llama.cpp with ROCm

Warning

This is a technical guide and assumes a certain level of technical knowledge. If there are confusing parts or you run into issues, I recommend using a strong LLM with research/grounding and reasoning abilities (eg Claude Sonnet 4) to assist.

While Vulkan can sometimes have faster tg speeds, it can run into "GGGG" issues in many situations, and if you want the fastest pp speeds, you probably will want to try the ROCm backend.

As of August 2005, the generally fastest/most stable llama.cpp ROCm combination:

  • build llama.cpp with rocWMMA: -DGGML_HIP_ROCWMMA_FATTN=ON
  • run llama.cpp with env to use hipBLASlt: ROCBLAS_USE_HIPBLASLT=1

There are still some GPU hangs, see:

If you are looking for pre-built llama.cpp ROCm binaries, first check out:

Building llama.cpp with ROCm

If you want or need to build it yourself, you can basically just follow the llama.cpp build guide:

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp

# build w/o rocWMMA
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151 -DCMAKE_BUILD_TYPE=Release && cmake --build build --config Release -- -j$(nproc)

# really, you want to build w/ rocWMMA
cmake -B build -S . -DGGML_HIP=ON -DAMDGPU_TARGETS="gfx1151" -DGGML_HIP_ROCWMMA_FATTN=ON && time cmake --build build --config Release -j$(nproc)

# after about 2 minutes you should have a freshly baked llama.cpp in build/bin:
build/bin/llama-bench --mmap 0 -fa 1 -m /models/gguf/llama-2-7b.Q4_K_M.gguf

Of course, to build, you need some dependencies sorted.

First, you should run the latest Linux (6.16+) and linux-firmware (git).

ROCm

You'll need ROCm installed first before you can build. For best performance you'll want to use the latest ROCm/TheRock nightlies. See: Guides/AI-Capabilities#rocm

To build, you may need to make sure your environment variables are properly set. If so, take a look at https://github.com/lhl/strix-halo-testing/blob/main/rocm-therock-env.sh for an example of what this might look like. Change ROCM_PATH to whatever your ROCm path is.

rocWMMA

Your ROCm probably has the rocWMMA libraries installed already. If not, you'll want them in your rocm folder. This is relatively straightforward (we only need the library installed, but you can refer to https://github.com/lhl/strix-halo-testing/blob/main/arch-torch/02-build-rocwwma.sh for building this.

If you are using a TheRock nightly build of ROCm, you may get some errors compiling. In that case, take a look at https://github.com/lhl/strix-halo-testing/blob/main/llm-bench/apply-rocwmma-fix.sh to apply the fixes necessary for a compile.

0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9