Principle:Togethercomputer Together python Model Download
| Attribute | Value |
|---|---|
| Principle Name | Model_Download |
| Overview | Mechanism for downloading fine-tuned model weights or checkpoint adapters from Together AI to local storage. |
| Domain | MLOps, Fine_Tuning |
| Repository | togethercomputer/together-python |
| Last Updated | 2026-02-15 16:00 GMT |
Description
Model download retrieves the trained model artifacts from a completed fine-tuning job. The download mechanism supports several configuration options to control what is downloaded and where it is stored:
Checkpoint Selection
A fine-tuning job can produce multiple checkpoints during training (controlled by the n_checkpoints parameter at job creation). The download mechanism supports:
- Final model -- Downloaded by default when no step is specified.
- Intermediate checkpoints -- Downloaded by specifying a step number, either via the
checkpoint_stepparameter or by using the compound ID format"ft-id:step".
Checkpoint Types
For LoRA-trained models, the download supports three output types:
- Default (
"default") -- For LoRA jobs, this is automatically treated as"merged". For full training jobs, this downloads the model output. - Merged (
"merged") -- Downloads the base model with LoRA weights merged into the full model parameters. This produces a standalone model that can be loaded without LoRA-specific infrastructure. - Adapter (
"adapter") -- Downloads only the LoRA adapter weights. This is significantly smaller than the merged model and requires the base model to be loaded separately for inference.
For full training jobs, only the "default" checkpoint type is allowed.
Transfer Handling
The download transfers a compressed archive from Together's servers to local storage. The process includes:
- Progress tracking -- The download manager provides progress indication during the transfer.
- Output path control -- The file can be saved to a custom path, or defaults to the current working directory with the model name as the filename.
- Metadata fetching -- The download manager retrieves file metadata to report the downloaded file size.
Usage
Use this after a fine-tuning job completes successfully, to retrieve the trained model for local deployment or further processing. The typical workflow:
- Monitor the fine-tuning job until
statusis"completed". - Call
client.fine_tuning.download(id)to download the final model. - Optionally specify a
checkpoint_stepto download an intermediate checkpoint. - Use the downloaded model files for local inference, evaluation, or further training.
For LoRA jobs, choosing between "merged" and "adapter" checkpoint types depends on the deployment target:
- Use
"merged"for standalone deployment where the base model will not be separately loaded. - Use
"adapter"for efficient storage when the base model is already available, or for further experimentation with different LoRA adapter combinations.
Theoretical Basis
Downloading trained model artifacts is the final step in the fine-tuning lifecycle, bridging cloud-based training with local deployment or evaluation. The separation of merged and adapter downloads for LoRA models reflects the mathematical structure of LoRA itself: the weight update is decomposed as W + BA where B and A are the low-rank matrices. The adapter download contains only B and A, while the merged download contains the full W + BA.
Offering intermediate checkpoints for download enables practitioners to select the best-performing checkpoint based on validation metrics, implementing a form of model selection that can yield better results than simply using the final training step.