AGENT SKILL

module-and-trees

Use Equinox modules, fields, PyTree filtering, and tree surgery correctly.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
module-and-trees
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use Equinox modules, fields, PyTree filtering, and tree surgery correctly.
disable-model-invocation
true

02 资源文件 2

03 SKILL.md 正文

Module and Trees

Use this sub-skill for Equinox’s core model-as-PyTree abstraction and the utilities that manipulate mixed PyTrees.

Use when

  • The user is defining or debugging an eqx.Module subclass.
  • A model has JAX arrays plus Python functions, strings, booleans, or other non-array leaves.
  • The task mentions eqx.field, converters, static fields, AbstractVar, AbstractClassVar, or __check_init__.
  • The task asks for model surgery with tree_at, trainable/static partitioning, tree_equal, tree_check, apply_updates, or Partial.
  • A task needs to prepare a model PyTree for Optax, JAX transforms, or serialization.

Route elsewhere

  • Use ../filtered-transformations/ when the main issue is filter_jit, filter_grad, filter_vmap, filter_pmap, AD, callbacks, or sharding.
  • Use ../nn-and-state/ for built-in equinox.nn layers, stateful modules, Shared, and inference-mode behavior.
  • Use ../diagnostics-and-serialization/ for saving/loading leaves, runtime errors, pretty printing, progress meters, or debug tools.
  • Use ../internal-advanced/ only for semi-public equinox.internal helpers.

Read first

Core workflow

  1. Identify the PyTree boundary. Decide which values are array leaves, non-array dynamic objects, or static metadata.
  2. Define an eqx.Module subclass with annotated fields. Use a custom __init__ only when construction logic needs it; otherwise rely on dataclass initialization.
  3. Mark metadata such as strings, booleans, shapes, or callables with eqx.field(static=True) when it should not be a PyTree leaf.
  4. Use eqx.field(converter=...) for safe field normalization at initialization time.
  5. Use __check_init__ for invariant checks; do not mutate fields inside it.
  6. Split trainable and static leaves with eqx.partition or eqx.filter, using predicates such as eqx.is_inexact_array, before optimizers or raw JAX transforms.
  7. Use eqx.tree_at for out-of-place surgery such as replacing a final layer, changing an initializer result, or toggling nested leaves.
  8. Validate structure with eqx.tree_check and compare expected outputs with eqx.tree_equal.

Minimal validation

From the generated skill root (set EQUINOX_SKILL_ROOT to its absolute path), run the bundled smoke helper after installing Equinox:

cd "$EQUINOX_SKILL_ROOT" && python scripts/smoke.py --mode module

This checks module construction, static-field preservation, partition/combine, tree_at, and tree_check using only the installed package and bundled runtime files.

Scope note

This sub-skill distills Equinox's public module, filter, tree-manipulation, and training-update behavior into the bundled references. Future agents should use references/module-and-trees.md, references/troubleshooting.md, and scripts/smoke.py; the source repository was construction evidence only, not a runtime dependency.

Key cautions

  • An Equinox module is a frozen dataclass and a PyTree. Treat object identity as secondary to tree value and structure.
  • Plain jax.jit and jax.grad are valid only when their dynamic arguments are all JAX-compatible leaves. Use filtered transforms for mixed modules.
  • Do not store bound methods on a module field; use a property or wrapper.
  • Do not mark trainable arrays static just to silence JAX errors; filter the model instead.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351