README.md

nanoptq/

The core library. Everything runnable lives here.

Structure

nanoptq/
├── core/          # Quantization math primitives
├── model/         # QuantLinear layer + HF model loading
├── algorithms/    # RTN, AWQ, GPTQ implementations
├── io/            # Save/load quantized checkpoints
├── eval/          # Perplexity and latency benchmarks
├── data/          # Calibration/eval dataset loader
└── cli.py         # Entry point: nanoptq quantize / eval / compare

Reading Order

If you are learning, read in this order:

Step	File	What you learn
1	core/quant_primitives.py	Symmetric, asymmetric, fake_quant math
2	core/group_quant.py	Why group-wise dramatically improves int4
3	model/quant_linear.py	Unified quantized layer; dequant-on-the-fly
4	algorithms/rtn.py	Baseline: round and done
5	algorithms/awq_lite.py	Activation-aware improvement
6	algorithms/gptq_lite.py	Hessian-based compensation

Entry Point

cli.py ties everything together. Each subcommand imports only what it needs.

nanoptq quantize  →  algorithms/ + io/
nanoptq eval      →  io/ + eval/
nanoptq compare   →  algorithms/ + eval/