Tutorials

Interactive, runnable tutorials for SNEPPX-Algo. Each tutorial has a companion Jupyter notebook under docs/tutorials/notebooks/ that you can download and run:

$env:PYTHONPATH = "bindings/python"
pip install jupyter
jupyter notebook docs/tutorials/notebooks/classification.ipynb

The notebooks use the pure-Python / NumPy fallback where possible so they run without a compiled C backend. Steps that require the C backend (backward(), Trainer.fit, real model forward on large weights) are explicitly marked :material-alert-decagram: **C backend required** and guarded with a runtime check.

Tutorial map

Tutorial	Notebook	Skill	What you build
Classification	classification.ipynb	Beginner	MLP on MNIST-style data
Text Generation	generation.ipynb	Beginner–Int	Greedy / sampling / beam search
RLHF Fine-Tuning	fine_tuning_rlhf.ipynb	Advanced	LoRA + DPOTrainer
Quantization + Serving	quantization_serving.ipynb	Intermediate	INT4/AWQ quant + sneppx-serve
Distributed Training	distributed_training.ipynb	Advanced	ZeRO-1 + DDP
MoE SER Routing	moe_ser_routing.ipynb	Intermediate	8-expert top-2 routing
Security Scanning	security_scanning.ipynb	Intermediate	sneppx-analyze + S0 crypto
Profiling & Benchmarks	profiling_benchmarks.ipynb	Intermediate	Profiler + sneppx-bench
Data Pipeline	data_pipeline.ipynb	Beginner	Tokenizer + DataLoader + streaming
Model Conversion	model_conversion.ipynb	Intermediate	HF → SNEPPX checkpoints

Prerequisites

# Build (for real training/gradient steps)
cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
$env:PYTHONPATH = "bindings/python"

# Verify
python -c "import SneppX_ALG as s; print('C backend:', s._HAS_C_BACKEND)"