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 module

max.tensor.sharding.per_shard_dim

Wraps one Dim per device, so a dimension can vary by shard.

A PerShardDim bundles one Dim per mesh shard (device index along a Sharded mesh axis). Arithmetic distributes element-wise. make_per_shard_dim() collapses to its first cell when every cell is equal, no global_dim is attached, and force_wrap is false. Shape rules project each wrapper per shard before any Dim.to_mlir() call, so a wrapper never reaches the MLIR graph.

Per-shard dimensions

PerShardDim	A Dim listing one cell per mesh shard.

global_dim	Folds a PerShardDim into a single global Dim.
global_shape	Returns the global shape by summing every PerShardDim axis.
is_per_shard_dim	True if d is a PerShardDim.
is_static	True if d's global extent is a static (compile-time) size.
is_symbolic	True if d's global extent is a symbolic (named) dim.
local_dim_at	Returns the dimension value at shard.
local_shape_at	Projects every axis of shape onto a single shard index.
make_per_shard_dim	Wraps per_shard in a PerShardDim, or collapses to a plain Dim.