vLLM has released version 0.21.0, featuring 367 commits from 202 contributors. This update deprecates support for Transformers v4, requiring users to migrate to v5, and introduces a C++20 build requirement for PyTorch compatibility. Key enhancements include the integration of KV offloading with the Hybrid Memory Allocator and speculative decoding improvements. The release also adds support for new model architectures and NVIDIA Blackwell GPUs, aiming to boost performance and expand capabilities.
Read originalLlama.cpp's latest update introduces speculative decoding support for GLM-5.2, enhancing its capabilities with NextN/MTP features. This addition allows for more efficient tensor loading and context management, particularly benefiting models using the GLM_DSA architecture. The update also includes options for exporting models with or without the MTP feature, providing flexibility for developers. This release marks a step forward in optimizing model performance and adaptability, especially for those leveraging the GLM-5.2 framework.
The b10178 release of llama.cpp enhances its server capabilities by adding trace logging for slot similarity checking, offering developers detailed insights into prompt cache slot selection processes. This update includes specifics on skip reasons and similarity calculations, which can aid in performance optimization. While no new model architectures are introduced, the release continues to support a wide array of platforms, such as macOS with KleidiAI, Ubuntu with ROCm 7.2, and Windows with CUDA 12 and 13. This makes llama.cpp a more versatile tool for developers working on different systems, reinforcing its position as a comprehensive inference runtime.