ggml-org’s llama.cpp release b11977 fixes how its OpenVINO backend handles SCALE operations on sliced tensor views. Previously, a strided view could be treated as its full source tensor, producing an oversized result and potentially crashing tests.
The fix processes the view as a slice and rejects in-place SCALE on non-contiguous views, where copying the result back could write to the wrong elements. The release also lists builds for platforms including macOS, Linux, Android and Windows.
Why it matters
For developers using llama.cpp’s OpenVINO backend with sliced tensor inputs, this fixes a case that could crash tests or write results to the wrong elements.
