renderers

Renderers convert chat-style messages into token sequences for training and generation, bridging user prompts to model inputs. They provide a base API to format and parse data, ensuring inputs are correctly tokenized for large language models.

A renderer registry provides a centralized factory to obtain named renderers for specific model families. This allows you to dynamically select and manage the correct tokenization strategy, ensuring proper prompt formatting across different language models.

Yes, renderers support model fine-tuning workflows and integrate vision inputs. They are designed to handle multimodal data, transforming both text and image-based chat messages into unified token sequences for training and inference.

A pickleable renderer is a serializable interface that maintains its state across distributed execution environments. This serialization is required to safely transfer rendering logic across multiple compute nodes during distributed training.

Yes, the renderer interface fully supports TrainOnWhat settings. This integration allows you to precisely control which parts of the chat messages are tokenized and included in the training loss calculation.

The best way to customize tokenization is to use the renderer base class and its registry. This allows you to define model-family compatible formatting rules and integrate specific tools, ensuring prompts are parsed correctly for your target architecture.