b11108 b11109
medium risk
ggml-org/llama.cppSep 22, 2026 → Sep 22, 20260 days between releases
What's New
Adds backend safety checks for MUL_MAT_ID operations to prevent failures
Breaking Changes
- MUL_MAT_ID operations with F32 source precision may now fail on Metal, CUDA, and Vulkan backends.
New Features
- Added precision checks to gate MUL_MAT_ID rescaling operations.
Improvements
- Metal backend logic updated to handle MUL_MAT_ID operations safely.
Bug Fixes
- Fixed backend failures by rejecting unsupported MUL_MAT_ID operations.
Upgrade Notes
If your application uses MUL_MAT_ID operations with F32 source precision on Metal, CUDA, or Vulkan, ensure the operation is supported or it will be rejected. This change prevents crashes in these specific backends.
The Two Releases
b11108
Pre-release
b11108
Sep 22, 2026· 30 days ago
Optimizes IQ1_M quantization performance by building prefix sums once per block instead of repeatedly.
Full release notesb11109
Pre-release
b11109
Sep 22, 2026· 30 days ago
Restricts MUL_MAT_ID operations in Metal, CUDA, and Vulkan backends based on source precision to prevent backend failures.
Full release notes