The llama.cpp release adds topk_mask support to its OpenCL backend and offers downloadable builds for several platforms.
NoteBy OossaPublished by Oossa: 1 min read
The ggml-org team has released llama.cpp b12052, a new version of the software used to run language models. Its highlighted change is “add topk_mask” for OpenCL, a framework for running computation across different hardware. The release page also lists builds for macOS, iOS and several Linux systems.
People running llama.cpp with OpenCL may be able to use top-k masking, though the release notes do not say which models or tasks benefit.
Models
The 753-billion-parameter model’s weights are available under the MIT License. Deep Principle reports gains on several science tests, but has not released its full training data or recipe.
Models
The Google–Meta project lets developers run PyTorch workloads on Google’s chips with limited code changes, alongside tools such as vLLM and Hugging Face.
Models
From December, Plus, Pro and Business users will draw from one token budget for ChatGPT and Codex. Instant-model chats are excluded.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
