b8420
📦 llama-cppView on GitHub →
🐛 1 fixes🔧 2 symbols
Summary
This release addresses a critical bug in RotaryPositionEmbedding on CANN hardware related to in-place operations on non-contiguous f32 tensors. It also provides numerous updated pre-compiled binaries for diverse hardware and operating system configurations.
🐛 Bug Fixes
- Fixed an issue where RotaryPositionEmbedding on CANN failed when source and destination tensors shared the same non-contiguous buffer during in-place operations by implementing a fallback that uses contiguous temporary buffers for computation and copying results back.