← Back to events
ActiveAIRelease9.0

viable/strict/1789252151: [inductor] Gate CUDA TMA tests on CUDA devices (#196636)

What happened

has_triton_tma_device() reports tensor-descriptor support when any installed Triton backend supports it, including CPU and XPU. A ROCm build that also bundles the Triton CPU backend can therefore return true even though its AMD GPU cannot run CUDA TMA templates. Add a CUDA-specific predicate requiring a non-HIP CUDA device with compute capability 9.0 or newer, and use it for the CUDA-only TMA tests. Keep the existin…

Summary assembled by rule from the sources below

Why it's spreading

Sources