The llama.cpp project released version b9952, which modifies the handling of key-query masks for DeepSeek V4 models. The update ensures all KQ masks are set to f16 when Flash Attention is used, while also removing the zero attention bias.

  • Removed dead code that previously repeated the unified raw_k cache for each stream in DeepSeek V4.
  • Provided binaries for macOS (Apple Silicon and Intel), iOS, Linux, Android, Windows, and openEuler across CPU, GPU, and specialized hardware backends.