Change8

b7632

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

Summary

This release focuses on bug fixes, primarily addressing issues within the Vulkan backend related to quantization overflow and matrix multiplication failures across different hardware configurations.

🐛 Bug Fixes

  • Handled quantize_q8_1 overflowing the max workgroup count in Vulkan backend.
  • Fixed small tile size matrix multiplication on Lavapipe.
  • Fixed mul_mat_id failures.

🔧 Affected Symbols

vulkanquantize_q8_1matmulmul_mat_id