Release b9274 (May 21) fixes a critical VRAM leak in llama.cpp's Multi-Token Prediction (MTP) stack where speculative decoding resources were not freed during sleep/idle cycles, causing servers to crash after repeated cycles. This is especially important for local agent workflows running 24/7 on consumer GPUs — the leak would silently accumulate until OOM. Users running Qwen3.6 MTP models were most affected.