ggml-org/llama.cpp is a local LLM inference engine written in C/C++. Over the course of 2026-10-29 KST, 9 build tags went up from b11963 to b11982, and no separate commit batch was counted. The content centers on compute optimizations in the SYCL and Hexagon backends, followed by the addition of an OpenCL path for Adreno and a few bug fixes.
The most notable are the three SYCL items. b11963 applies a workgroup argmax (an operation that finds the maximum value within a row) to short TOP_K rows, introducing a computation method different from the existing path, and the release notes include a performance table comparing multiple row sizes and k values on Arc Pro B70 with oneAPI 2026.1. b11979 changed the dequantize-and-store step for rearranged-layout quantization into vector operations, raising throughput for the Q4_0 and K-quant families, and b11980 added a flash attention kernel (a method that bundles keys and values to compute the Attention Score) that runs directly on q4_0 and q8_0 key caches, handling decoding and small batches with 1~31 query tokens.
On the Hexagon backend, b11981 increased DDR bandwidth usage for transpose-and-concatenate (CONCAT), doubling the row group size and widening the buffer scratch space, and b11982 newly supports the UPSCALE operation, refining nearest-neighbor 2x and 4x interpolation and bilinear 4x upscaling. On the Adreno OpenCL side, b11971 added a tiled q4_k matrix multiplication path so that the language model output layer uses tiled operations when the batch is greater than 1. Elsewhere, b11966 switched the s390x continuous integration runner to Ubuntu 26.04, b11977 fixed a problem where OpenVINO SCALE multiplied the entire source tensor for strided view inputs and overflowed the result buffer, and b11978 corrected the CUDA graph reservation slack calculation test.
That said, most of the day's changes are fine-grained optimizations for specific backends. The performance comparison in b11963 also shows results slower than before under some conditions, the kernel in b11980 is used only for decoding and small batches and keeps f16 for the V.P operation, and the UPSCALE kernel in b11982 is not fully vectorized either. Since overall throughput figures are not in the release notes, it is still too early to judge how large the actual gain is.