b10181
📦 llama-cppView on GitHub →
🐛 1 fixes🔧 2 symbols
Summary
This release addresses a critical bug in ggml-cuda by disabling MMQ on devices with insufficient shared memory, preventing crashes and ensuring a fallback to the BLAS path. Various pre-compiled binaries for different platforms and hardware accelerators are also provided.
🐛 Bug Fixes
- Disabled MMQ on devices with less than 48 KiB shared memory to prevent crashes and ensure fallback to BLAS path. This affects prefill (batch > 1) on devices like Moore Threads MTT S70 with 28 KiB shared memory, which previously caused GGML_ABORT.