Guides & Fixes
Installation notes, configurations, and fixes from things I have worked through myself.
Getting started
Building llama.cpp and setting up each test machine.
- Run llama.cpp on AMD Strix Halo: ROCm and Vulkan Setup
How I set up llama.cpp on Strix Halo with ROCm and Vulkan, including build commands and shared-memory settings.
- Build llama.cpp with CUDA for NVIDIA Blackwell
Build llama.cpp for Blackwell GPUs on Linux, with the CUDA settings and commands I used.
- Install ROCm 7.14 on AMD Strix Halo / Ryzen AI Max+
Install ROCm 7.14 for Strix Halo's Radeon 8060S and verify the setup with rocminfo and llama.cpp.
- Set Up Vulkan Compute on AMD Strix Halo / Radeon 8060S
Set up Mesa RADV and build llama.cpp's Vulkan backend for Strix Halo's Radeon 8060S.
Troubleshooting and hardware
Fixes for problems I hit, and checks for the hardware underneath.
- Fix NVIDIA Multi-GPU P2P Corruption on Intel Linux
Investigate corrupt multi-GPU llama.cpp output using kernel logs and tests for IOMMU-related problems.
- Check Your GPU's Real PCIe Bandwidth on Linux
Check your GPU's negotiated PCIe speed and link width, understand what the reported values mean, and measure real transfer bandwidth.
- RTX PRO 5000 Blackwell eGPU on Linux with Sonnet 850T5
Setup and troubleshooting notes for an RTX PRO 5000 in a Sonnet 850T5 Thunderbolt 5 enclosure on Linux.
Model notes
Options and commands
For command-line flags, environment variables, and other lookup material, see Reference.