Saving and Loading

Guide to saving trained model weights and loading them for inference.

Overview

Weaver provides simple APIs for:

• Saving model weights after training
• Loading saved weights for inference
• Managing multiple model checkpoints

Saving Weights

save_weights_for_sampler()

Save model weights and get a sampling client:

from weaver import ServiceClient

# After training
sampling_client = training_client.save_weights_and_get_sampling_client(
    name="my-trained-model"
)

# Now you can sample from the trained model
result = sampling_client.sample(...)

Parameters:

• name (str): Name for this saved model

Returns:

A SamplingClient ready to generate text from the trained model.

Save Path

The method returns the path where weights were saved:

save_path = training_client.save_weights_for_sampler(name="checkpoint-100")
print(f"Weights saved to: {save_path}")

This path can be used later to create a sampling client.

Loading Weights

create_sampling_client()

Create a sampling client from saved weights:

from weaver import ServiceClient

service_client = ServiceClient()

# Load from saved path
sampling_client = service_client.create_sampling_client(
    model_path="/path/to/saved/weights",
    base_model="Qwen/Qwen3-8B",
)

# Sample from loaded model
result = sampling_client.sample(...)

Parameters:

• model_path (str): Path to saved model weights
• base_model (str): Base model identifier
• model_id (str, optional): Model ID for tracking

Next Steps

• Learn about Training and Sampling - Core APIs
• Explore Loss Functions - Available losses
• Check Model Lineup - Supported models