DevOnBike.Overfit 10.0.18

Sources/Main

Core Overfit runtime and library code.

Current runtime areas

Autograd      Reverse-mode graph and training operations
DeepLearning  Layers, Sequential, train/eval state, save/load
Inference     Prepared zero-allocation inference facade
Ops           TensorMath and graph-aware training math
Tensors       Tensor storage, views and memory abstractions
Evolutionary  Population-based gradient-free optimization
Onnx          Focused ONNX import MVP for PyTorch-exported inference models

Inference path

The preferred production-style inference API is:

engine.Run(input, output);

The hot path should avoid:

model.Forward(...)
AutogradNode
ComputationGraph
new arrays per call
ToArray()
LINQ in runtime code

InferenceEngine.Run(...) uses caller-owned input/output buffers and prepared reusable internal buffers.

ONNX import

The ONNX importer is a focused load-time feature. It is not a full ONNX runtime.

Current MVP goal:

PyTorch-exported eval-mode ONNX CNN
-> OnnxImporter.Load(path)
-> Sequential
-> InferenceEngine.Run(...)

Supported MVP operators:

Conv
Relu
MaxPool
Reshape / Flatten
Gemm

Constraints:

FP32 only
NCHW layout
Linear topology only
Concrete shapes required
No branching / skip connections
No grouped/depthwise conv
No quantized or FP16 tensors

Example:

using DevOnBike.Overfit.Onnx;
using DevOnBike.Overfit.Inference;

var model = OnnxImporter.Load("mnist_cnn.onnx");
model.Eval();

using var engine = InferenceEngine.FromSequential(
    model,
    inputSize: 1 * 28 * 28,
    outputSize: 10);

var input = new float[1 * 28 * 28];
var output = new float[10];

engine.Run(input, output);

Main-project coding rules

Runtime code in Sources/Main should stay conservative:

• Avoid LINQ in hot/runtime paths.
• Prefer explicit loops and spans.
• Avoid hidden allocations.
• Keep import-time allocations out of inference hot paths.
• Keep training/graph allocation policy separate from inference policy.
• Do not add dependencies unless there is a clear architectural reason.

Benchmark status

Current verified hot-path inference results include:

Linear(784,10): ~250-300 ns/op, 0 B
Linear(4096,10): ~1.08 us, 0 B
MLP 784->256->128->10: ~10-12 us, 0 B
Small CNN: ~5-6.5 us, 0 B
Imported ONNX MNIST CNN: ~7.5 us, 0 B

See Sources/Benchmark/README.md and docs/InferenceBenchmarkSummary.md for full benchmark tables.