Record 039 · Runtimes · 2016
67%16.39.U
ROCm
AMD
“AMD's compute stack — the reason a 7900 XTX can run a model without NVIDIA.”
ROCm's rocky years still matter: llama.cpp, vLLM, and PyTorch-on-AMD are how non-NVIDIA local boxes exist. Strix Halo depends on this path.
Opened local GPU inference off the green team.
- Open GPU compute for AMD
- PyTorch/ROCm
- llama.cpp HIP
CUDA, if you brought an AMD card.
