jax
dfbd6c9b - [partitioning] Canonicalize reduce and gather all-reduce axes to mesh order.

Commit
2 days ago
[partitioning] Canonicalize reduce and gather all-reduce axes to mesh order. `_reduce_collective_rule` and `_gather_collective_rule` previously built their all-reduce (`psum` / `pmax`) axis tuples in operand spec order, emitting mixed intra-slice / cross-slice replica groups when an operand spec interleaved intra-slice and cross-slice mesh axes. Canonicalize both via `core.order_wrt_mesh`. PiperOrigin-RevId: 1022965967
Author
Parents
Loading