All notable changes to Comprexx are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

• Benchmarking module (comprexx.benchmark): cx.benchmark() measures real inference latency with configurable warmup/iters, reporting mean, median, std, p50/p90/p99, min/max, and throughput. cx.compare_benchmarks() returns a before/after comparison with speedup and latency/throughput deltas. Quantized models are automatically run on CPU. New comprexx bench CLI command.
• Example notebooks with cell outputs: ResNet18 edge deployment (fuse, prune, benchmark, ONNX export) and linear layer compression (SVD, weight-only INT4, dynamic INT8).
• GitHub Actions CI workflow running pytest on Python 3.10, 3.11, 3.12 plus a ruff check lint job.
• CHANGELOG.md with history for v0.1.0 and v0.2.0.

• Silenced the torch.ao.quantization is deprecated warning inside the PTQ dynamic and static stages. The underlying API is still used, with a TODO marking the upcoming migration to torchao.quantization.
• Fixed the package __version__ to report the correct version.
• Tightened the codebase against ruff check and added a per-file ignore for E741 in tests.

• Unstructured pruning stage: magnitude or random element-wise pruning with global/local scope and optional gradual cubic schedule.
• N:M sparsity stage: structured N-of-M sparsity (default 2:4) for NVIDIA Ampere sparse tensor cores.
• Weight-only quantization stage: group-wise INT4/INT8 with symmetric or asymmetric scaling for Linear and Conv2d layers.
• Low-rank decomposition stage: truncated SVD factorization of Linear layers, with fixed rank-ratio or energy-threshold selection modes.
• Operator fusion stage: Conv2d + BatchNorm2d folding via torch.fx with graceful fallback on non-traceable models.
• Weight clustering stage: per-layer k-means codebook clustering.
• cx.analyze_sensitivity(): per-layer sensitivity probing via prune or noise perturbation. Returns a SensitivityReport that ranks layers by metric drop and can suggest exclude_layers above a threshold.
• New techniques are wired through the recipe schema and loader, and exported from comprexx.stages.

• 163 passing (up from 91).

Initial release.

• Model analysis and profiling via cx.analyze().
• Structured pruning with L1/L2/random criteria and global/local scope.
• Post-training dynamic and static INT8 quantization.
• ONNX export with manifest and optional onnxruntime validation.
• Recipe-driven pipelines (YAML) validated via Pydantic.
• CLI commands: comprexx analyze, compress, export.
• Accuracy guards with halt/warn actions.
• Per-stage compression reports persisted under comprexx_runs/.