The test added in #2899 lowers with an explicit `tvm.target.Target("cuda")`, which
reaches `tilelang/cuda/transform/__init__.py`. In a `USE_CUDA=OFF` build only the
"ALWAYS" CUDA sources compile, so `AnnotateDeviceBoundTmaCopies` is never registered
and the test fails on ROCm:
AttributeError: module 'tilelang.cuda._ffi_api' has no attribute
'AnnotateDeviceBoundTmaCopies'
Same pattern as #2863. `requires_cuda` is the weakest sufficient gate: the test is
compile-only (`enable_device_compile=False`), so it needs no particular GPU
generation and keeps running on the NVIDIA leg.
Verified in CI: SKIPPED on the gfx942 leg, PASSED on CUDA-auto.
Yuxi Chi