Change8

b7592

📦 llama-cppView on GitHub →
3 features🐛 3 fixes🔧 3 symbols

Summary

This release introduces the count_equal operator to the Metal backend and includes several optimizations and documentation cleanups for Apple Silicon performance.

✨ New Features

  • Added count_equal operator for Metal backend
  • Implemented multi-thread group (multi tg) and templating for Metal
  • Added memset to initialize destination to 0 in Metal operations

🐛 Bug Fixes

  • Removed BLAS support references from Metal documentation
  • Fixed shared memory type by changing shmem to i32
  • Cleaned up trailing whitespace and general Metal backend code

🔧 Affected Symbols

ggml_metalcount_equalshmem