The ggml-org team released llama.cpp b12367 with a CUDA change: it adds SM103 SASS, or GPU machine code, to non-native builds. That can let those builds run code prepared for SM103 hardware without being compiled specifically on that target.
The release also includes downloadable builds for macOS, iOS and several Linux systems. The notes don’t identify which GPUs use SM103 or give performance results.
Why it matters
People running llama.cpp on SM103-targeted Nvidia hardware may be able to use non-native CUDA builds, but the release notes don’t specify affected GPU models or measure any speed change.
