Skip to content

Request: AVX512-off build — Arrow Lake CPUID reports AVX512=1 but hardware doesn't support it #10

@Go-Memphis

Description

@Go-Memphis

System:

  • OS: Windows 11
  • CPU: Intel Core Ultra 9 275HX (Arrow Lake-HX)
  • GPU: RTX 5080 Laptop (sm_120, 16GB VRAM)
  • CUDA Toolkit: 13.0
  • Python: 3.12.12

Problem:
Using llama_cpp_python-0.3.20+cuda13.0.sm100.sm120.blackwell-py3-none-win_amd64.whl

Crashes with 0xC000001D (Illegal Instruction) at import time.

CPU-Z confirms Arrow Lake has NO AVX512 in its instruction set.
However, llama.cpp system_info incorrectly reports AVX512=1 via CPUID.
This causes llama.cpp to enable AVX512 code paths that the hardware
cannot execute.

bcdedit /set xsavedisable 1 does not help — it also disables AVX/AVX2,
making the crash worse.

Request:
A build with -DGGML_AVX512=off for the same configuration:

  • Blackwell sm_120
  • CUDA 13.0
  • Python 3.12
  • Windows

This is also reported in issue #4 by other users with similar hardware.

Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions