Seeking guidance for build and inference of b1_58-2B-4T with TL2 on Windows

[Jimmy-sheep (Jimmy-sheep)]

I am currently trying to build and run the b1_58-2B-4T model using the tl2 kernel on Windows, but I have encountered persistent build and runtime issues. Although I managed to resolve tons compilation errors with the help of AI-assisted patches, the model inference finally outputs only question marks (???), leaving me uncertain about the validity of those modifications.

As a beginner in this field looking to learn by getting this project running locally, my current goal is simply to establish a working baseline. I suspect the issue might stem from recent changes or compatibility differences in the 3rdparty/llama.cpp submodule.

Has anyone successfully compiled and run b1_58-2B-4T with the tl2 kernel on Windows(not with WSL)? If so, could someone share a verified setup or workflow that works in this environment I would greatly appreciate.

Environment:
OS: Windows 11 Home (Build 10.0.26200)
Compiler / Toolchain: Visual Studio 2026 + Clang 22.1.3 (x86_64-pc-windows-msvc)
CMake Version: 4.3.1-msvc1
Python Version: 3.10.11 (in venv)