TL;DR: the recipe's image-build mods aren't actually public – I reconstructed them from the public kernels (with Claude) – and you have to build vLLM at the author's exact pinned ref or the real AWQ weights crash on load. Running now at ~9.4 tok/s on my own 4× GB10. Saw a link on X to CosmicRaisins' GLM-5.2 stack for 4× GB10: vLLM TP=4, MTP speculative decode, ported sparse-MLA Triton kernels (the Hopper-only _flashmla_C path doesn't exist on sm_121), and a data-free 15% expert prune so the AW...