boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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

pub mod bucket_manager;
pub mod bucketed_trainer;
pub mod fused_optimizer;
pub mod grad_sync;
pub mod pipeline;
pub mod trainer;
pub mod zero;
pub mod zero2;
pub mod zero2_trainer;
pub mod zero3;
pub mod zero3_trainer;
pub mod zero_base;
pub mod zero_trainer;
pub(crate) mod zero_trainer_base;

pub use bucket_manager::{GradientBucketManager, param_order_from_graph};
pub use bucketed_trainer::BucketedTrainer;
pub use fused_optimizer::{FusedDistributedOptimizer, FusedOptimizerConfig};
pub use grad_sync::{all_reduce_grads, broadcast_params};
pub use pipeline::gpipe::PipelineSchedule;
pub use pipeline::{
    GpipeSchedule, PipelineStage, Schedule1F1B, ScheduleInterleaved1F1B, ScheduleZeroBubble,
    TrainablePipelineStage, ZeroBubbleStage,
};
pub use trainer::DistributedTrainer;
pub use zero::ZeroStage1;
pub use zero_base::ZeroOptimizer;
pub use zero_trainer::ZeroTrainer;
pub use zero2::ZeroStage2;
pub use zero2_trainer::Zero2Trainer;
pub use zero3::ZeroStage3;
pub use zero3_trainer::Zero3Trainer;