weights

To merge LoRA adapters into a base model, the weights Skill uses the build_hf_model function. It takes adapter_path inputs and a base_model selection to combine them into a deployable HuggingFace artifact.

Downloading trained adapters from Tinker is handled by the weights Skill's download function. It automates retrieving the trained LoRA adapter files from Tinker and stores them locally for subsequent processing.

Yes, you can publish a merged model to HuggingFace Hub using the publish_to_hf_hub function. It supports uploading finished artifacts directly to HFHub, including options for private repository publishing.

Automating the fine-tuned model weight lifecycle is streamlined by standardizing adapter downloads, LoRA merging, and HuggingFace publishing. This reduces manual overhead and errors when moving weights between systems for deployment.

Yes, the weight lifecycle workflow supports dtype options during the merging process. When using build_hf_model to merge LoRA adapters into the base model, you can specify the desired data type for the output artifact.