llama.cpp

Reset schedule earlier to allow overlap with ggml graph computation on device

#6933

Merged