Paul Durkin
Solo developer building AI training and inference workflows for consumer-grade AMD hardware — specifically the Ryzen AI MAX+ 395 / Radeon 8060S Strix Halo platform with 128 GB of unified memory.
Most of my current work is open-source. The big-picture stuff:
- strix-halo-llm-finetune-guide — production-tested guide for getting Strix Halo working end-to-end for fine-tuning 27B+ class LLMs. ROCm 7 nightly setup, mainline kernel patches, llama.cpp HIP + Vulkan builds, the actual training script that works, troubleshooting for every wall I hit
- ROCm vs Vulkan on Strix Halo — long-form benchmark writeup; backend choice depends on quantization (Vulkan wins Q4 decode by 22%, ROCm wins BF16 decode by 117%)
- NorthstarAurora — my org; benchmark datasets and (eventually) reproducible fine-tunes land there
Currently working on:
- Fine-tuning Qwen3.5-27B (hybrid GatedDeltaNet architecture) with bf16 LoRA on a single Strix Halo box
- A/B benchmarking ROCm vs Vulkan across precision levels on gfx1151
- Contributing fixes upstream to Unsloth's Strix Halo support (see unsloth#5301, #5303, #5517)
Find me:
Most of my own model and dataset repos here are private — the public surface is the parts that are useful to the broader Strix Halo / open-LLM community.
