llama.cpp Build b11330 Adds Multi-Token Prediction for Qwen4Exp
llama.cpp has released build b11330, introducing support for Multi-Token Prediction (MTP) in Qwen4Exp models. The update also refactors context buffer checks and cleans up recurrent memory structures. Multi-Token Prediction allows large language models to generate multiple tokens per forward pass, significantly accelerating inference speed on local hardware. Extending MTP support to experimental Qwen architecture variants ensures that open-source inference engines can run modern optimized model architectures efficiently. Pull request #29761 implements MTP for Qwen4Exp by replacing the `has_state` member check with a verification of non-empty `ctx_bufs`, while cleaning up code comments and recurrent memory handling. Pre-compiled binaries for build b11330 are provided across platforms including macOS, Linux (CUDA 12/13, ROCm, Vulkan, OpenVINO, SYCL, Snapdragon), Windows, and Android.
## BACKGROUND
Standard Large Language Models generate text autoregressively via next-token prediction, producing only one token at a time per forward pass. Multi-Token Prediction (MTP) is an inference technique that predicts multiple tokens simultaneously, reducing decoding latency. llama.cpp is a widely used open-source C/C++ inference framework designed for efficient execution of LLMs on consumer hardware across diverse CPU and GPU backends.