Struct Zero3PerformanceStats 

pub struct Zero3PerformanceStats {

    pub forward_passes: u64,
    pub backward_passes: u64,
    pub optimizer_steps: u64,
    pub total_forward_time: Duration,
    pub total_backward_time: Duration,
    pub total_optimizer_time: Duration,
    pub parameter_transfer_time: Duration,
    pub gradient_sync_time: Duration,
    pub layer_timings: HashMap<String, LayerTimingStats>,
    pub throughput_metrics: ThroughputMetrics,
    pub memory_transfer_metrics: MemoryTransferMetrics,
    pub communication_stats: CommunicationStats,
    pub optimization_efficiency: OptimizationEfficiency,
}

Comprehensive performance statistics for ZeRO-3 operations

Tracks all aspects of ZeRO-3 performance including:

• Forward and backward pass timing
• Parameter transfer and optimization statistics
• Memory management performance
• Distributed synchronization metrics
• Throughput and efficiency measurements

Fields

forward_passes: u64

Number of forward passes completed

backward_passes: u64

Number of backward passes completed

optimizer_steps: u64

Number of optimizer steps completed

total_forward_time: Duration

Total time spent in forward passes

total_backward_time: Duration

Total time spent in backward passes

total_optimizer_time: Duration

Total time spent in optimizer steps

parameter_transfer_time: Duration

Time spent transferring parameters between CPU/GPU

gradient_sync_time: Duration

Time spent synchronizing gradients across ranks

layer_timings: HashMap<String, LayerTimingStats>

Per-layer execution timings

throughput_metrics: ThroughputMetrics

Throughput metrics

memory_transfer_metrics: MemoryTransferMetrics

Memory transfer performance

communication_stats: CommunicationStats

Distributed communication statistics

optimization_efficiency: OptimizationEfficiency

Optimization efficiency metrics

Implementations

impl Zero3PerformanceStats

pub fn new() -> Self

Create new performance statistics

pub fn record_forward_pass(&mut self, duration: Duration, num_tokens: usize)

Record a completed forward pass

pub fn record_backward_pass(&mut self, duration: Duration, num_tokens: usize)

Record a completed backward pass

pub fn record_optimizer_step(&mut self, duration: Duration, num_params: usize)

Record a completed optimizer step

pub fn record_layer_execution(&mut self, layer_name: String, duration: Duration)

Record layer execution timing

pub fn record_layer_backward(&mut self, layer_name: String, duration: Duration)

Record layer backward pass timing

pub fn record_parameter_transfer( &mut self, duration: Duration, bytes_transferred: usize, direction: TransferDirection, )

Record parameter transfer operation

pub fn record_gradient_sync( &mut self, duration: Duration, num_gradients: usize, world_size: usize, )

Record gradient synchronization

pub fn record_communication( &mut self, operation: CommunicationOperation, duration: Duration, bytes: usize, )

Record distributed communication operation

pub fn average_forward_time(&self) -> Duration

Get average forward pass time

pub fn average_backward_time(&self) -> Duration

Get average backward pass time

pub fn average_optimizer_time(&self) -> Duration

Get average optimizer step time

pub fn get_tokens_per_second(&self) -> f64

Get tokens per second throughput

pub fn get_memory_bandwidth_gbps(&self) -> f64

Get memory transfer bandwidth in GB/s

pub fn get_communication_efficiency(&self) -> f64

Get communication efficiency metrics

pub fn get_training_efficiency(&self) -> f64

Get overall training efficiency score (0.0 to 1.0)

pub fn get_performance_summary(&self) -> PerformanceSummary

Get detailed performance summary

pub fn reset(&mut self)

Reset all statistics

pub fn merge(&mut self, other: &Zero3PerformanceStats)

Merge statistics from another instance (useful for distributed aggregation)

Trait Implementations

impl Clone for Zero3PerformanceStats

fn clone(&self) -> Zero3PerformanceStats

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Zero3PerformanceStats

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

Formats the value using the given formatter.

impl Default for Zero3PerformanceStats

fn default() -> Self

Returns the “default value” for a type.