CompiledTapeTrainingStep<T>
Helpers & Utilities · AiDotNet.Training
Compiled training step — auto-compiles the forward + backward pass on the first step, then replays the compiled plan on subsequent steps for near-zero overhead training.
How it works:
- Step 1 (tracing): Enables GraphMode, traces the forward pass + loss computation
through the layer stack, compiles a CompiledTrainingPlan with backward pass, and executes it.
- Steps 2+ (replay): Calls plan.Step() which replays the compiled forward + backward
as flat delegate arrays with pre-allocated gradient buffers. Zero allocation, zero dispatch overhead.
Recompilation triggers:
- Input shape changes (different batch size, sequence length, etc.)
- Explicit Invalidate() call (model structure changed)
- Compilation failure (falls back to eager TapeTrainingStep for that shape)
Methods
| Method | Summary |
|---|---|
Exception GetLastFallbackException() |
AiDotNet#1395: read the last exception that caused TryStepWithFusedOptimizer to fall back, or null if the most recent fallback was due to one of the explicit return-false paths (plan switch, config drift, EnableCompilation=false, etc.)… |
T Step(IReadOnlyList<ITrainableLayer<T>> layers, Tensor<T> input, Tensor<T> target, T learningRate, Func<Tensor<T>, Tensor<T>> forward, Func<Tensor<T>, Tensor<T>, Tensor<T>> computeLoss, object owner) |
Executes a single compiled training step. |
long GetFusedStepCount() |
Gets the count of successful fused-step executions on the calling thread. |
void Invalidate() |
|
void ResetFusedStepCount() |
Resets the fused-step counter on the calling thread to zero. |