A fix for DeepSeek conversations past 32,768 tokens on AMD GPUs and NVIDIA RTX 20 cards.
- DeepSeek past 32K tokens on AMD (Strix Halo, Radeon) and RTX 20: since v1.1.22, once a conversation passed
32,768 tokens the engine stopped (ds4 gguf mv: invalid argument). The long-context top-k now sizes its buffer from
each GPU's shared memory - the same picks, the same speed. NVIDIA V100, RTX 30 and newer were not affected. - Maya-XS is no longer labelled experimental in the setup.