Change8

b9014

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

Summary

This release introduces layer norm operations support for ggml-webgpu, including stability improvements and stride adjustments in the shader implementation. It also provides updated pre-compiled binaries for macOS, Linux, Android, Windows, and openEuler.

Migration Steps

  1. If using ggml-webgpu, note that the implementation for layer norm ops was changed from using Kahan summation back to the original implementation.

✨ New Features

  • Added layer norm operations support for ggml-webgpu.

🐛 Bug Fixes

  • Stabilized floating point computation for layer norm ops using Kahan summation (though later reverted).
  • Removed non-contiguous strides in shader (norm) implementation.

Affected Symbols