---

name: inference-and-rendering description: "Load DeTikZify models and adapters, build inference pipelines, generate TikZ, compile and rasterize documents, and troubleshoot programmatic synthesis workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Inference and Rendering

Use this sub-skill when the task is about DeTikZify's Python inference surface: model loading, processor handling, adapter-backed text conditioning, DetikzifyPipeline, DetikzifyGenerator, TikzDocument, compile/rasterize/save, or MCTS-backed sampling.

Route away from this sub-skill when the main task is the browser UI, training loops, evaluation metrics, or dataset builders. Those workflows have dedicated sub-skills.

Fast Path

1. Confirm the package imports and the public API surface is present:

   python scripts/api_smoke.py
   

2. For a rendering-specific smoke, compile a tiny document through the bundled helper:

   python scripts/tikz_smoke.py
   

3. If you need search-based synthesis or tree-search behavior, run the MCTS helper:

   python scripts/mcts_smoke.py
   

What This Sub-Skill Owns

• detikzify.model.load(...) and detikzify.model.load_adapter(...)
• DetikzifyProcessor and AdapterProcessor
• DetikzifyPipeline.sample(...) and DetikzifyPipeline.simulate(...)
• DetikzifyGenerator and the MCTS bridge used inside the pipeline
• TikzDocument.compile(), rasterize(), save(), and error inspection
• image loading / preprocessing helpers that feed inference or rendering
• text-conditioned generation when an adapter is loaded

Common Decisions

• Use load(model_name_or_path, ...) for image-conditioned generation with the default processor flow.
• Use load_adapter(...) or a model that already carries an adapter when the request needs text prompts.
• Use sample(...) for a single best TikZ program and simulate(...) when the user wants search / multiple candidates.
• Treat metric="model" as the perceptual path and metric="fast" as the compiler-diagnostics path.
• Set preprocess=True when you want the image trimmed and padded to a square before generation.
• Check compile_timeout and mcts_timeout separately; they control different stages.

Bundled References

• references/api-reference.md: signatures and object roles for the public inference surface.
• references/workflows.md: image-conditioned, text-conditioned, save/rasterize, and search-based workflows.
• references/troubleshooting.md: adapter, CUDA, image-source, and compile/rasterize failure modes.

Related Helpers

• ../../scripts/api_smoke.py: safe import and signature snapshot.
• ../../scripts/tikz_smoke.py: safe TeX compile/rasterize smoke.
• ../../scripts/mcts_smoke.py: safe dummy-state MCTS sanity check.

Guardrails

• A CPU import does not prove the GPU path works.
• Do not assume text prompts work unless an adapter-capable load path is active.
• Do not treat compile success as proof that the output is non-empty; check the rasterized result too.