trunk/296101d35a7ab06f3cc8a4c3ec70f4bb020d5a9e: [caffe2] Use thread-local cache for N-bit embedding bags (#200898)
- PyTorch: 2312 events in the last 90 days
- PyTorch: 2282th Release in the last 90 days
- Previous: earlier the same day · trunk/4290a8548a523d960de21ae86fec8eaa9803c5b1: [inductor] avoid_32bit_indexing_guards: int64 over hint guards (#200811)
What happened
Summary: Opts the quantized CPU embedding-bag implementation into FBGEMM's thread-local dense and row-wise-sparse N-bit kernel caches, removing global cache lock contention from the production lookup path while retaining compatibility with older FBGEMM versions. Test Plan: All CPU-applicable quantized embedding-bag tests pass; GPU-only cases are not applicable to this CPU-path change. The supporting cache tests also…
Summary assembled by rule from the sources below