[Prajwal7-2002]

Summary

• Runnable Example Scripts:
  • \examples/01_quant_linear_inference.py: Drop-in QuantLinear replacement, bit-packing memory footprint, and recursive transformer conversion with float \lm_head\ preservation.
  • \examples/02_advanced_calibration.py: SmoothQuant continuous mathematical invariance verification, AWQ Frobenius grid search, and GPTQ second-order Hessian error compensation with 1.48x error reduction.
  • \examples/03_servebench_roofline_eval.py: Hardware roofline modeling on consumer GPUs (RTX 3050, T4), demonstrating autoregressive decode arithmetic intensity = 2 / \text{bytes_per_param}$ and .4%$ DRAM bandwidth traffic reduction under INT4.
• Contributing Guide:
  • Added \CONTRIBUTING.md\ covering \uv\ development setup, invariant testing protocols, zero compiled C++/CUDA principles, and pull request conventions.
• Documentation Polish:
  • Updated \README.md\ with PyPI installation instructions, test suite metrics (47 tests, 95% coverage), and references to examples and contribution guidelines.