Remove padding and multiple D2D copies for MTP by gaugarg-nv · Pull Request #24086 · ggml-org/llama.cpp

In three linesPull request on llama.cpp optimizing MTP (Multi-Token Prediction) by removing padding and redundant D2D copies. Performance improvement for multi-token inference.

Summary generated by Claude — human-verified