The llama.cpp project released version b11630 on Oct 15, 2026. It fixes a layout bug in the conv_1d operation when processing batched inputs. The previous code stored results as [OL, N, OC] but told later steps they were [OL, OC, N], causing channels to mix with batches. The fix reshapes to the correct layout and permutes dimensions, so all seven test geometries now pass. The change matters for downstream audio tools that run several synthesis lanes in the same graph.
Why it matters
Audio developers can now safely batch multiple vocoder lanes in one llama.cpp graph without output corruption.
