am17an
4 storiesam17an is a contributor to the llama.cpp project, mainly submitting code related to MoE expert caching and model conversion. Recent coverage includes a merged pull request (#31028) that merges tensors during GGUF conversion, which may require users to reconvert existing files; PR #29887 adding a GPU cache for MoE experts kept in host memory; and llama.cpp's merged MTP support for Qwen4Exp, with GGUF quantizations published on Hugging Face.
Related topics
llama.cpp pull request #30966 adds llama_moe_cache, which runs missed MoE experts on the CPU, plus a tuning script. It is seen as a significant speedup for low-VRAM users offloading MoE models.
Open-source project llama.cpp merged a pull request (#31028) by contributor am17an that merges tensors during conversion. The change may require users to reconvert existing GGUF model files.
Developer am17an submitted PR #29887 to llama.cpp adding a GPU cache for MoE experts kept in host memory, potentially delivering big speedups for Mixture-of-Experts models that don't fully fit in VRAM.
A pull request adding MTP support for Qwen4Exp was merged into llama.cpp after 17 hours of work, enabling MTP with Qwen Flash Next. Corresponding GGUF quantizations have been published on Hugging Face.
That is everything