skippy-runtime 0.77.0

pub use skippy_ffi::Status;
#[cfg(test)]
pub(crate) use skippy_ffi::TensorRole;

mod activation;
mod capability_probe;
mod checkpoint;
mod config;
mod decision;
mod devices;
mod error;
mod gguf_writer;
mod kv_pages;
mod laya;
mod logging;
mod media;
mod native;
mod native_mtp;
mod native_test_evidence;
mod ngram;
pub mod package;
mod path_cstring;
mod runtime_event_reporter;
mod runtime_events;
mod session;
mod stage_planning;
mod types;

pub use activation::{
    DecodeFrameBatchRequest, IterationBatchPhase, IterationBatchRequest, PartialBatchExecution,
};
pub use capability_probe::{CapabilityReport, probe_capabilities};
pub use checkpoint::{CheckpointQuantization, is_safetensors_checkpoint};
pub use config::{
    FlashAttentionType, GGML_TYPE_F16, GGML_TYPE_F32, GGML_TYPE_I32, GGML_TYPE_Q4_0,
    GGML_TYPE_Q8_0, GlmDsaPolicy, LLAMA_SERVER_DEFAULT_N_BATCH, LLAMA_SERVER_DEFAULT_N_UBATCH,
    MtpSource, RuntimeConfig, SKIPPY_UNIFIED_KV_DEFAULT_N_BATCH, SplitMode, parse_cache_type,
};
pub use decision::{
    DecisionError, DecisionModel, DecisionOutput, DecisionQuestion, DecisionQuestionKind,
    DecisionRequest, DecisionValue,
};
pub use devices::{BackendDevice, BackendDeviceType, backend_devices};
pub(crate) use error::ensure_ok;
pub use gguf_writer::{
    ModelInfo, write_gguf_from_parts, write_gguf_from_parts_consuming,
    write_gguf_metadata_from_parts,
};
pub use kv_pages::{decode_cachegen_kv_page, encode_cachegen_kv_page};
pub use laya::{LayaMemory, LayaModel, LayaModelInfo};
pub use logging::{
    LLAMA_LOG_LEVEL_DEBUG, MeasuredNativeBuffers, NativeLogEvent, NativeLogParserMode,
    NativeLogParserPolicy, configure_native_log_parser, disable_verbose_native_logs,
    enable_verbose_native_logs, measured_native_buffers, redirect_native_logs_to_file,
    register_filtered_native_logs, restore_native_logs, set_filtered_native_logs_enabled,
    suppress_native_logs, unregister_filtered_native_logs, write_native_log_note,
};
pub use media::{SpeechAudio, SpeechOutputFormat, SpeechSynthesisConfig};
pub use native::{
    ModelWorkload, PoolingType, StageModel, StageModelReader, SystemOneReadSlot, WorkloadInfo,
};
pub use native_mtp::NativeMtpDraft;
pub use native_test_evidence::write_evidence_marker;
pub use ngram::{Cache as NgramCache, NGRAM_CACHE_MAX_NGRAM};
pub use runtime_event_reporter::{
    RECORD_RING_CAPACITY, buffered_runtime_events, clear_runtime_event_reporter,
    deliver_runtime_event_for_test, drain_runtime_events, dropped_runtime_events,
    install_runtime_event_reporter, rejected_runtime_events, runtime_event_reporter_installed,
};
pub use runtime_events::{
    INLINE_DETAIL_BYTES, MODEL_OPEN_RECORD_CAPACITY, ModelOpenEventQueue, NativeEventRecord,
    OperationId, RecordRejection, RuntimeEvent, RuntimeEventCategory, RuntimeEventEmitterKind,
    RuntimeEventFailureCode, RuntimeEventKind, RuntimeEventProgressUnit, next_operation_id,
};
pub use session::{DecodeBatchRequest, StageSession};
pub use skippy_ffi::LoadMode as RuntimeLoadMode;
pub use skippy_ffi::MAX_DRY_SEQUENCE_BREAKER_BYTES;
pub use skippy_ffi::{
    ACTIVATION_FRAME_VERSION, ACTIVATION_IDENTITY_BYTES, ACTIVATION_MAX_DIMS, ACTIVATION_MAX_PARTS,
    ACTIVATION_PART_OPTIONAL,
};
pub use stage_planning::{
    GgufStageRuntimePlan, gguf_shard_paths, plan_gguf_stage_resident_tensor_names,
    plan_gguf_stage_resident_tensor_names_for_range, plan_gguf_stage_runtime_plan_for_range,
    plan_gguf_stage_runtime_plans,
};
// KV page descriptor flags. Re-exported so callers can read a page's layout
// without taking a direct dependency on the raw ABI crate.
pub use skippy_ffi::{KV_PAGE_FLAG_HAS_K_IDX, KV_PAGE_FLAG_V_TRANSPOSED};
pub use types::{
    ACTIVATION_BOUNDARY_DESC_VERSION, ActivationBoundaryDesc, ActivationDesc, ActivationFrame,
    ActivationPartDesc, ChatReasoningFormat, ChatTemplateJsonOptions, ChatTemplateJsonResult,
    ChatTemplateMessage, ChatTemplateOptions, DecodeFrameBatchOutput, DrySamplingConfig,
    GenerationSignalWindow, IterationBatchOutput, IterationSample, LoadedModelCapability,
    LogitBias, MAX_LOGIT_BIAS, MediaInput, MediaPrefill, MediaPrefillChunkFrame, MediaPrefillFrame,
    ModelStateKind, ReasoningBudget, RuntimeKvPage, RuntimeKvPageDesc, SamplingConfig, TensorInfo,
    TokenSignal, XtcSamplingConfig,
};

#[cfg(feature = "dynamic-native-runtime")]
pub use skippy_ffi::{
    NativeRuntimeLoadError, load_native_runtime_libraries, load_native_runtime_library,
    native_runtime_loaded,
};

#[cfg(test)]
use error::format_skippy_error;

#[cfg(not(feature = "dynamic-native-runtime"))]
pub fn native_runtime_loaded() -> bool {
    true
}

#[cfg(not(feature = "dynamic-native-runtime"))]
/// No-op for statically linked Skippy runtime builds.
///
/// # Safety
///
/// Static builds resolve the native ABI at process link/load time, so this
/// function does not dereference the supplied path or mutate loader state.
pub unsafe fn load_native_runtime_library(
    _path: impl AsRef<std::path::Path>,
) -> Result<(), skippy_ffi::NativeRuntimeLoadError> {
    Ok(())
}

#[cfg(not(feature = "dynamic-native-runtime"))]
/// No-op for statically linked Skippy runtime builds.
///
/// # Safety
///
/// Static builds resolve the native ABI at process link/load time, so this
/// function does not dereference the supplied paths or mutate loader state.
pub unsafe fn load_native_runtime_libraries<I, P>(
    _paths: I,
) -> Result<(), skippy_ffi::NativeRuntimeLoadError>
where
    I: IntoIterator<Item = P>,
    P: AsRef<std::path::Path>,
{
    Ok(())
}

#[cfg(test)]
include!("tests.rs");