AGENT SKILL

multimodal-generation

Use UniLM Kosmos, TextDiffuser, WavLM, BEATs, SpeechT5, SpeechLM, VALL-E, and LatentLM for safe multimodal generation, OCR/markdown, audio, and latent diffusion workflow planning.

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

01 前言 Front Matter

name
multimodal-generation
license
MIT
metadata
{ "disco-role": "operating" }
description
Use UniLM Kosmos, TextDiffuser, WavLM, BEATs, SpeechT5, SpeechLM, VALL-E, and LatentLM for safe multimodal generation, OCR/markdown, audio, and latent diffusion workflow planning.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Multimodal Generation

Use this sub-skill when a task names Kosmos, Kosmos-2.5, TextDiffuser, TextDiffuser-2, WavLM, BEATs, SpeechT5, SpeechLM, VALL-E, or LatentLM, or when the request is about open-ended image/text generation, grounded multimodal generation, literate OCR/markdown generation, audio representation extraction, speech generation, or latent diffusion sampling/evaluation.

Route Here

  • Validate or plan Kosmos-2 grounded image-text generation, phrase grounding, captioning, VQA, GRIT-style data preparation, or fairseq generation commands.
  • Validate or plan Kosmos-2.5 OCR and image-to-markdown inference with --do_ocr or --do_md, local image/checkpoint inputs, optional aspect-ratio preprocessing, and OCR bounding-box visualization.
  • Choose TextDiffuser v1 text-to-image, template-conditioned generation, text inpainting, MARIO-Eval sampling, or TextDiffuser-2 full/LoRA/inpainting command shapes.
  • Plan WavLM, BEATs, SpeechT5, SpeechLM, or VALL-E audio/speech workflows, including checkpoint loading, 16 kHz audio assumptions, fairseq manifests, TTS/VC generation, and reference-only zero-shot TTS notes.
  • Validate LatentLM sampling, many-sample generation, training, or FID evaluation inputs, including checkpoint layout, VAE, ImageFolder/Hugging Face dataset choices, and reference-stat requirements.

Route Elsewhere

  • Use vision-document-ai for LayoutLM, LayoutLMv2/v3, LayoutXLM, MarkupLM, XDoc, DiT, TrOCR, BEiT, VLMo, or document OCR/layout tasks unless the request specifically names Kosmos-2.5 literate OCR or image-to-markdown.
  • Use embeddings-retrieval for E5, SimLM, dense retrieval, reranking, hard-negative mining, or text embedding evaluation.
  • Use architectures-training for architecture-only training mechanics, long-context models, optimizer schedules, TorchScale/DeepNet/Magneto internals, or Deepspeed/Apex mechanics not tied to these multimodal workflows.

Read Or Run

  • Read references/workflows.md to select the right model family and construct safe command templates for Kosmos, TextDiffuser, audio/speech, or LatentLM tasks.
  • Read references/input-checklists.md before any native run to confirm required images, prompts, checkpoints, audio manifests, datasets, and outputs.
  • Read references/troubleshooting.md when checkpoints, tokenizers, CUDA/FlashAttention/xformers, fairseq, Hugging Face, audio sample rates, or LatentLM FID inputs fail.
  • Run scripts/check_multimodal_inputs.py to validate mode-specific local arguments and print a safe run checklist without importing heavy frameworks, downloading, or launching generation.

Safe Default Workflow

  1. Classify the request by family: Kosmos grounding/OCR, TextDiffuser text rendering, audio/speech representation or generation, or LatentLM latent diffusion.
  2. Use the references to pick the native entry-point shape and required inputs; keep original inference/training scripts reference-only until the user confirms data, checkpoints, dependencies, network access, GPU readiness, and output location.
  3. Run the bundled checker in the closest mode to catch missing local files and unsafe assumptions before constructing a native command.
  4. If the checker passes, present the command as a template to run from a user-provided compatible source checkout and environment, not from this skill directory.
  5. Prefer small validation or --help checks before long generation, distributed training, downloads, remote demos, or metric evaluation.

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