llama.cpp

ggml-cuda: fix ROCm multi-GPU illegal memory access in recurrent state restore

#21170

Open