Hey,

Your recent Windows cu128 builds seem to be broken. I get a STATUS_ILLEGAL_INSTRUCTION error when initializing the model context.

Error:

Traceback (most recent call last):
  File "F:\IA\toolcalltest\vl\tool_vl.py", line 175, in <module>
    llm = Llama(**llm_params)
  File "F:\IA\toolcalltest\venv\lib\site-packages\llama_cpp\llama.py", line 455, in __init__
    internals.LlamaContext(
  File "F:\IA\toolcalltest\venv\lib\site-packages\llama_cpp\_internals.py", line 360, in __init__
    ctx = llama_cpp.llama_init_from_model(self.model.model, self.params)
OSError: [WinError -1073741795] Windows Error 0xc000001d

Working build:
llama_cpp_python-0.3.18-cp310-cp310-win_amd64.whl (cu128-Basic-win-20251223) - Works flawlessly.
Broken builds:
llama_cpp_python-0.3.22-cp310-cp310-win_amd64.whl (cu128-Basic-win-20260118) - Crashes with the error above.

Environment:

OS: Windows 10/11
Python: 3.10
GPU: NVIDIA (CUDA 12.8)
CPU: AMD Ryzen (Family 23, Model 113) — does not support AVX-512
The 0xc000001d error code (STATUS_ILLEGAL_INSTRUCTION) suggests the recent builds may have been compiled with CPU instructions (e.g. AVX-512) that are not available on all processors, whereas the older "Basic" builds did not have this issue.