GUIDES
AMD consumer hardware can run serious AI workloads. Getting there means fighting undocumented failures nobody has written up yet. These are the write-ups — the exact errors, the exact fixes, and the versions they were true for.
The full path from a throttled APU to working LTX-2 generation — ROCm install, performance mode, the nightly PyTorch build, the tensor patch, and the env vars that stop the hangs.
COMFYUInumpy has no bfloat16, so the cast produces NaN and your render comes out black. How to tell which of the three failure modes you have, and the FP32 VAE decode fix.
PYTORCHPorting a CUDA-assuming ML research harness to a consumer APU: the FA3 to SDPA swap, the dtype fix that unblocks fullgraph compile, and the two mandatory ROCm env vars.
VIDEO GENWhy unified memory removes the CPU-offload problem entirely, and why patching the transformer blocks works where patching forward crashes on LTXAV.
COMFYUISubprocess-based lip sync nodes leave temp-file litter and orphaned processes. An in-process approach that works on ROCm and CUDA alike.
AI PRACTICEFive markdown files give an assistant identity across sessions. What goes in each, why it changes the answers you get, and how to set it up in about five minutes.
Every guide links back to working code on GitHub. Versions and commands are quoted exactly as tested — where something was true only for a specific build, the build is named.