trunk/376702f84049e0d74551e656432395979fd14da5: [ROCm] Round gfx942 TF32 GEMM inputs to nearest (#200788) (#200788)
- PyTorch: 2573 events in the last 90 days
- PyTorch: 2538th Release in the last 90 days
- Previous: earlier the same day · ciflow/trunk/201491: [INITIAL] Update
What happened
Summary: This diff adds opt-in round-to-nearest-even preprocessing for native FP32 GEMMs on AMD gfx942 so MI300 TF32 input conversion tracks NVIDIA TF32 semantics. The path is private and controlled by TORCH_GFX942_ENABLE_RN=1 ; unset or 0 preserves current behavior. Context: T16 NVIDIA-to-AMD parity validation found a significant dpa_product_cvr_model NE increase only on MI300 with TF32 enabled. MI300/gfx942 native…
Summary assembled by rule from the sources below