AGENT SKILL

quark0-darts

Use the original DARTS research code for differentiable architecture search, CNN/RNN genotypes, CIFAR/ImageNet/PTB/WT2 workflows, and legacy PyTorch troubleshooting.

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

01 前言 Front Matter

name
quark0-darts
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use the original DARTS research code for differentiable architecture search, CNN/RNN genotypes, CIFAR/ImageNet/PTB/WT2 workflows, and legacy PyTorch troubleshooting.
disable-model-invocation
true

02 资源文件 7

03 SKILL.md 正文

DARTS

Use this repo skill when a task is about the original DARTS: Differentiable Architecture Search codebase: convolutional architecture search/evaluation, recurrent language-modeling search/evaluation, genotype schemas, visualization, dataset/checkpoint layout, or legacy PyTorch/CUDA failures.

This is a source-backed operating skill for a script-style research repository. It does not assume the original checkout is available, does not install a package, and does not claim that paper-scale native runs were executed during skill production.

Start here

  1. Read legacy runtime before promising native execution. The original code targets Python 3.5-era PyTorch 0.3.1/torchvision 0.2.0 and CUDA.
  2. Read data and checkpoints when a workflow needs CIFAR-10, ImageNet, PTB, WikiText-2, or pretrained model files.
  3. Read troubleshooting for cross-cutting failures such as async=True syntax errors, missing CUDA, dataset assertions, checkpoint-format mismatches, and Graphviz issues.
  4. Check repo provenance before treating this skill as current for a different DARTS checkout.

Route by task

User task Read
CIFAR-10 DARTS architecture search, CIFAR-10 training/test, ImageNet training/test, CNN model/cell/operation facts, or CNN CUDA/OOM/nondeterminism failures cnn-architectures
Penn Treebank or WikiText-2 recurrent DARTS search/training/test, corpus layout, checkpoint behavior, ASGD schedule, perplexity interpretation, or RNN CUDA/hidden-size/data failures rnn-language-modeling
CNN/RNN genotype schemas, built-in DARTS/NASNet/AmoebaNet/ENAS genotypes, converting search outputs, validating custom genotypes, or DOT visualization genotypes-and-visualization
Construct a native command without launching a job darts_command_builder.py
Inspect a local DARTS-style source tree before running or porting it darts_static_inspector.py

Safe helper scripts

  • Run python scripts/darts_command_builder.py list to see workflow ids.
  • Run python scripts/darts_command_builder.py build cnn-search --smoke --gpu 0 to print a wiring-check command and prerequisites. The helper does not run training or download data.
  • Run python scripts/darts_static_inspector.py --repo-root <path> to statically check an external DARTS source tree. The helper does not import the repo.
  • Run python sub-skills/genotypes-and-visualization/scripts/darts_genotype_tools.py list to inspect bundled genotype catalogs and emit DOT without Graphviz.

Operating rules

  • Distinguish architecture search from final evaluation. Search validation accuracy/perplexity is not the paper result.
  • Treat full native workflows as legacy CUDA jobs. If the user is on modern Python/PyTorch, route to porting guidance instead of implying direct compatibility.
  • Never report smoke-mode output as paper accuracy or perplexity.
  • Keep CNN and RNN genotype schemas separate; both define DARTS, but the tuple fields and valid edges differ.
  • Do not tell future agents to open original repo scripts or README pages for routine operation. Use the bundled references and scripts in this skill.
  • If a requested dataset, pretrained checkpoint, CUDA runtime, or legacy PyTorch build is missing, report the missing prerequisite clearly and stop before inventing a result.

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