feat: multiturn support for ChatML

[ysjprojects]

This pull request enhances the prompt formatting system to better support multi-turn conversations and training/inference scenarios for several prompt styles. The most important changes include adding a supports_multiturn attribute, introducing an add_generation_prompt parameter, and refactoring prompt application logic for consistency and extensibility.

Multi-turn and Prompt Formatting Enhancements:

• Added a supports_multiturn attribute to PromptStyle and its subclasses (Llama3, R1Base, ChatML), indicating whether each style supports multi-turn conversations. [1] [2] [3] [4]
• Updated the apply method signature for multi-turn styles to accept an add_generation_prompt argument, allowing control over whether a generation prompt is appended (for inference vs. training). [1] [2] [3]

Prompt Logic Refactoring:

• Moved the has_system_prompt utility function to the module level for reuse and clarity, and updated its usage in prompt formatting logic. [1] [2] [3] [4]

Training/Inference Handling:

• Modified prompt formatting logic in Llama3, R1Base, and ChatML to conditionally add the generation prompt based on the add_generation_prompt parameter, supporting both inference and training workflows. [1] [2] [3]

Testing Improvements:

• Added and updated tests to verify that prompt formatting behaves correctly for both inference and training cases, specifically checking the behavior of the add_generation_prompt parameter.