PretrainedDType

Enums · AiDotNet.ModelLoading.Pretrained

The numeric precision a pretrained checkpoint’s weights are decoded into when loaded.

For Beginners

Model files store their weights in a specific precision (32-bit float, 16-bit half, or 16-bit “brain float”). This picks what precision AiDotNet decodes them to. The default, Auto, keeps whatever the file used — which is almost always what you want.

Fields

Field Summary
PretrainedDType Auto Keep the checkpoint’s on-disk precision (recommended default).
PretrainedDType BFloat16 Decode weights to bfloat16 (BF16) — the dtype most modern LLMs ship in.
PretrainedDType Float16 Decode weights to IEEE 754 half precision (F16).
PretrainedDType Float32 Decode weights to 32-bit single precision.
int value__