#[cfg(feature = "distributed")]
pub mod collective;
#[cfg(feature = "distributed")]
pub mod context;
#[cfg(feature = "distributed")]
pub mod data_parallel;
#[cfg(feature = "distributed")]
pub mod fsdp;
#[cfg(feature = "distributed")]
pub mod mesh;
#[cfg(feature = "distributed")]
pub mod pipeline;
pub mod placement;
#[cfg(feature = "distributed")]
pub mod plan;
#[cfg(feature = "distributed")]
pub mod rule;
#[cfg(feature = "distributed")]
pub mod tensor_parallel;
#[cfg(feature = "distributed")]
pub use collective::{
CollectiveDType, CollectiveError, CollectiveKind, CollectiveReductionDType, GroupId, StreamId,
validate_collective_dtype, validate_collective_reduction,
};
#[cfg(feature = "distributed")]
pub use context::{
ContextError, ContextFailure, DistributedContext, DistributedContextHandle,
DistributedContextState, DistributedIdentity, LOCAL_CUDA_DEVICE_ENV, RANK_ENV,
RENDEZVOUS_ADDR_ENV, RENDEZVOUS_TIMEOUT_MS_ENV, RUN_ID_ENV, RunId, StaticTwoRank,
TWO_RANK_WORLD, WORLD_SIZE_ENV,
};
#[cfg(all(feature = "distributed", feature = "std"))]
pub use context::{
DynRendezvousConfig, RankLaunch, RendezvousEndpoint, StaticRendezvousConfig, TwoRankLaunchPlan,
};
#[cfg(feature = "distributed")]
pub use data_parallel::{
DataParallelDType, DataParallelError, DataParallelPlan, DataParallelPlanBuilder,
GradientDescriptor, GradientId, TwoRankDataParallel, validate_data_parallel_dtype,
};
#[cfg(feature = "distributed")]
pub use fsdp::{
FsdpError, FsdpMemoryReport, FsdpParameterDescriptor, FsdpParameterId, FsdpPlan,
FsdpPlanBuilder, ZeROStage,
};
#[cfg(feature = "distributed")]
pub use pipeline::{
ActivationCheckpoint, GPipe, OneForwardOneBackward, PipelineAction, PipelineBoundaryId,
PipelineClock, PipelineDType, PipelineError, PipelinePhase, PipelinePlan, PipelinePlanBuilder,
PipelineSchedule, PipelineScheduleDescriptor, PipelineTransfer, PipelineTransferDescriptor,
StaticPipelineSchedule, TwoRankPipeline, validate_microbatches, validate_pipeline_dtype,
};
pub use placement::{ConstPlacement, Local, Placement, PlacementKind};
#[cfg(feature = "distributed")]
pub use placement::{
Max, Mean, Min, Partial, PartialReduction, PipelineStage, PlacementAxis, PlacementBuf,
PlacementOn, Prod, Replicated, Sharded, Sum,
};
#[cfg(feature = "distributed")]
pub use plan::{
AgreedPlan, CollectiveDescriptor, CollectivePlan, CollectivePlanBuilder, CollectiveTag,
CommunicationEvidence, HybridPlanDType, HybridPlanError, HybridPlanReport, HybridPlanner,
HybridWorkload, MemoryLimit, ParallelOptions, ParallelStrategy, ParallelStrategyKind,
PlanError, PlanObjective, PlanSummary, PlannedCollectiveTransition, PlanningCollectiveKind,
RejectedStrategy, SequenceToken, ShardEvidence, StaticParallelOptions, StrategyCandidate,
StrategyRejection, StrategySet, TwoRankPlanningTopology, WorkloadField, preflight,
validate_hybrid_plan_dtype,
};
#[cfg(feature = "distributed")]
pub use rule::{
CompletePlacement, DistributedError, DistributedInputs, DistributedRule, ElementwisePlacement,
LegalTransition, PlacementTransition, PlacementTransitionRule, ReduceShardedAxis,
ShardDivisible, ShardRemainderPolicy, ValidatedDistributed, validate_pipeline_stage,
validate_shard, validate_transition,
};
#[cfg(feature = "distributed")]
pub use tensor_parallel::{
TensorParallelCollective, TensorParallelDType, TensorParallelDescriptor,
TensorParallelDimension, TensorParallelError, TensorParallelId, TensorParallelPlan,
TensorParallelPlanBuilder, TwoRankTensorParallel, TwoWayShard, validate_tensor_parallel_dtype,
validate_two_way_extent,
};