●Change8

b7657

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

Summary

This release introduces targeted Vulkan warptile tuning optimizations for Intel Xe2/Xe3 hardware and resolves a regression related to bf16 matrix multiplications.

Migration Steps

  1. If using Intel hardware, be aware that warptile tuning parameters have been adjusted (m_warptile changed to l_warptile).

✨ New Features

  • Improved Vulkan warptile tuning specifically for Intel Xe2/Xe3 architectures.

🐛 Bug Fixes

  • Fixed a regression in bf16 matmuls by changing warptile tuning from m_warptile to l_warptile.
  • Fixed back formatting issues.

🔧 Affected Symbols

warptile tuningbf16 matmuls