Struct ParallelismConfig 

pub struct ParallelismConfig {
    pub dp_size: usize,
    pub mp_size: usize,
    pub pp_size: usize,
    pub num_micro_batches: usize,
    pub gradient_accumulation: bool,
    pub accumulation_steps: usize,
    pub activation_checkpointing: bool,
    pub comm_backend: CommBackend,
    pub pipeline_schedule: PipelineSchedule,
    pub memory_optimization: MemoryOptimization,
}

3D Parallelism Configuration

Combines data parallelism (DP), model parallelism (MP), and pipeline parallelism (PP) to efficiently scale transformer training across multiple GPUs and nodes.

Key concepts:

• Data Parallelism: Each process has a full copy of the model and trains on different data
• Model Parallelism: Model parameters are split across processes within a layer
• Pipeline Parallelism: Model layers are split across processes, enabling pipeline execution

The total number of processes = dp_size * mp_size * pp_size

Fields

dp_size: usize

Data parallel group size

mp_size: usize

Model parallel group size

pp_size: usize

Pipeline parallel group size

num_micro_batches: usize

Number of micro-batches for pipeline parallelism

gradient_accumulation: bool

Whether to use gradient accumulation

accumulation_steps: usize

Number of gradient accumulation steps

activation_checkpointing: bool

Whether to use activation checkpointing

comm_backend: CommBackend

Communication backend preference

pipeline_schedule: PipelineSchedule

Pipeline scheduling strategy

memory_optimization: MemoryOptimization

Memory optimization level

Trait Implementations

impl Clone for ParallelismConfig

fn clone(&self) -> ParallelismConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ParallelismConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ParallelismConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for ParallelismConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for ParallelismConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.