ciflow/trunk/199610: [T273452478] Cache FBGEMM embedding kernel handles (#199610)
- PyTorch: 1981 events in the last 90 days
- PyTorch: 1954th Release in the last 90 days
- Previous: earlier the same day · ciflow/torchtitan/199988
What happened
Summary: Pull Request resolved: #199610 Adds separate thread-local single-entry caches for dense and row-wise-sparse FBGEMM embedding kernels, keyed by D and has_weight . Steady-state calls reuse the handle while key changes regenerate through the existing path. One entry per template instantiation intentionally bounds thread-local state and optimizes stable shapes; alternating keys regenerate. FBGEMM dispatch overr…
Summary assembled by rule from the sources below