Success. I got the ROCM 7.1 build working in Arch Linux with Ryzen 8700GE (780M) chip

[litan1106]

First, thank you for the llama.cpp.
I want to share my build.

Building llama.cpp for ROCm on Arch Linux

This outlines the steps that worked for compiling the repository with HIP targeting the integrated Radeon 780M (gfx1100) inside the Ryzen Pro 7 8700GE on Arch Linux with ROCm 7.1 installed via the AUR.

1. Record the GPU target:

   hipconfig -l
   rocminfo | grep gfx

   → the messaging confirmed a gfx1100 RDNA3 GPU string.

2. Prepare the ROCm toolchain variables:

   HIP_PATH="$(hipconfig -R)"
   HIPCXX="$(hipconfig -l)/clang"

3. Configure CMake with HIP and the proper target:

   cmake -S . -B build \;
     -DGGML_HIP=ON -DGPU_TARGETS=gfx1100 \;
     -DGGML_HIP_ROCWMMA_FATTN=OFF -DCMAKE_BUILD_TYPE=Release \
     -DGGML_HIP=ON

   This ensures the HIP backend is built for the 780M without rocWMMA support (the headers were missing).

4. Build the project:

   cmake --build build --config Release -- -j$(nproc)

   The second invocation finished the remaining targets after the first run timed out at 120 s.

5. Source fixes in ggml/src/ggml-hip/CMakeLists.txt

   • Added find_package(hipblaslt REQUIRED) after find_package(rocblas REQUIRED)
   • Updated line from target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas) to target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas roc::hipblaslt) so the HIP extension symbols resolve during the link stage.

6. Runtime notes:

   • Use GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 when you need UMA for the iGPU.
   • Control visibility via HIP_VISIBLE_DEVICES or force the GPU string with HSA_OVERRIDE_GFX_VERSION=11.0.0 if ROCm misdetects the integrated GPU.

The binaries live under build/bin (e.g., llama-cli, llama-server).

7. Profit

HIP_VISIBLE_DEVICES=0 GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 \
    ./build/bin/llama-cli \
      --model models/Llama-3.2-8B-Instruct-Q4_K_M.gguf \
      --prompt "Hello again" \
      --gpu-layers 2 \
      --tensor-split 2,6 \
      --ctx-size 2048 \
      --flash-attn off \
      --threads 8
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon 780M Graphics, gfx1100 (0x1100), VMM: no, Wave Size: 32

Loading model...


▄▄ ▄▄
██ ██
██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
                                    ██    ██
                                    ▀▀    ▀▀

build      : b7728-afd8929b
model      : Llama-3.2-8B-Instruct-Q4_K_M.gguf
modalities : text

available commands:
  /exit or Ctrl+C     stop or exit
  /regen              regenerate the last response
  /clear              clear the chat history
  /read               add a text file


> Hello again

It's nice to chat with you again. Is there something on your mind that you'd like to discuss, or would you like some recommendations for topics?

[ Prompt: 29.7 t/s | Generation: 10.7 t/s ]