nvalchemi.distributed.Reduce#
- class nvalchemi.distributed.Reduce(*values)[source]#
How an output’s per-rank value is combined into the global value.
Passed inside
OutputSpec. Mirrors the three consolidation branches inoutput_consolidation.Members#
- NONE
Default per-kind consolidation (e.g. an autograd per-node output is halo-reverse-summed to owners; a per-graph output passes through).
- ALL_REDUCE
Each rank holds a partial; sum across the mesh to the global value. (Maps to
MLIPSpec.all_reduce_outputs.)- OWNED_ONLY
Already globally-correct on every rank; slice/passthrough, no cross-rank reduce. (Maps to
MLIPSpec.owned_only_outputs.)