AGENT SKILL

nn-and-state

Use Equinox neural-network layers, Sequential models, stateful layers, and inference mode.

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

01 前言 Front Matter

name
nn-and-state
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use Equinox neural-network layers, Sequential models, stateful layers, and inference mode.
disable-model-invocation
true

02 资源文件 2

03 SKILL.md 正文

Neural Networks and State

Use this sub-skill for equinox.nn layers and stateful model workflows.

Use when

  • The user asks how to build a model with eqx.nn.Linear, MLP, Sequential, Conv*, MultiheadAttention, Embedding, pooling, RNN cells, or normalization layers.
  • A task needs Dropout, BatchNorm, SpectralNorm, WeightNorm, inference_mode, or explicit PRNG key handling.
  • The user needs eqx.nn.State, StateIndex, make_with_state, delete_init_state, or a custom stateful layer.
  • A model needs tied parameters with eqx.nn.Shared.
  • A training loop combines Equinox layers with filter_grad, Optax, and apply_updates.

Route elsewhere

  • Use ../module-and-trees/ for raw eqx.Module field design, tree surgery, filtering, and optimizer parameter selection.
  • Use ../filtered-transformations/ when the dominant issue is JIT, AD, vmap, pmap, callbacks, or sharding.
  • Use ../diagnostics-and-serialization/ for saving/loading, runtime errors, debug tracing, or pretty printing.
  • Use ../internal-advanced/ for semi-public internal loops and primitives.

Read first

Core workflow

  1. Choose the layer family and confirm whether it is stateless, stochastic, or stateful.
  2. Remember that most equinox.nn layers consume one example. Use jax.vmap or eqx.filter_vmap around layer calls for batches.
  3. Use Sequential for ordered layers. Wrap activation functions with eqx.nn.Lambda when they live inside Sequential.
  4. Pass PRNG keys to constructors and to stochastic calls such as dropout.
  5. For stateful layers, construct with eqx.nn.make_with_state(ModelClass) and thread the returned State through every call.
  6. Toggle inference behavior with eqx.nn.inference_mode(model, value=True).
  7. Use eqx.nn.Shared for intentional tied weights instead of relying on object identity in a PyTree.
  8. For training, combine this sub-skill with filtered-transformations and module-and-trees.

Minimal validation

From the generated skill root (set EQUINOX_SKILL_ROOT to its absolute path), run:

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

This validates tiny MLP, Sequential, custom state, and Shared examples. It does not run expensive training or download datasets.

Scope note

This sub-skill distills Equinox's built-in layer, state, and sharing workflows into the bundled references and smoke helper. The source repository's tests, docs, and notebooks were construction evidence only, not runtime dependencies.

Key cautions

  • BatchNorm must run inside a vmap or pmap with a matching axis_name for training statistics.
  • State.set and State.update return a new state and invalidate the old one.
  • MLP(..., scan=True) changes the internal representation for compile-time reasons; do not call the scanned hidden layer stack directly.
  • Shared requires matching source/destination structure, shapes, and dtypes.
  • Stochastic layers use explicit keys; do not hide key generation inside a JITed training step without a reproducible split plan.

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