← Back to events
ActiveAIRelease

trunk/40e238cee98065a2526d8420affade6be07912d7: [XPU] Fix accuracy issue in addmm for bf16/f16 dtypes (#174864)

Photo: PyTorch Releases

What happened

Fixes intel/torch-xpu-ops#2837 Fixes precision loss in XPU addmm and baddbmm for bf16/f16 dtypes with non-trivial alpha/beta. The oneDNN 3-step post-op chain rounds intermediates to reduced precision at each stage. The fix pre-copies self into result and uses post_sum which accumulates in oneDNN's internal f32 accumulator, matching CPU/CUDA behavior. Pull Request resolved: #174864 Approved by: https://github.com/Eik…

Summary assembled by rule from the sources below

Why it's spreading

Sources