EveryO v0.2.0 — Scaling, Interoperability and GPU Runtime

EveryO v0.2.0 advances the framework from a CPU-first foundation into a broader research runtime for modern neural networks, interoperability, profiling, distributed execution, and persistent GPU inference.

Highlights

Models and automatic differentiation

• Convolution and pooling layers
• Normalization, recurrent, embedding, attention, and transformer building blocks
• Autograd coverage for the operations used by the expanded model stack
• Unified operator dispatch across NumPy and CUDA-capable backends

Performance and runtime

• Mixed-precision training utilities and gradient scaling
• GPU-resident CUDA inference tensors with explicit host/device transfers
• Dynamic INT8 model quantization
• Structured model profiling with Chrome/Perfetto trace export

Scale and interoperability

• Single-machine data parallelism
• Multi-node TCP process groups for trusted private networks
• ONNX layer export plus tracing-based export for recurrent and attention models
• Optional TensorFlow reference and interoperability paths

Engineering quality

• Cross-platform CI for Python 3.9–3.12
• Dedicated TensorFlow, ONNX, and CUDA build/import checks
• 709 tests documented across core, numerical, integration, and optional-backend behavior
• Expanded governance, security, issue-routing, and contributor documentation

Upgrade from v0.1.0

git fetch --tags
git checkout v0.2.0
python -m pip install -e ".[dev]"
python -m pytest -q

Important boundaries

• EveryO remains an educational and research-oriented framework, not a production replacement for PyTorch or TensorFlow.
• The GPU-resident tensor API is currently inference-oriented.
• The TCP distributed backend is designed for trusted private networks and does not provide NCCL, RDMA, or transport encryption.
• CUDA numerical execution requires supported NVIDIA hardware; CPU fallback remains available for core features.

Full changelog: v0.1.0...v0.2.0

Target commit: 670eba8