AGENT SKILL

project-workflows

Route Align-Anything satellite project workflows and decide runnable versus extension versus reference-only use.

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

01 前言 Front Matter

name
project-workflows
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Route Align-Anything satellite project workflows and decide runnable versus extension versus reference-only use.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Project Workflows

Use this sub-skill when a task mentions an Align-Anything satellite project under projects/, project-local scripts, Janus workflows, InterMT, language-feedback generation, text-image-to-text-image Chameleon workflows, any-to-text model initialization, or the bundled Eval-Anything project. This sub-skill helps decide whether the project material is immediately runnable, an extension pattern for core Align-Anything trainers, or reference-only evidence.

Do not use this sub-skill as the primary guide for ordinary Align-Anything trainer modules, core package APIs, or generic installation. Route those through the root skill or the more specific training/evaluation sub-skills, then return here only for project-folder evidence.

First Decision

  1. Identify the project signal in the user request: any_to_text, janus, intermt, lang_feedback, text_image_to_text_image, or eval-anything.

  2. If working inside an Align-Anything checkout and the user asks what is available, run the bundled discovery script. It parses files and shell snippets but does not import or execute project code:

    python scripts/list_project_entrypoints.py --root <repository-root>
    

    Use --json for machine-readable output.

  3. Read references/project-map.md for routing status, prerequisites, data-shape expectations, and runnable/extension/reference-only decisions.

  4. For eval-anything, also read references/eval-anything-notes.md; treat it as a separate package surface unless a dedicated runtime has been prepared.

  5. When something fails or appears inconsistent with the project README, read references/troubleshooting.md before assuming the core package is broken.

Routing Matrix

Project signal Default treatment Runnable only when Use as extension/reference when
projects/any_to_text Runnable builder scripts plus training-pattern evidence. Align-Anything imports work, Transformers/Torch can load the requested base LLM, CLIP vision tower, and optional CLAP audio tower, and the output directory is intentionally chosen. Designing custom multimodal initializers or interpreting staged any-to-text training flags.
projects/janus and scripts/janus Optional Janus workflow. The separate Janus-compatible package is installed, Janus model weights are available, GPU capacity is planned, and tokenized .pt data conventions match the target trainer. Understanding how Align-Anything wires Janus SFT/DPO generation and understanding trainer shell patterns.
projects/intermt Reference-only by default. Only after a separate InterMT-Bench/data runtime is intentionally prepared. Explaining InterMT dataset/benchmark intent and selecting multi-turn multimodal preference-alignment evidence.
projects/lang_feedback Internal/development workflow; cautious runnable pattern. vLLM, a multimodal model, GPU tensor parallel plan, images, and the expected JSON fields are available. Reusing the base → critique → refine data-generation pattern without claiming stable public support.
projects/text_image_to_text_image Chameleon preprocessing/training pattern with optional runtime. A Chameleon-capable Transformers fork/model and GPU memory are available, and the dataset schema matches the selected tokenizer script. Planning text-image interleaved SFT/DPO/RM/PPO workflows, especially pre-tokenization before core trainers.
projects/eval-anything Separate package/CLI/pipeline reference unless a heavy runtime is prepared. A Python 3.11-compatible Eval-Anything environment with vLLM/HF/API backend dependencies, model weights or API credentials, datasets, and GPU or backend resources is intentionally prepared. Mapping safety benchmark configuration, package entry points, model backends, and pipeline extension points.

Operating Rules

  • Treat project README commands as examples, not proof that a runtime is ready. Confirm optional packages, model identifiers, dataset schema, device count, and output directories first.
  • Do not run repository project scripts merely to discover capabilities. Use the bundled discovery script or inspect files statically.
  • Do not preserve placeholder path tokens in runnable commands; replace them with user-supplied locations or stop and ask.
  • Keep eval-anything separate from the main align_anything package. It has its own package metadata, CLI, configs, benchmark registry, model backend registry, and optional VLA dependency group.
  • If a project depends on another forked repository, external datasets, large model downloads, or API credentials, mark it optional or blocked until those are explicitly prepared.

Handoff Checklist

When handing a project-workflow decision to another sub-skill or to execution, include:

  • Project folder and entrypoint name.
  • Decision: runnable, extension pattern, or reference-only.
  • Required package/runtime deltas beyond the base Align-Anything install.
  • Expected input schema and output artifact type.
  • Device/backend assumptions and whether CPU substitution is unsupported, partial, or unknown.
  • Any known script caveat from references/troubleshooting.md.

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