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 class

NamedMapping

NamedMapping

class max.tensor.sharding.NamedMapping(mesh=None, spec=(), *, unreduced=())

source

Bases: DeviceMapping

Builds a DeviceMapping from a JAX-style named spec.

Each spec entry corresponds to a tensor dim and names the mesh axis that shards it (or None for replicated). Mesh-axis names not present on the target mesh resolve to Replicated. unreduced names mesh axes carrying a pending reduction; each becomes a Partial placement.

The resolved placements are stored on the instance, so it behaves as a regular DeviceMapping. It also keeps the original spec, which _resolve uses to rebuild the mapping on another mesh.

Parameters:

• mesh (DeviceMesh) – The target device mesh, or None to take the one set by default_device().
• spec (tuple[SpecEntry, ...]) – One entry per tensor dimension.
• unreduced (Iterable[str]) – Mesh axes carrying pending reductions.

Raises:

• ValueError – If mesh is None and no default_device() is active, leaving the spec with nothing to resolve against.
• ConversionError – If one mesh axis is named for more than one tensor dimension, or if an axis appears in both spec and unreduced.