AMD Ryzen AI Halo: LM Studio setup and local AI demos

This AMD-sponsored review examines the Linux version of the Ryzen AI Halo, which the presenter describes as having 128 GB of unified LPDDR5X memory. The CPU and GPU share that pool, with settings to adjust the GPU allocation. Windows is also an option, but the demonstrations use Linux with drivers and AI software already configured.

The LM Studio setup requires installing its ROCm llama.cpp runtime; the presenter says it is not selected by default. His local LLM tests report above 50 tokens per second for Qwen 3.6 35B with a drafter and around 30 for GPT-OSS-120B. These are results from the demonstrated workloads. He favors the larger memory capacity for loading big models, while noting that his earlier discrete GPU workstation can be faster with smaller ones.

ComfyUI provides the image and video workflow. A script varies prompts and seeds, then sends generated images to LTX for animation. The presenter reports roughly 19 to 20 seconds per image in his Qwen image workflow. Hermes Agent uses LM Studio for function calling and skills, with multiple models loaded for switching between tasks.

The final demo follows an Unsloth playbook for 50 training steps on a Gemma 4B model and merges LoRA weights into a 16-bit model. Larger fine-tunes require careful memory allocation and may need smaller batches with gradient accumulation. The review presents capacity and preconfigured software as the machine's practical strengths, with speed depending on the workload.