first commit

2025-07-28 19:44:18 +01:00
parent d9a961a568
commit c76c69d8d8
3 changed files with 182 additions and 0 deletions
@@ -0,0 +1,35 @@
+FROM fedora:rawhide
+
+# Install build dependencies and tools
+RUN dnf install -y \
+       make gcc cmake lld clang clang-devel compiler-rt libcurl-devel \
+       rocminfo radeontop 'rocm-*' 'rocblas-*' 'hipblas' 'hipblas-*' \
+       git vim \
+    && dnf clean all
+
+# Set up working directory
+WORKDIR /opt/llama.cpp
+
+# Clone llama.cpp repository (with submodules)
+RUN git clone --recursive https://github.com/ggerganov/llama.cpp.git .
+
+# Build llama.cpp with HIP support
+RUN git clean -xdf \
+    && git pull \
+    && git submodule update --recursive \
+    && \
+    # Configure and compile with HIP toolchain
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
+      cmake -S . -B build \
+            -DGGML_HIP=ON \
+            -DAMDGPU_TARGETS=gfx1151 \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DLLAMA_HIP_UMA=ON \
+    && cmake --build build --config Release -- -j$(nproc) \
+    && cmake --install build --config Release
+
+RUN find /opt/llama.cpp/build -type f -name 'lib*.so*' -exec cp {} /usr/lib64/ \; \
+ && ldconfig
+
+# Default to interactive shell
+CMD ["/bin/bash"]
@@ -0,0 +1,30 @@
+FROM fedora:rawhide
+
+# Install build tools, Vulkan headers/loader, and glslc
+RUN dnf install -y \
+      git vim \
+      make gcc cmake ninja-build lld clang clang-devel compiler-rt libcurl-devel \
+      vulkan-loader-devel mesa-vulkan-drivers \
+      glslc \
+    && dnf clean all
+
+WORKDIR /opt/llama.cpp
+
+# Clone llama.cpp
+RUN git clone --recursive https://github.com/ggerganov/llama.cpp.git .
+
+# Build with Vulkan support
+RUN git clean -xdf \
+ && git pull \
+ && git submodule update --recursive \
+ && cmake -S . -B build -G Ninja \
+      -DGGML_VULKAN=ON \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_INSTALL_PREFIX=/usr \
+      -DLLAMA_BUILD_TESTS=OFF \
+      -DLLAMA_BUILD_EXAMPLES=ON \
+      -DLLAMA_BUILD_SERVER=ON \
+ && cmake --build build --config Release \
+ && cmake --install build --config Release
+
+CMD ["/bin/bash"]
@@ -1 +1,118 @@
 # amd-strix-halo-toolboxes
+
+This repository provides two Fedora Rawhide–based Dockerfiles for building isolated dev containers with AMD’s new **Strix Halo** GPUs (gfx1151):
+
+- **Docker.rocm**: Builds `llama.cpp` with ROCm (HIP) support targeting gfx1151
+- **Docker.vulkan**: Builds `llama.cpp` with Vulkan compute support
+
+Both images load the latest ROCm/Vulkan libraries from Fedora Rawhide to ensure compatibility with Strix Halo.
+
+## Repository Structure
+
+```
+.
+├── Docker.rocm        # HIP-based build for ROCm (gfx1151)
+├── Docker.vulkan      # Vulkan-based build
+└── README.md          # This documentation
+```
+
+## Prerequisites
+
+- **Podman** (or Docker aliased to Podman) installed on a Fedora Rawhide (or compatible) host
+- **Fedora Toolbox** (for creating dev containers)
+- AMD GPU drivers (ROCm/Vulkan) installed on the host
+
+
+## 1. Building the Images
+
+### ROCm (HIP) Image
+
+```bash
+podman build -t llama-rocm -f Docker.rocm .
+```
+
+### Vulkan Image
+
+```bash
+podman build -t llama-vulkan -f Docker.vulkan .
+```
+
+
+## 2. Creating Toolbox Containers
+
+Toolbox will automatically mount your home directory, map your UID:GID, enable X11, and use the host network. You only need to pass through the GPU devices and relax seccomp.
+
+### ROCm Toolbox
+
+```bash
+toolbox create llama-rocm \
+  --image localhost/llama-rocm:latest \
+  -- \
+    --device /dev/kfd \
+    --device /dev/dri \
+    --group-add video \
+    --security-opt seccomp=unconfined
+```
+
+### Vulkan Toolbox
+
+```bash
+toolbox create llama-vulkan \
+  --image localhost/llama-vulkan:latest \
+  -- \
+    --device /dev/dri \
+    --group-add video \
+    --security-opt seccomp=unconfined
+```
+
+
+## 3. Entering and Testing
+
+After creation, enter each container and verify that your GPU and libraries are accessible.
+
+### ROCm (HIP) Test
+
+```bash
+toolbox enter llama-rocm
+# inside container:
+llama-cli --list-devices
+```
+
+You should see your Strix Halo gfx1151 device listed.
+
+### Vulkan Test
+
+```bash
+toolbox enter llama-vulkan
+# inside container:
+vulkaninfo | head -n 10
+llama-cli --help  
+```
+
+If `vulkaninfo` reports your GPU and `llama-cli` runs without errors, the Vulkan build is working.
+
+
+## Host Configuration
+
+The following host details and kernel settings ensure optimal performance and unified memory access for Strix Halo:
+
+- **Machine**: HP Z2 Mini G1a  
+- **Memory**: 128 GB RAM, with 512 MB allocated to the GPU in BIOS  
+- **Host OS**: Fedora 42, kernel 6.15.6-200.fc42.x86_64  
+- **Kernel boot parameters** (in `/etc/default/grub`):  
+  ```text
+  amd_iommu=off amdgpu.gttsize=131072 ttm.pages_limit=335544321
+  ```  
+  - `amd_iommu=off` disables IOMMU for lower latency and avoids address translation overhead.  
+  - `amdgpu.gttsize=131072` sets the GPU GTT (Graphics Translation Table) size, enabling a unified memory window so the GPU can directly access up to 128 GB of system RAM.  
+  - `ttm.pages_limit=335544321` raises the TTM (Translation Table Maps) page limit to allow larger pinned allocations.  
+- **Apply parameters**: after editing `/etc/default/grub`, run  
+  ```bash
+  sudo grub2-mkconfig -o /boot/grub2/grub.cfg
+  ```  
+
+## Notes  
+
+Both images pull Fedora Rawhide packages for the newest ROCm/Vulkan support.  
+
+