ciflow/trunk/201190: ATen: Support output-shaped FP8 residual bias on ROCm (#201190)
- PyTorch: 2487 events in the last 90 days
- PyTorch: 2454th Release in the last 90 days
- Previous: earlier the same day · ciflow/torchtitan/201289: [UPDATE] Update
What happened
Summary: Support a BF16 output-shaped residual through the v1 torch._scaled_mm bias argument on ROCm with rowwise input scales. Add the residual in the GEMM epilogue before BF16 rounding. Detect matrix residuals by their two-dimensional output shape, including single-row outputs; element count alone must not route [1, N] residuals as vector bias. Preserve vector-bias compatibility and CUDA matrix-bias rejection. Rej…
Summary assembled by rule from the sources below