---

name: runtime-and-models description: "Runtime, dependency, CLIP model, cache, backend, and inspection guidance for deep-daze." disable-model-invocation: true metadata: disco-role: operating license: MIT

runtime-and-models

Use this sub-skill when the task is to validate or troubleshoot a deep-daze runtime before any image generation, especially around installation identity, imports, package data, CLIP model names, tokenizer behavior, checkpoint cache behavior, and CPU/GPU availability.

Owns

• Distribution/import identity: package distribution deep-daze, import module deep_daze, exports DeepDaze and Imagine, and console script imagine.
• Dependency expectations for import, tokenization, CLI startup, CLIP loading, and generation-time optimizer/network components.
• CLIP model registry and tokenizer facts, including default model ViT-B/32, model names, token context length, and bundled BPE vocabulary.
• Runtime device behavior: CUDA when available, CPU fallback otherwise, and the practical limits of CPU-only generation.
• Checkpoint cache/download/checksum behavior and safe preflight inspection.
• Cross-cutting runtime troubleshooting before handing off to workflow-specific CLI or Python API instructions.

Start here

1. For a safe environment preflight that does not download CLIP checkpoints or generate images, run:

   python scripts/check_deep_daze_runtime.py
   

   Use python scripts/check_deep_daze_runtime.py --help for options.

2. Read references/runtime-reference.md for runtime facts, model names, cache semantics, and backend expectations.

3. Read references/troubleshooting.md for dependency, BPE vocabulary, checkpoint, network, JIT, CPU/GPU, and resource failure modes.

Route elsewhere

• For prompt-to-image command-line recipes, flags, and generation workflows, use ../cli-workflows/SKILL.md.
• For programmatic Imagine/DeepDaze construction and Python generation recipes, use ../python-api/SKILL.md.

Do not use this as a generation smoke test

This sub-skill intentionally avoids Imagine(...), deep_daze.clip.load(...), and imagine ... generation commands. Those actions can download large CLIP checkpoints, perform expensive optimization, write output files, and require interactive or display-related behavior. Use this sub-skill only to decide whether the runtime is plausibly ready and what must be fixed before generation.