Change8

b8104

📦 llama-cppView on GitHub →
2 features🐛 1 fixes🔧 4 symbols

Summary

This release fixes an issue where an extra newline was inserted between text and media markers in MTMD chat output by introducing a specific `media_marker` type. This resolves token count discrepancies when comparing llama-server output with HF implementations for vision models.

Migration Steps

  1. Refactor to use explicit per type ifs (implied code change related to handling media markers).
  2. Update common/chat.cpp (implied code change).
  3. Update common_chat_templates_apply_legacy (implied code change).

✨ New Features

  • Introduced a new type `media_marker` in MTMD chat processing to correctly handle media markers.
  • Added logic to prevent insertion of newlines before and after media markers during JSON serialization for chat compatibility.

🐛 Bug Fixes

  • Fixed an issue causing an extra newline character (\n) between text and media markers in MTMD chat output, ensuring token count consistency with HF implementations for vision models.

Affected Symbols