Saving and Loading

This page covers the low-level MinT checkpoint primitives. For end-to-end operational workflows, see the dedicated Advanced Checkpoint pages.

Save for inference

sampler_path = training_client.save_weights_for_sampler(name="0000").result().path

Use this when you want a weights-only checkpoint for sampling or export.

Save for training resume

resume_path = training_client.save_state(name="0010").result().path

Use this when you want to preserve both weights and optimizer state.

Load for full training resume

For optimizer-preserving resume, create a fresh LoRA training client with the same model/rank/options, then call load_state_with_optimizer(...):

training_client = service_client.create_lora_training_client(
    base_model=model,
    rank=rank,
    train_mlp=True,
    train_attn=True,
    train_unembed=True,
)
training_client.load_state_with_optimizer(resume_path).result()

Load weights only

training_client.load_state(resume_path).result()

load_state(...) restores weights only. It does not preserve optimizer state, so do not use it as the full training-resume path.

Related advanced workflows

• Save
• Download
• Upload
• Resume