ferrum-interfaces 0.8.4

Core trait contracts for the Ferrum LLM inference engine
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100

//! Core interface definitions for the Ferrum inference framework
//!
//! This crate carries the stable, GPU-free trait contracts shared across
//! the workspace: model execution, scheduling, KV cache management,
//! tokenization, sampling, and the lifecycle/modality engine traits.
//! Hardware backends live in `ferrum-kernels` (the `Backend<B>` trait
//! and its supertraits); only types that compile without GPU features
//! belong here.

#![allow(async_fn_in_trait)]

pub mod engine;
pub mod kv_cache;
pub mod kv_dtype;
pub mod model_executor;
pub mod recurrent_state;
pub mod sampler;
pub mod scheduler;
pub mod tensor;
pub mod tokenizer;
#[allow(
    dead_code,
    unused_imports,
    reason = "reviewed vNext contracts stay isolated and byte-stable until the product runtime migration"
)]
pub mod vnext;

// Re-export core traits and important types
pub use engine::InferenceEngine;
pub use kv_cache::{
    AllocationRequest, BlockTable, CacheHandleStats, KvCacheHandle, KvCacheManager,
};
pub use kv_dtype::{KvBf16, KvDtypeKind, KvFp16, KvFp8, KvInt8};
pub use model_executor::{
    DecodeInput, DecodeOutput, ExecutorAdmissionEpochs, ExecutorPrefillAdmission,
    ExecutorPrefillAdmissionDecision, ExecutorPrefillAdmissionReceipt,
    ExecutorPrefillMaintenanceBlocker, ExecutorPrefillMaintenanceDeferral,
    ExecutorPrefillMaintenanceOutcome, ExecutorPrefillMaintenanceStage, ExecutorSamplingOutput,
    ModelExecutor, PlanRuntimeBatchDecodeOutcome, PlanRuntimeBatchPrefillOutcome,
    PlanRuntimeDecodeInput, PlanRuntimeDecodeOutput, PlanRuntimePrefillAuthority,
    PlanRuntimePrefillCompletion, PlanRuntimePrefillInput, PlanRuntimePrefillOutcome,
    PlanRuntimePrefillOutput, PlanRuntimePrefillProduct, PrefillInput, PrefillOutput,
};
pub use recurrent_state::{
    RecurrentStateHandle, RecurrentStateHandleStats, RecurrentStateManager,
    RecurrentStateManagerStats, RecurrentStateResumePolicy, RecurrentStateSpec,
    RecurrentStateTensorSpec,
};
pub use sampler::{
    LogitsProcessor, Sampler, SamplingConfig, SamplingContext, SamplingRng,
    SAMPLING_RNG_ALGORITHM_ID,
};
pub use scheduler::{BatchHint, BatchPlan, Scheduler as SchedulerInterface};
pub use tensor::{TensorFactory, TensorLike, TensorOps, TensorRef};
pub use tokenizer::{IncrementalTokenizer, Tokenizer, TokenizerFactory, TokenizerInfo};

// Re-export types from ferrum-types, avoiding conflicts
pub use ferrum_types::{
    config::BackendConfig,
    // Config types - use fully qualified names to avoid conflicts
    config::EngineConfig,
    config::SchedulerConfig,
    config::TokenizerConfig,
    BatchId,
    BlockId,
    ClientId,
    ComponentHealth,
    ComponentStatus,
    DataType,
    // Device types
    Device,
    EngineMetrics,
    EngineStatus,
    FerrumError,
    FinishReason,
    HealthStatus,
    // Requests and responses
    InferenceRequest,
    InferenceResponse,
    // Metrics
    MemoryUsage,
    ModelId,
    // Model types
    ModelInfo,
    ModelSource,
    ModelType,
    Priority,
    // IDs
    RequestId,
    Result,
    // Sampling
    SamplingParams,
    SchedulerStats,
    SessionId,
    SpecialTokens,
    StreamChunk,
    TaskId,
    // Basic types
    TokenId,
};