SubModel<T>
Models & Types · AiDotNet.Interfaces
Represents a contiguous sub-model extracted from a larger ILayeredModel.
Contains a slice of layers that can be independently forwarded through and further partitioned.
Implements: ILayeredModel<T>
For Beginners
When you split a neural network across GPUs for pipeline parallelism, each GPU gets a sub-model - a consecutive sequence of layers from the original network. This class represents that slice.
How It Works
A sub-model implements ILayeredModel itself, so you can extract
sub-models of sub-models, enabling hierarchical partitioning (e.g., virtual pipeline stages
in Megatron-LM).
Reference: Inspired by PyTorch’s split_module() which returns
nn.Sequential modules for each pipeline stage.
Constructors
| Constructor | Summary |
|---|---|
SubModel(IReadOnlyList<ILayer<T>> layers, IReadOnlyList<LayerInfo<T>> layerInfos, int startIndex, int endIndex) |
Creates a new sub-model from the specified layers and metadata. |
Properties
| Property | Summary |
|---|---|
IReadOnlyList<ILayer<T>> Layers { get; } |
Gets the ordered list of layers in this sub-model. |
IReadOnlyList<LayerInfo<T>> LayerInfos { get; } |
Gets the layer metadata for all layers in this sub-model. |
int EndIndex { get; } |
Gets the end index (inclusive) of this sub-model within the parent model. |
int LayerCount { get; } |
Gets the number of layers in this sub-model. |
int StartIndex { get; } |
Gets the start index of this sub-model within the parent model. |
long ParameterCount { get; } |
Gets the total parameter count across all layers in this sub-model. |
long TotalEstimatedFlops { get; } |
Gets the total estimated FLOPs across all layers in this sub-model. |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<LayerInfo<T>> GetAllLayerInfo() |
|
LayerInfo<T> GetLayerInfo(int layerIndex) |
|
SubModel<T> ExtractSubModel(int startLayer, int endLayer) |
|
Tensor<T> Forward(Tensor<T> input) |
Performs a sequential forward pass through all layers in this sub-model. |
bool ValidatePartitionPoint(int afterLayerIndex) |
|
int[] GetInputShape() |
Gets the input shape expected by the first layer in this sub-model. |
int[] GetOutputShape() |
Gets the output shape produced by the last layer in this sub-model. |