b10690
📦 llama-cppView on GitHub →
🐛 1 fixes🔧 2 symbols
Summary
This release addresses a memory-related crash during context shifts for unquantized K caches. It ensures the Hadamard matrix is copied to the k_rot tensor only when a buffer is assigned, preventing potential issues.
🐛 Bug Fixes
- Prevented crashes during context shift of unquantized K cache by ensuring the Hadamard matrix is copied to the k_rot tensor only if it has a buffer assigned.