Zum Inhalt

Customized Models

Customized Models are the outputs produced by DeepExtension's model training workflows — typically PEFT adapters or checkpoints that are still linked to their original base models. These intermediate artifacts represent models that have completed training but have not yet been merged into standalone full models.

Overview

Customized models are automatically saved upon successful completion of a training job in Model Training.

On the main Customized Models page, you will find a list of all stored adapter-based training results. Each entry includes:

  • Base Model Name: Identifies which base model this adapter is built on
  • Train Name: The name provided when the training job was submitted
  • Model Card: Metadata extracted from the training configuration and evaluation summary
  • Auto-Generated Customized Model Name: Formatted as:
[Customized_model_name] = [base_model_technical_name]_[train_name]_[YYYYMMDD]_[first4ofTrainingUUID]

Currently, only adapter-based artifacts are stored at this stage — full checkpoints are not yet supported but may be added in future releases.

Save a Customized Model

To promote a Customized model to a standalone Saved Model:

  1. Click "Save" on the desired Customized model
  2. Confirm the automatically generated parameters — no additional manual input is required
  3. Click "OK" to submit the saving job

The process runs in background mode, and once complete, the new model will appear in the Complete Models section.

Note: Quantization is not supported during the saving process for the following reasons:

  • There is currently no standardized format for quantized complete models.
  • Quantization can be more appropriately and flexibly applied during the deployment stage instead.

Deploy a Customized Model

To make a Customized model available for real-time use:

  1. Click "Deploy" on the desired Customized model
  2. Provide any required extra parameters (e.g., quantization configuration, deployment environment)
  3. DeepExtension will forward deployment requests to your configured LLM deployment tool
  4. Upon successful deployment, a new entry will appear under Live Models

Note:

  • Deployment requires prior integration with an external deployment backend. DeepExtension itself does not serve models.
  • To deploy via Ollama, a valid Deployment Template File is required. See Base Models for more details.
  • The Deployment Environment must be correctly configured. See Deployment Tool Configuration.
  • Currently, only the following quantization modes are supported: no_quantization, q8_0, q4_K_M, and q4_K_S — as these are the only options supported by the Ollama API.

Delete a Customized Model

To remove a Customized model:

  • Click "Delete" on the model entry
  • This will permanently delete only the adapter or checkpoint file

If this model was already saved as a full model in the Complete Models section, that entry will remain unaffected.

Customized Model Behavior

  • Customized models are always tied to their base models and cannot operate independently until saved and merged
  • The saved name format ensures traceability and version control
  • Checkpoint-based training result saving is not yet available but under consideration for future support

DeepExtension — Manage your training artifacts with precision and traceability