Expand description
Configuration types for Ferrum components
Structs§
- Backend
Config - Backend configuration
- Batch
Config - Engine
Config - Engine configuration
- Engine
Model Config - KvCache
Config - KV Cache configuration
- Memory
Capacity Budget - Resolved device-wide memory budget consumed by runtime planning.
- Memory
Config - Memory management configuration
- Monitoring
Config - Padding
Config - Padding configuration
- Reusable
Execution Capture Config - Product policy for bounded reusable-execution capture.
- Runtime
Knobs - Engine runtime knobs the CLI/autosizer resolves and injects via the
runtime-config snapshot. The continuous engine reads these from the typed
config instead of
std::env::vars(), so the env bridge stays at the composition root and tests can vary the knobs perEngineConfig. - Sampling
Config - Scheduler
Config - Scheduler configuration
- Security
Config - Sampling configuration presets Security configuration
- Tokenizer
Config - Tokenizer configuration
- Truncation
Config - Truncation configuration
- VNext
Checkpoint Capture Config - Explicit diagnostic capture of semantic vNext activations. Product paths leave this unset; release tooling must supply a dedicated empty directory.
- VNext
Teacher Forcing Config - Bounded token history used only by explicit vNext checkpoint diagnostics.
Enums§
- Backend
Type - Supported backend types
- KvCache
Dtype - KV Cache element dtype (Dim 5 polymorphism point).
- KvCache
Type - KV Cache implementation types
- Padding
Strategy - Padding strategies
- Reusable
Execution Preparation Mode - When a supported backend prepares reusable device programs.
- Scheduling
Policy - Scheduling policies
- Sequence
FitPolicy - Product policy for proving that a sequence fits before prefill admission.
- Tokenizer
Type - Tokenizer algorithms
- Truncation
Strategy - Truncation strategies
- VNext
Diagnostic Fault - Explicit one-shot faults used to prove product-path failure attribution. These are never inferred and remain disabled in normal execution.
Constants§
- DEFAULT_
MAXIMUM_ AUTOMATIC_ EXACT_ DECODE_ WIDTH - Default upper bound for automatically generated exact decode capture widths.
- MAXIMUM_
REUSABLE_ EXECUTION_ STARTUP_ CAPTURE_ WIDTH - Non-configurable upper bound for startup decode capture width.
- MAX_
VNEXT_ TEACHER_ FORCED_ TOKENS