Module checkpoint 

Saving and restoring LoRA adapters and training state.

Adapters are written as safetensors (F32) with their LoRA settings in the file’s metadata, so ferrotensor-cli generate --lora can rebuild the same adapter shapes. TrainingCheckpoint (JSON) holds the step counter, optimizer moments and loss-scaler state for resuming.

Structs

LoraCheckpoint

Tensors and metadata read from an adapter file.

LoraMetadata

LoRA settings stored in an adapter file’s metadata.

TrainingCheckpoint

Everything besides the adapter weights needed to resume a run.

TrainingState

Progress and settings of a training run (also read by ferrotensor-monitor from training_state.json).

Functions

apply_checkpoint

Copy checkpoint values into the matching tensors.

load_lora_checkpoint

Load LoRA adapter weights from a safetensors file (F32, F16 or BF16 tensors are accepted and returned as F32).

load_weights

Overwrite each tensor in tensors with the weight of the same name in weights_ctx; names that are missing are logged and skipped.

save_lora_checkpoint

Save LoRA adapter weights (F32 safetensors) to path, with metadata in the header. The file is written to a temporary name and renamed, so a crash never leaves a truncated adapter behind.