UniformPartitionStrategy<T>
Models & Types · AiDotNet.DistributedTraining
Divides model parameters evenly across pipeline stages.
Implements: IPipelinePartitionStrategy<T>
For Beginners
This is the default strategy. It splits the model like cutting a cake into equal slices. It works well when all layers have similar computational cost, but can cause imbalance when some layers (like attention) are much heavier than others.
How It Works
This is the simplest partitioning strategy: each stage gets approximately the same number of parameters. When the total isn’t evenly divisible, earlier stages get one extra parameter each.
Constructors
| Constructor | Summary |
|---|---|
UniformPartitionStrategy() |
Methods
| Method | Summary |
|---|---|
ValueTuple<int, int>[] ComputePartition(int totalParameters, int numStages) |