Expand description
Core value objects from the ubiquitous language.
Enums§
- Device
Target - Requested device class;
Autois resolved by the Hardware & Delegate context at runtime. - Model
Format - Supported on-disk model formats (ADR-002).
.pte/GGUF-via-C++ is gone; Candle reads GGUF and safetensors natively, ONNX viatract. - Phase
- Inference session state machine (ADR-001).
- Runtime
Kind - The execution engine that owns a loaded model (ADR-002).
- Safety
Mode - Tiered safety strategy (ADR-005), budget-gated by device profile.
- Speculation
Mode - Speculative decoding strategy (ADR-002 / context 3). Default is
Off. - Stop
Reason - Why generation stopped.
Type Aliases§
- Token
- A vocabulary id produced or consumed by the model — the atomic generation unit.