ggml/llama.cpp releases: b10254

TL;DR

New templates for DeepSeek V4 Flash 0731 were added to ggml/llama.cpp, aligning them with official encoders and updating the chat functionality. This update ensures compatibility and improves parser behavior without altering existing settings.

Detailed Summary

The ggml/llama.cpp project updated its chat template to align with DeepSeek V4 Flash 0731, specifically addressing the default drop_thinking setting for DeepSeek V4 history. This update involves contributors who are part of the ggml community and impacts how conversations are managed within the DeepSeek V4 framework, enhancing consistency with official encoders while maintaining parser behavior integrity.

Key Points

• • chat : add new template for DeepSeek V4 Flash 0731
• • common/chat: update DeepSeek V4 templates
• • Default drop_thinking for DeepSeek V4 history so prior thinking is not included