v0.1.0

einf v0.1.0 Release Notes

Released: 2026-02-19

Highlights

• First public release of einf, a tensor DSL built around reusable, typed TensorOp objects.
• Axis-signature driven API for defining transformations once and reusing them across calls.
• Unified operation surface for common tensor workflows:
  • rearrange
  • reduce
  • repeat
  • contract
  • einop
  • view

What is included in v0.1.0

• First-class TensorOp construction with configuration methods such as with_sizes(...) and reduce_by(...).
• Variadic axis packs support (packs(...)) for matching zero or more trailing/leading axes.
• Deterministic planning pipeline based on abstract plan -> symbolic plan -> specialized runtime steps.
• Cached operation construction and warm-path reuse behavior for repeated identical specs.
• Typed public API (.pyi and py.typed) for LSP/type-checker support.

Runtime and Dependencies

• Python >=3.10
• Runtime dependencies:
  • array-api-compat>=1.13
  • opt_einsum>=3.4
• Dev extras:
  • pytest>=8.0
  • einops
  • einx

Documentation

• Project overview and examples: README.md
• Internal execution model note: docs/architecture.md

Notes

• TensorOp construction is the cold path (planning/lowering work).
• Reusing constructed op instances is the recommended pattern for best warm-path latency.