einops Repo Skill

Use this repo skill when a task involves the einops Python package: readable and shape-checked tensor manipulation, named axes, deep-learning tensor refactors, reversible token/feature packing, optional framework layers, or maintainer workflows for the einops repository.

Install and Minimal Check

Public package install:

pip install einops

einops declares no runtime dependencies. Install NumPy, PyTorch, TensorFlow, JAX, or another tensor framework separately according to the user's project. For a minimal core check with NumPy installed:

import numpy as np
from einops import rearrange, reduce, repeat

x = np.arange(2 * 3 * 4).reshape(2, 3, 4)
assert rearrange(x, "batch channel time -> batch time channel").shape == (2, 4, 3)
assert reduce(x, "batch channel time -> batch time", "sum").shape == (2, 4)
assert repeat(x, "batch channel time -> batch channel time copy", copy=2).shape == (2, 3, 4, 2)

For a bundled diagnostic, run scripts/check_einops_install.py after installing einops and, for the full smoke, numpy.

Route Map

• sub-skills/tensor-operations/: use rearrange, reduce, repeat, parse_shape, and asnumpy for core tensor shape transformations, pooling, broadcasting, stack/concatenate, ellipsis, and pattern troubleshooting.
• sub-skills/named-einsum-and-packing/: use einsum, pack, and unpack for named-axis contractions, attention-like dot products, class-token/multimodal packing, packed-shape (PS) handling, and reversible split/merge workflows.
• sub-skills/framework-integrations/: use optional backend dispatch, Array API functions, framework layers, EinMix, torch scripting/compilation notes, and missing optional dependency diagnostics.
• sub-skills/repo-development/: use maintainer guidance for focused tests, backend selection, EINOPS_TEST_BACKENDS, docs/notebook checks, formatting/type checks, CI matrix interpretation, and release/deploy boundaries.

Core Decision Points

• If the user wants a readable replacement for reshape, view, permute, transpose, pooling, tiling, broadcasting, or shape parsing, start with tensor-operations.
• If the operation is a mathematical contraction or a dot product with named axes, use named-einsum-and-packing and remember that einops.einsum takes tensors first and the pattern last.
• If multiple tensors must be concatenated and later split without losing their heterogeneous middle dimensions, use pack/unpack rather than manual slices.
• If the transform belongs inside a neural-network model definition, serialized layer, Keras model, Flax module, or Torch Sequential, use framework-integrations.
• If the task is about changing the repository, running native tests, or building docs, use repo-development instead of package-usage routes.

Shared References

• references/repo-provenance.md: source commit, tag, package version, evidence paths, and refresh cues.
• references/repo-routing-metadata.json: structured scenario metadata for DisCo repo-skill routing if imported later.
• references/troubleshooting.md: cross-cutting install/import, optional dependency, backend, and stale-skill troubleshooting.

Safe Operating Rules

• Keep runtime guidance self-contained. Do not require future agents to open or run original repository notebooks, tests, docs, or scripts to complete normal package-usage tasks.
• Treat optional frameworks as optional. A CPU NumPy smoke does not prove CUDA, ROCm, MPS, TensorFlow, JAX, or PyTorch runtime behavior.
• Prefer explicit semantic axis names (batch, channel, height, width) in examples that explain user intent.
• When decomposing axes such as (height h2), supply or validate lengths that cannot be inferred safely.
• For maintainer commands that can mutate an environment or docs tree, use the repo-development sub-skill's dry-run helpers before executing.

Refresh Triggers

Read references/repo-provenance.md before deciding this skill is current for a checkout. Refresh the repo skill if the package version, public API exports, pattern grammar, backend list, layer modules, native test runner, docs scripts, CI matrix, or relevant evidence paths changed since the recorded snapshot.