Select the storage format for model coefficients and weights: 8-bit, 16-bit, or 32-bit. Lower bit depths (8 or 16) help minimize model size and optimize device resource usage. For 8-bit and 16-bit storage, calculations can be performed in both floats and integers. For 32-bit storage, only floats are used. By default, this setting matches your dataset type, but you can adjust it as needed.
Note
When the LiteRT framework is used, the value is predefined to 32-bit floating point and cannot be changed. Training uses 32-bit floating point, but the final model is quantized.