Technical write-up of getting TRELLIS.2 running on a Ryzen AI MAX+ 395
iGPU under ROCm, cutting generation from ~22 min to ~40 s, and wiring it
into a queue-backed workflow (vision LLM → background removal →
generation → auto-rig → publish).
Includes the two gfx1151 segfaults and their fixes, the AOTriton env-var
speedup, the decimation_target/UV-unwrap profiling that produced the
sub-2-minute draft tier, and five UI screenshots.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gabriel Vidal
693fd240a4
