How's your real-world AI (code generation) on the Framework Desktop?

[Squiggler]

Have you tried this? It looks like it uses ROCm:

hipEngine: Fast Native Qwen 3.6 Inference for RDNA3 (including Strix Halo)

Hey everyone, I wanted to share GitHub - shisa-ai/hipEngine · GitHub - a new open source (from-scratch) inference engine that is a pure AMD ROCm implementation (all hot-path code is HIP/C++, all AMD libraries, PyTorch is expressly not a dependency) for RDNA3 GPUs. It’s been mostly tuned on a spare gfx1100 (W7900, 7900 XTX) GPU, but I did an initial pass for gfx1151 (Strix Halo) support, and it ends up being faster than llama.cpp (HIP or Vulkan) basically across the board for Qwen 3.6 35B-A3B (b…