v0.2.0

What's new in v0.2.0

GPU Acceleration Layer

• New ComputeBackend protocol covering the two O(V×D²) GEMMs that dominate encode/decode cost: rotate, rotate_inverse, project, normalize_rows, restore_norms
• NumPyBackend — always available, delegates to canonical implementations to prevent drift
• TorchBackend — CUDA (NVIDIA) and MPS (Apple Silicon via PyTorch); pip install semafold[torch]
• MLXBackend — Metal (Apple Silicon via MLX); pip install semafold[mlx]
• Thread-safe auto-detection registry with priority chain: CUDA → MPS → MLX → NumPy
• On-device matrix cache keyed by (ctypes.data, shape) — bounded to ≤ 64 entries, safe against id() reuse after LRU eviction
• list_backends() probes availability without instantiating backend objects
• Plain pip install semafold unchanged — NumPy remains the default, no new required dependencies

Typed Enum Surface

• EncodeObjective, EncodeMetric, EncodingSegmentKind added to the stable root export
• EncodingBoundType, WorkloadSuitability added to the stable root export
• All enum fields accept both enum instances and string values — existing string-based code continues to work

Quality

• Windows compatibility verified
• 189 tests passing across Python 3.10–3.13
• Pyright 0 errors

Install

pip install semafold              # NumPy core — no GPU required
pip install semafold[torch]       # + NVIDIA CUDA / Apple MPS
pip install semafold[mlx]         # + Apple Silicon Metal

## Upgrade notes

No breaking changes. String values for enum fields continue to work — enums are the preferred form going forward.

v0.1.0

Initial release of semafold — vector-first compression for embeddings, retrieval, and KV-cache workloads.

What's included

• Stable root API: CompressionBudget, CompressionEstimate, CompressionFootprint, CompressionGuarantee, ValidationEvidence
• PassthroughVectorCodec and ScalarReferenceVectorCodec
• Preview-only TurboQuant family under deep import semafold.turboquant:
  • TurboQuantMSEVectorCodec — reconstruction-optimized quantization
  • TurboQuantProdVectorCodec — inner-product-optimized quantization with QJL sign sketch
  • TurboQuantKVPreviewCodec — KV-cache compression (keys: Prod, values: MSE)
• Explicit byte accounting and validation evidence on all encodings
• Deterministic synthetic benchmarks

Install

pip install semafold==0.1.0