On a self-hosted instance, you can configure settings to determine whether to the auxiliary model, and model use on GPU.
To improve overall inference, you can configure whether Textual uses the en_core_web_sm auxiliary NER model.
The auxiliary model detects the following types:
EVENT
LANGUAGE
LAW
NRP
NUMERIC_VALUE
PRODUCT
WORK_OF_ART
To configure whether to use the auxiliary model, you use the environment variable TEXTUAL_AUX_MODEL.
The available values are:
en_core_web_sm - This is the default value.
none - Indicates to not use the auxiliary model.
When you use a textual-ml-gpu container on accelerated hardware, you can configure:
Whether to use the auxiliary model,
Whether to use the date synthesis model
To configure whether to use the auxiliary model for GPU, you configure the environment variable TEXTUAL_AUX_MODEL_GPU.
By default, on GPU, Textual does not use the auxiliary model, and TEXTUAL_AUX_MODEL_GPU is false.
To use the auxiliary model for GPU, based on the configuration of TEXTUAL_AUX_MODEL, set TEXTUAL_AUX_MODEL_GPU to true.
When TEXTUAL_AUX_MODEL_GPU is true, and TEXTUAL_MULTI_LINGUAL is true, Textual also loads the multilingual models on GPU.
By default, on GPU, Textual loads the date synthesis model on GPU.
Note that this model requires 600MB of GPU RAM for each machine learning worker.
To not load the date synthesis model on GPU, set TEXTUAL_DATE_SYNTH_GPU to false.