Expand description
Architecture-declared parallel parameter semantics and local layouts. Backend-neutral semantic parameter sharding and rank-local layouts.
Architectures describe physical checkpoint members in logical groups. An execution backend may then realize the resulting placement without knowing projection names, attention geometry, or other model-family semantics.
Structs§
- Local
Model Layout - Complete rank-local model geometry produced alongside checkpoint placement.
- Local
Tensor Layout - Rank-local shape and placement for one planned physical tensor.
- Parallel
Model Info - Architecture-neutral information for one rank-local parallel model.
- Parameter
Group Spec - Atomic logical parameter and all of its physical checkpoint companions.
- Parameter
Member Spec - One physical tensor belonging to a logical parameter group.
Enums§
- Member
Sharding - Rank-local selection rule for one physical checkpoint tensor.
- Parallel
Plan Error - Invalid architecture-declared parallel semantics.
- Parameter
Role - Semantic role of a logical parameter group.
- Projection
Sharding - Logical sharding behavior for a parameterized affine projection.
- Sharding
Policy - Behavior when a requested shard is not legal for the current TP size.
- Tensor
Placement - Backend-neutral placement decision for one physical tensor.
Functions§
- aligned_
partition_ units - Returns the finest legal logical-unit count for an aligned partition.
- expand_
linear_ format_ parameter_ groups - Rewrites semantic dense matrix declarations into their authoritative physical checkpoint representation and publishes every required companion in the same atomic parameter group.
- module_
parameter_ group - Describes every parameter in a neutral module as one logical group.
- partitioned_
module_ parameter_ group - Describes every parameter in a neutral module as one shared logical partition.
- partitioned_
projection_ group - Describes projections that consume one shared logical partition.
- projection_
parameter_ group - Describes one affine projection and all encoding companions.
- segmented_
projection_ group - Describes a component-major fused column projection and its row-parallel output as one shared logical partition.