Skip to main content

Crate sokuto_runtime

Crate sokuto_runtime 

Source
Expand description

§sokuto-runtime

高速トークナイズおよび ONNX Runtime による超低遅延推論パイプラインを提供するクレート。

§概要

  • Hugging Face tokenizers を用いたトークナイズ処理。
  • ONNX Runtime (ort) を用いた単一フォワードパス決定ヘッド推論。
  • 候補数が多い場合の粗密探索(Coarse-to-Fine)やバッチ推論。

Re-exports§

pub use dag::DagDefinition;
pub use dag::DagError;
pub use dag::DagExecutionResult;
pub use dag::DagNode;
pub use dag::DagNodeType;
pub use dag::InProcessDagExecutor;
pub use dag::LowLatencyResourcePool;
pub use dag::LowLatencySessionPool;
pub use dag::NodeCondition;
pub use dag::NodeConditionOp;
pub use dag::PooledResource;
pub use dag::PooledSession;
pub use dag::StepExecutionResult;
pub use dag::validate_dag;
pub use engine::AlignedBuffer64;
pub use engine::BatchScratchpad;
pub use engine::CacheCompressionMode;
pub use engine::CachedStateFeatures;
pub use engine::CachedTensorPayload;
pub use engine::CandidateEmbeddingCache;
pub use engine::CoarseScorer;
pub use engine::CoarseToFineConfig;
pub use engine::DEFAULT_COARSE_THRESHOLD;
pub use engine::DEFAULT_DUAL_MAX_BATCH_CHUNK_SIZE;
pub use engine::DEFAULT_MAX_BATCH_CHUNK_SIZE;
pub use engine::DEFAULT_NEGATIVE_KEYS;
pub use engine::DEFAULT_TOP_M_CANDIDATES;
pub use engine::DualSessionEngine;
pub use engine::EmbeddingCoarseScorer;
pub use engine::ExecutionProvider;
pub use engine::FilteredCandidates;
pub use engine::HierarchicalCoarseScorer;
pub use engine::InferenceEngine;
pub use engine::LexicalCoarseScorer;
pub use engine::ModelDrivenCoarseScorer;
pub use engine::OptimizationLevel;
pub use engine::SessionConfig;
pub use engine::StateCacheManager;
pub use engine::filter_top_candidates;
pub use engine::is_aligned_64;
pub use engine::is_negative_candidate;
pub use engine::reconstruct_probabilities;
pub use error::Result;
pub use error::RuntimeError;
pub use hierarchical::CoarseToFineRouter;
pub use hierarchical::DEFAULT_BEAM_MARGIN_THRESHOLD;
pub use hierarchical::DEFAULT_CATCHALL_THRESHOLD;
pub use hierarchical::DEFAULT_COARSE_POWER_ALPHA;
pub use hierarchical::DEFAULT_ENTROPY_TEMP_GAMMA;
pub use hierarchical::DEFAULT_ESCALATE_CONFIDENCE_THRESHOLD;
pub use hierarchical::DirichletCalibrator;
pub use hierarchical::HierarchicalEngine;
pub use hierarchical::HierarchicalExecutionTrace;
pub use hierarchical::HierarchicalMapping;
pub use hierarchical::HierarchicalMappingBuilder;
pub use hierarchical::HierarchicalResult;
pub use hierarchical::HierarchicalRouterConfig;
pub use hierarchical::combine_and_reconstruct_probabilities;
pub use hierarchical::compute_fine_temperature;
pub use hierarchical::compute_normalized_entropy;
pub use hierarchical::softmax_f64;
pub use hierarchical::temperature_scaled_softmax;
pub use tokenizer::BatchTokenizedQuestions;
pub use tokenizer::JevTokenizer;
pub use tokenizer::TokenizedQuestion;
pub use sokuto_core as core;

Modules§

dag
インプロセス DAG 実行器モジュール
engine
推論エンジンモジュール
error
ランタイムエラー型定義モジュール
hierarchical
粗密二段階階層ルーティング (Coarse-to-Fine Hierarchical Routing)
tokenizer
高速トークナイザー統合モジュール

Functions§

version
ランタイムの初期化確認用関数。