ciflow/torchtitan/201029: Keep a zero exact through a low-precision rounding cast
- PyTorch: 2484 events in the last 90 days
- PyTorch: 2451th Release in the last 90 days
- Previous: earlier the same day · ciflow/torchtitan/201319: Update
What happened
Rounding a constant zero to float16 or bfloat16 gives exactly zero, so the prologue zero-mask analysis can keep it, as it did before #197132 . A prologue whose only low-precision op is a rounding cast then keeps its masked zeros and gets no extra tl.where, matching 2.14's generated code. Other values still become unknown. The end-to-end test now checks the generated code for the mask, the same way test_upcast does.
Summary assembled by rule from the sources below