← Back to events
ActiveAIRelease

viable/strict/1788216615: [inductor] Enable BF16 index_add lowering on SM90+ and ROCm (#195192)

What happened

The Inductor index_add decomposition retained a stale OSS-only BF16 fallback after Triton added BF16 atomic_add support. This enables the decomposition on ROCm and on NVIDIA SM90+, allowing index_add to participate in fusion instead of becoming an ATen fallback kernel. Triton support was added in triton-lang/triton#6519 and was tested upstream on MI300 as well as NVIDIA GPUs. NVIDIA remains gated to SM90+ here becau…

Summary assembled by rule from the sources below

Why it's spreading

Sources