ExLlamaSharp 1.3.1
Windows LLM server (EXL3 / ExLlamaV3) with OpenAI-compatible API and Blazor Admin.
Fixes
- Admin session auth (stale/revoked cookies no longer leave Chat on 401)
- Qwen3 Chat hang (thinking block consumed all tokens with empty visible output)
- Ollama
options.repeat_penalty: 1.0no longer maps tofrequency_penalty=1(which stalled the sampler) - CUDA device sanitization for single-GPU desktops
- Persisted ASP.NET DataProtection keys under ProgramData
Install
Run ExLlamaSharp-Setup-win-x64.exe as Administrator (desktop/Tray mode recommended).
PyTorch is downloaded at install time; ExLlamaV3 CUDA wheel is bundled.