Struct GradientCheckpointer 

pub struct GradientCheckpointer {
    pub config: CheckpointerConfig,
    pub global_step: u64,
    /* private fields */
}

Production-grade gradient checkpointing engine.

Collects GcGradientTensor objects via accumulate, combines them into a GcGradientCheckpoint via flush, and maintains a bounded history of the most recent checkpoints.

Fields

config: CheckpointerConfig

Runtime configuration.

global_step: u64

Cumulative step counter. Incremented once per accumulate call.

Implementations

impl GradientCheckpointer

pub fn new(config: CheckpointerConfig) -> Self

Create a new checkpointer with the given configuration.

pub fn accumulate(&mut self, gradient: GcGradientTensor)

Push a gradient tensor into the accumulation buffer for its layer.

The global step counter is incremented once per call.

pub fn flush( &mut self, now: u64, ) -> Result<GcGradientCheckpoint, GradientCheckpointerError>

Flush accumulated gradients into a new checkpoint.

Applies the configured accumulation mode over all pending tensors for each layer, optionally clips the resulting global norm, computes a FNV-1a checksum, and stores the checkpoint. If the checkpoint count exceeds max_checkpoints, the oldest entry is evicted.

Errors

Returns GradientCheckpointerError::NoAccumulatedGradients when there are no pending tensors.

pub fn replay(&self, checkpoint: &GcGradientCheckpoint) -> Vec<GcGradientTensor>

Return gradient tensors from checkpoint sorted by layer_id.

pub fn latest_checkpoint(&self) -> Option<&GcGradientCheckpoint>

Return a reference to the most recently flushed checkpoint, if any.

pub fn checkpoint_by_id( &self, id: CheckpointId, ) -> Option<&GcGradientCheckpoint>

Return a reference to the checkpoint with the given id, if it still resides in the bounded history.

pub fn diff( &self, a: &GcGradientCheckpoint, b: &GcGradientCheckpoint, ) -> HashMap<String, f64>

Compute the per-layer L2 distance between two checkpoints.

For each layer present in either checkpoint the function returns the L2 distance between the corresponding gradient value vectors. If a layer is absent in one of the checkpoints its contribution is 0.0.

pub fn global_norm(&self, checkpoint: &GcGradientCheckpoint) -> f64

Compute the global L2 norm of a checkpoint (sqrt of sum of squared per-layer norms).

pub fn pending_layers(&self) -> Vec<&str>

Return a sorted list of layer IDs that have pending (unflushed) gradients.

pub fn pending_count(&self) -> usize

Return the total count of pending gradient tensors across all layers.

pub fn stats(&self) -> GcCheckpointerStats

Return a snapshot of current checkpointer statistics.