trunk/6a3a43fbf2d65204345bdb1789a02a4c0142d821: [Inductor] Match eager float-to-int narrowing in strict modes (#199788)
- PyTorch: 1956 events in the last 90 days
- PyTorch: 1930th Release in the last 90 days
- Previous: earlier the same day · trunk/6dfcc83c6095cc4d4f018e317fbbdee7aebaecc9
What happened
Floating-point casts to byte/char/short can disagree with CUDA eager for overflow, NaNs, and infinities. Match eager by converting through int64 for byte and int32 for char/short before narrowing. Apply this to Triton under numerics="strict_pointwise" and "strict". Add regression coverage across floating dtypes, integer inputs, numerics modes, and precision-upcast settings. Remove nine forward xfails. Assisted with…
Summary assembled by rule from the sources below