SNEPPX-Algo Cookbook¶
Copy-paste recipes for getting things done with SNEPPX-Algo. Every snippet
uses the verified Python API (import paths confirmed against
SneppX_ALG). Recipes are grouped by category; each has an intent, a
snippet, and notes (gotchas, C-backend requirements).
Set your path once per session:
$env:PYTHONPATH = "bindings/python"
Categories¶
| Category | Recipes | When to reach for it |
|---|---|---|
| Tensors | 10 | Creating, reshaping, math, autograd |
| Models & Layers | 8 | Building networks with Module/nn |
| Training | 6 | Loops, loss, checkpoints, Trainer |
| Optimizers | 5 | SGD/AdamW/Lion/LAMB + schedulers |
| Quantization | 8 | INT8/INT4/FP8/AWQ/GPTQ, serving |
| Distributed | 5 | ZeRO, DDP, launch, sampler |
| Serving & Inference | 4 | sneppx-serve, FastAPI, batching |
| Data & Tokenization | 5 | Dataset, DataLoader, Tokenizer |
| Security | 5 | Scan, PQ crypto, key vault, attestation |
| Profiling | 3 | Profiler, timeit, MemoryTracker |
| Checkpointing | 3 | Save/load, async, fault tolerance |
| Conversion | 3 | HF ↔ SNEPPX, safetensors, ONNX |
| Generation | 4 | generate, sampling, beam search, streaming |
Total: 71 recipes. Last updated by the docs maintainer.
Legend¶
- :material-alert-decagram: C backend required — raises
RuntimeError: C backend not availablewithout_SNEPPX_c. - :material-cpu-chip: CPU-safe — runs on pure NumPy, no build needed.
- :material-gpu: GPU — needs
SNEPPX_BUILD_CUDA=ONand a CUDA device.
Import quick-ref¶
from SneppX_ALG import Tensor, TensorDataset, AdamW, Linear, Trainer, Transformer
# Sub-module (not re-exported via *):
from SneppX_ALG.interface_bindings.data_loader import DataLoader # DataLoader
from SneppX_ALG.interface_bindings.tokenizer import Tokenizer # HuggingFace/byte-level
from SneppX_ALG.interface_bindings.generation import generate, GenerationConfig, TextStreamer
from SneppX_ALG.interface_bindings.quantized_serve import quantize_model_weights, QuantizedModelConfig