b8392

The maintainers of the massively popular llama.cpp project, a cornerstone of the open-source AI inference ecosystem, have pushed a critical fix. Commit b8392 resolves a bug in the KLEIDIAI backend where the `supports_op()` function was incorrectly rejecting matrix multiplication (MUL_MAT) operations with three-dimensional (batched) inputs. This rejection occurred despite the underlying `compute_forward_qx()` implementation being fully capable of handling such inputs via a loop. The bug specifically caused models using Q4_0 or Q8_0 quantization—common formats for efficient deployment—to crash during the graph scheduling phase when processing sequences longer than a single batch (n_seq_max > 1).

The root cause was a mismatch between weight loading and runtime execution. During loading with 2D test inputs, weights were correctly placed in KLEIDIAI buffers. However, the runtime's use of 3D inputs triggered a failure because the operation was incorrectly flagged as unsupported. The fix adjusts the logic to correctly return `true` for 3D inputs and relaxes a buffer check to allow `supports_op()` to be called during the weight loading phase when a source buffer might be NULL. This stability patch is essential for the project's broad compatibility, which spans macOS (Apple Silicon/Intel), Linux (CPU, Vulkan, ROCm), Windows (CPU, CUDA, Vulkan), and specialized platforms like openEuler.