+ "ݳj FRt^RIt]P!]4tRsRRltR#)aq Workaround for a Triton miscompile that produces spurious NaN/Inf in compiled float8 backward kernels (e.g. FSDP2 float8 training with torch.compile). Root cause: PyTorch inductor PR pytorch/pytorch#186933 changed the Triton codegen for ``minimum``/``maximum`` from the ``tl.where``-based ``triton_helpers.{minimum, maximum}`` to ``tl.{minimum,maximum}(a, b, tl.PropagateNan.ALL)``. The two forms are numerically identical (both propagate NaN), but the ``PropagateNan.ALL`` form lowers to PTX ``min.NaN``/``max.NaN`` instructions that make Triton mis-lower a neighboring transposed, vectorized 1-byte (fp8) store. The stored value is correct, but the transposed store writes garbage bytes that decode as NaN/Inf in fp8. See https://github.com/triton-lang/triton/issues/11111. This monkeypatch reverts the inductor min/max codegen to the numerically-identical ``tl.where`` form, which does not emit the ``min.NaN``/``max.NaN`` instructions and so avoids tripping the Triton store bug. It is applied from ``convert_to_float8_training`` so it only affects callers who actually use the float8 training product, not every ``import torchao``. The ``tl.where`` form is always correct, so this is a no-op on numerics even once Triton fixes the underlying bug. Note: we must emit the ``tl.where`` expression *inline* rather than delegate to ``triton_helpers.{minimum,maximum}``. #186933's follow-up also rewrote those helper functions to ``tl.{minimum,maximum}(..., propagate_nan=tl.PropagateNan.ALL)``, so routing through them would still emit ``min.NaN``/``max.NaN`` and fail to work around the bug (confirmed against torch 2.14.0.dev / triton 3.8.0). NFcV^8dQhRR/#)returnN)formats"_/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/float8/_inductor_patch.py __annotate__r*s%%%c\'dR#^RIHp\ R4p\ R4pY nY0nRs\P R4R# \d"p\P RT4Rp?R#Rp?ii;i)zRevert inductor's Triton min/max codegen to the ``tl.where`` form. Idempotent and defensive: if inductor internals have moved, this logs and returns rather than breaking ``import torchao.float8``. N)TritonOverridesz5float8: could not import TritonOverrides to patch: %sc ,RV RV RV RV RV RV R2 #) tl.where((z < ) | ( != ), , )rabs&&rminimum0_patch_inductor_min_max_codegen..minimumA/A3c!E!D3qcA3a@@r c ,RV RV RV RV RV RV R2 #)r z > rrrrrrrs&&rmaximum0_patch_inductor_min_max_codegen..maximumErr Tzqfloat8: patched inductor Triton min/max codegen to work around https://github.com/triton-lang/triton/issues/11111) _PATCHEDtorch._inductor.codegen.tritonr Exceptionloggerdebug staticmethodrr)r errs r_patch_inductor_min_max_codegenr#*sxBAAAA&%H LL =-  LaPsA A?A::A?)__doc__logging getLogger__name__rrr#rr rr(s)8   8 $ %r