llama.cpp issue #26208: ROCm backend allocates host RAM instead of VRAM on gfx1201

8x RADEON AI Pro 9700 XT (gfx1201) on Debian with ROCm 7.11-7.14 and 10.1.0: nvtop shows no VRAM allocation at all, compute runs but reads weights over PCIe from host RAM; a 30 GB model at 256k context consumed over 256 GB of RAM. Reproduces with the official ROCm tarball builds and official build artifacts, and inside the official ROCm container, and is not fixed by any GGML_HIP_* build flag tried. Another user with 2x R9700 on Manjaro reports normal behaviour. Closed as not-reproducible-by-others, so the mechanism is still unknown.