AMD ROCm 7.0 Install for Mi50 32GB | Ubuntu 24.04 LTS скачать в хорошем качестве

AMD ROCm 7.0 Install for Mi50 32GB | Ubuntu 24.04 LTS

This graphics card is getting popular for AI as a cheap option for 32GB of VRAM but AMD doesn't "officially" support anymore i.e. they stopped shipping the tensor files. In this guide, I show you how to manually add them back in, which makes 7.0.2 fully functional with llama.cpp. This guide can be adapted for 6.4 if more stability is needed when working with PyTorch or vllm. Most performance improvements were already present in 6.4 (roughly 20-30% over 6.3), so 7.0 serves to offer more compatibility together with the latest AMD cards :) Content: 00:00 - ROCm quick install 01:04 - 6.4 Rocblas download from the AUR 01:15 - Extract the zip file 01:23 - Copy the tensor files 02:20 - Reboot 02:31 - Check if it worked Guide: 1. Run the commands from the ROCm quick install: https://rocm.docs.amd.com/projects/install... 2. Before rebooting to complete the install, download the 6.4 rocblas from the AUR: https://archlinux.org/packages/extra/x86_6... 3. Extract it 4. Copy all files that contain the filename "gfx906" in rocblas-6.4.3-3-x86_64.pkg/opt/rocm/lib/rocblas/library to /opt/rocm/lib/rocblas/library 5. Reboot, enrolling MOK if needed 6. Check by running sudo update-alternatives --display rocm Now you can build llama.cpp with ROCm + flash attention (adjust j value according to number of threads): HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \ cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx906 -DGGML_HIP_ROCWMMA_FATTN=ON -DCMAKE_BUILD_TYPE=Release \ && cmake --build build --config Release -- -j 16 Note: Vulkan also works, but in my findings prompt processing seems to be better on ROCm.