Change8

b10702

📦 llama-cppView on GitHub →
🐛 3 fixes🔧 4 symbols

Summary

Optimized HIP's Q2_0 dot-product path for gfx1201 and improved MMQ unpack with native permutations. Restored MMQ tile index handling.

🐛 Bug Fixes

  • Optimized HIP's Q2_0 dot-product path for gfx1201 by using native amdgcn perm for vec_dot_q2_0_q8_1.
  • Optimized HIP Q2_0 MMQ unpack with native perm.
  • Restored MMQ tile index handling.

Affected Symbols