trunk/4290a8548a523d960de21ae86fec8eaa9803c5b1: [inductor] avoid_32bit_indexing_guards: int64 over hint guards (#200811)
- PyTorch: 2311 events in the last 90 days
- PyTorch: 2281th Release in the last 90 days
- Previous: earlier the same day · trunk/5232ed0e9aa7558842a3d79c8b59a009ecc2d411
What happened
Problem. For a dynamic size with no upper bound, Inductor picks int32 indexing when the compile-time hint fits and installs a size guard, so a pure codegen choice becomes a Dynamo specialization boundary. A precompile artifact captured on ordinary inputs then misses (and, under strict serving, refuses) on the first input past that bound. Fix. An opt-in config, torch._inductor.config.avoid_32bit_indexing_guards : wit…
Summary assembled by rule from the sources below