Change8

b9119

📦 llama-cppView on GitHub →
🐛 1 fixes

Summary

This release addresses a specific performance regression on Windows Vulkan workloads involving Intel GPUs and BF16 computations. A refactoring was implemented to conditionally use l_warptile.

Migration Steps

  1. Use l_warptile only when coopamt is available for BF16.

🐛 Bug Fixes

  • Fixed a Windows performance regression for Vulkan BF16 workloads on Intel Xe2 and newer GPUs.