Local Ai
b9827
Release b9827 of llama.cpp, released on June 27, 2026, added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy for improved CUDA tensor copying performance. When tensors are not fully contiguous but each
Release b9827 of llama.cpp, released on June 27, 2026, added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy for improved CUDA tensor copying performance. When tensors are not fully contiguous but each row is contiguous, it uses cudaMemcpy2DAsync instead of the slow element-wise scalar copy kernel.
Source: llama.cpp Releases | 2026-06-27