed with the target integer
precision if the device has hardware acceleration for that type. For example, quantized
int8 primitives are executed with int8 precision for both ACCURACY and
PERFORMANCE modes if the device provides higher compute bandwidth for 8-bit data types
compared to any available floating-point type. On the other hand, devices without hardware
acceleration for the int8 data type can keep such operators in floating point precision,
and the exact floating point type will be affected by execution_mode and
inference_precision properties.
Code examples:
.. tab-set::
.. tab-item:: Python :sync: py
.. doxygensnippet:: docs/articles_en/assets/snippets/ov_execution_mode.py
:language: python
:fragment: [ov:execution_mode:part0]
.. tab-item:: C++ :sync: cpp
.. doxygensnippet:: docs/articles_en/assets/snippets/ov_execution_mode.cpp
:language: cpp
:fragment: [ov:execution_mode:part0]
Inference Precision ###################
ov::hint::inference_precision precision is a lower-level property that allows you
to specify the exact precision the user wants, but is less portable. For example, CPU
supports f32 inference precision and bf16 on some platforms, GPU supports f32
and f16, so if a user wants to an application that uses multiple devices, they have
to handle such cases manually. So if possible, it is generally recommended to use high
level execution_mode property.
Another thing is that inference_precision is also a hint, so the value provided is not guaranteed
to be used by Runtime (mainly in cases where the current device does not have the required hardware
capabilities).
.. note::
All devices only support floating-point data types (f32, f16, bf16) as a value
for inference_precision attribute.
Activation Scaling ###################
Since f16 has a smaller dynamic range compared to f32 or bf16, overflow might occur when using f16 for inference_precision.
To address this issue, activation scaling divides the input of linear operations such as MatMul or Convolution by an activations scale factor, ensuring the layer's output does not exceed f16's dynamic range.
The layer's output is then multiplied back by the activations scale factor to restore its original value, but overflow can occur again during this process.
Activation scaling uses :doc:LPT (Low Precision Transformations) <../../../documentation/openvino-extensibility/openvino-plugin-library/advanced-guides/low-precision-transformations> to delay multiplication by the scale factor as much as possible, preventing overflow.
The activations scale factor can be specified in the rt_info of the model IR or via ov::hint::activations_scale_factor.
Currently, this property is supported on GPU.
.. scrollbox::
.. code-block:: cpp
<?xml version="1.0" ?>
<net name="model_file_name" version="10">
...
<rt_info>
...
<runtime_options>
<ACTIVATIONS_SCALE_FACTOR value="8.0" />
</runtime_options>
...
</rt_info>
</net>
.. _limited_inference_precision:
Limitation of the bf16 inference precision
++++++++++++++++++++++++++++++++++++++++++++++
It is important to mention that inferring FP16 and FP32 LLM models with the bf16 runtime
precision may result in higher accuracy loss than the pre-determined threshold of 0.5%.
Higher accuracy drop may occur when inferring dolly-v2-12b, dolly-v2-3b, and
gpt-neox-20b original Pytorch models with bf16, and is caused by a limited
precision representation.
To solve the issue, you might use an INT8 model and force the FP32 inference precision.
The accuracy of an INT8 model with FP32 is nearly the same as of an FP16 model with f32.
Additionally, selective FP32 execution of ops on CPU plugin together with the NNCF bf16
calibration could potentially mitigate the accuracy loss.
However, the solutions mentioned above would, unfortunately, also result in significant performance drop during a large batch size inference task on machines with Intel AMX-BF16 SPR. In such cases, the fused multiply-add operation (FMA) is used instead of AMX. Also, in a compute-bound case, such as the LLM batch inference/serving, these workarounds would drastically reduce the throughput by more than 60%.
Additional Resources ####################
Inference Devices and Modes <../inference-devices-and-modes>