Skip to main content

Crate cera

Crate cera 

Source

Re-exports§

pub use backend::cpu_features::CpuFeatures;
pub use backend::cpu_features::CpuTier;
pub use backend::cpu_features::cpu_features;
pub use backend::cpu_features::cpu_tier;
pub use classifier::BioesPrefix;
pub use classifier::EntitySpan;
pub use classifier::detect_pii;
pub use classifier::extract_spans;
pub use classifier::parse_bioes;
pub use classifier::viterbi_decode;
pub use engine::BackendPreference;
pub use engine::CeraEngine;
pub use engine::EngineConfig;
pub use engine::ModelBytes;
pub use engine::ModelFiles;
pub use engine::ModelMetadata;
pub use model::whisper::Conv1dWeights;
pub use model::whisper::WHISPER_CATALOG;
pub use model::whisper::WhisperConfig;
pub use model::whisper::WhisperDecoderBlockWeights;
pub use model::whisper::WhisperDecoderWeights;
pub use model::whisper::WhisperEncoderBlockWeights;
pub use model::whisper::WhisperEncoderWeights;
pub use model::whisper::WhisperModel;
pub use model::whisper::WhisperModelCatalogEntry;
pub use model::whisper::WhisperSpecialTokens;
pub use model::whisper::WhisperTranscribeOpts;
pub use model::whisper::WhisperWeights;
pub use model::whisper::find_whisper_catalog_entry;
pub use model::whisper::is_whisper_gguf;
pub use model::whisper::transcribe_pcm;
pub use model::whisper::transcribe_pcm_with_tokens;
pub use session::CeraError;
pub use session::FinishReason;
pub use session::GenerateOpts;
pub use session::GenerateSummary;
pub use session::ModalityCapabilities;
pub use session::ModalitySink;
pub use session::Session;
pub use session::SessionConfig;
pub use session::SpecDecode;
pub use sysmem::available_memory_bytes;
pub use sysmem::fits_in_available_memory;
pub use vad::SileroVad;
pub use vad::SpeechTimestamp;
pub use vad::VadConfig;
pub use vad::VadEvent;
pub use vad::VadIterator;
pub use vad::VadSampleRate;

Modules§

audio_engine
Audio-aware generation loop with text ↔ audio modality switching.
backend
bundle
Bundle fetching + caching.
classifier
Token classification and constrained BIOES Viterbi decoding.
convert
SafeTensors -> GGUF model conversion and on-the-fly streaming quantization.
engine
CeraEngine — the owning, loader-aware front door for the core crate.
gguf
grammar
GBNF grammar-constrained decoding (structured output).
kv_cache
lora
LoRA (Low-Rank Adaptation) adapters.
manifest
LeapBundles manifest parser.
model
par
Parallelism facade over rayon.
quant
GGUF quantization formats: block layouts, dequantization, and the vec_dot kernels that consume a block without materializing it.
sampler
session
Stateful, multimodal, cancellable inference session.
spec
Speculative decoding: drafting and greedy accept/reject helpers.
sysmem
Best-effort host physical-memory queries for proactive load pre-flight.
tensor
time
Cross-target time primitives.
tokenizer
tools
Tool (function) calling: tool definitions, chat-template plumbing, output parsing, and grammar generation for constrained tool-call decoding.
turboquant
TurboQuant KV cache key compression (arXiv:2504.19874).
vad
Voice Activity Detection (VAD) using Silero VAD v5.

Constants§

GIT_SHA
Short git SHA of this build, embedded by build.rs. Best-effort: "unknown" when git was unavailable at build time (a packaged source build). Can be pinned via the CERA_GIT_SHA build-env override.
VERSION
Crate version, sourced from Cargo.toml at compile time. Useful for FFI / wrapper crates that want to surface the core lib version to their consumers (e.g. cera-wasm::ceraVersion()) without re-reading the manifest themselves.

Functions§

build_info
Build provenance for telemetry — "<version>+<git-sha>", e.g. "0.4.0+1a2b3c4d5e6f". This is the analog of the llama.cpp build commit that a benchmark harness records alongside results to identify exactly which engine build produced them.