ROCm 7.2 working on AMD Vega 8 (Ryzen 5700G); could also work on Vega 56/64
A few months ago, when I tried to squeeze the maximum from my AMD Vega 8 APU — Ryzen 5700G, I was not able to find the latest custom-baked ROCm for LLM anywhere, so I decided to build one for Linux — here is ROCm 7.2 working on AMD Vega 8: https://github.com/daimonionnn/amd-vega-rocm-vulkan-llm-toolkit
This can also work on Vega56/64 (Vega10) since it is same architecture. Maybe just some minor changes in config are needed.
Tested on Qwen35B and smaller Gemma4 models. It was better in prefill than Vulkan, but since then Vulkan has improved even in prefill. I did not have time to extensively test it on more LLM models, and the results are a mixture of older and newer ROCm, Vulkan, different settings, and different Ubuntu versions/Docker images. My plan was to test and optimize it on Vega 56/64 (Vega 10), but my only Vega 56 died some time ago — I shorted it badly when I started the PC with the graphics card not fully seated in the PCIe slot. I also recently upgraded to a new MOBO, CPU, and 2x Radeon 9700 AI Pro (Asus ProArt Z890 and Intel Core Ultra 5 250K Plus) and I'm not planning to develop/optimize this anymore, but any pull requests or forks for Vega 56/64 + PyTorch/ComfyUI support, optimization, benchmarks are welcome. See https://github.com/daimonionnn/amd-vega-rocm-vulkan-llm-toolkit/blob/main/docs/ARCHITECTURE.md for details.