DeepSeek decodes faster on a single NVIDIA GPU and reaches its full 1M context, the chat stays steady while it writes
long code, and AMD and multi-GPU setups get new opt-in options.
- Faster DeepSeek decode on one NVIDIA GPU (#476, @RubenTorney): when the PC's RAM holds the experts, the GPU now
gets the experts it needs much more efficiently and keeps more of them close at hand. Up to about 2x decode reported
on an RTX 3090, +14% on a V100 here and +35% with less VRAM free; prompts read at the same speed. It is on for Linux
with one NVIDIA GPU; Windows, AMD and multi-GPU setups keep the previous path, and the start's log says which runs. - DeepSeek 512K and 1M contexts (#442, @needmorevram): the setup and the dashboard now offer the model's full
context. A longer context takes VRAM from the experts, and a prompt that long takes a while to read. - A steadier chat while it streams (#478, @needmorevram): long code answers no longer flash or jump, scrolling up
stops the chat following the answer, and the sidebar shows the model's logo and name. - GLM on AMD, opt-in (#454, #455, #456, #457, @pepuscz): the expert helper now works beside a layer split, Maya-M's
experts can run on a tuned RDNA kernel, the profiler shows attention in more detail, and alternative decode
attention kernels give the same results to the bit. With these settings off nothing changes (docs/SETTINGS.md). - DeepSeek on several GPUs, opt-in (#412, #415, @pepuscz): the DSpark drafter can be spread over a tensor-parallel
group to free room on the first GPU, and a helper GPU can take the experts VRAM cannot hold (Linux). Off by default.
Update: the dashboard's About > Update, or git pull, then ./setup.sh (Windows: START-MAYA.bat) - it recompiles only what changed; the model is not downloaded again.