For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python data

DEFAULT_TRANSITIONS

DEFAULT_TRANSITIONS

max.experimental.sharding.DEFAULT_TRANSITIONS = frozenset({(<class 'max.experimental.sharding.placements.Partial'>, <class 'max.experimental.sharding.placements.Replicated'>), (<class 'max.experimental.sharding.placements.Replicated'>, <class 'max.experimental.sharding.placements.Sharded'>), (<class 'max.experimental.sharding.placements.Sharded'>, <class 'max.experimental.sharding.placements.Replicated'>), (<class 'max.experimental.sharding.placements.Sharded'>, <class 'max.experimental.sharding.placements.Sharded'>)})

source

Every transition except (Partial, Sharded), so a partial sum resolves to Replicated by allreduce; ALL_TRANSITIONS also allows reduce-scatter.