Vulkan is an on-ramp, not a consolation prize

This site is about ROCm at home. It is still honest to say: a lot of people generated their first R9700 tokens on Vulkan.

llama.cpp can compile a HIP backend and a Vulkan backend. They are different compilers pointing at the same memory. HIP is the path that shares DNA with PyTorch, vLLM, and hipcc. Vulkan is the path that shares DNA with the desktop driver you already needed to get a picture on the monitor.

When Vulkan is the right first backend

• HIP CMake cannot find a ROCm that rocminfo insists is fine
• You are on Windows, where a full ROCm userspace may not exist, and Adrenalin + Vulkan does
• An iGPU/dGPU split is easier to see in llama-cli --list-devices as RADV GFX1201 vs RADV RAPHAEL_…
• You want a weekend result before you commit to PyTorch wheels

Community threads for the R9700 spend a lot of time on RADV vs AMDVLK, visible devices, and whether memory clocks are actually at the top P-state. That work is real. It is also not a substitute for AMD’s published bandwidth or TOPS figures.

When you should still install ROCm

You will want HIP/ROCm when you:

• Install PyTorch or vLLM the way AMD documents
• Care about rocm-smi counters in the same way a CUDA user cares about nvidia-smi
• Debug library-level issues (rocBLAS kernels, missing gfx1201 tensiles) instead of shader paths
• Run more than “one GGUF in a terminal”

Ollama sits in the middle: it may use ROCm, Vulkan, or surprise-CPU depending on the build and OS. If it is slow, look at the backend log before you blame RDNA 4.

A clean way to keep both

On Linux, it is reasonable to have:

1. Distro or AMD graphics stack with a working RADV ICD
2. One ROCm userspace for HIP apps
3. Two llama.cpp build directories: build-hip and build-vk

Do not measure them against each other on different models, different -c, and different days, then declare a winner on the internet. Measure them to learn which layer you broke.

The local LLM tutorial includes both configure lines. Start with whichever configure script exits zero. Graduate to the other when you need it — not because a brand hierarchy told you to.