Draft PR: Minify LLM-Guard & Lazy-Load Heavy Dependencies

[JuanFKurucz]

This PR is work-in-progress and intentionally opened as a draft as an idea/necessity on Issue #250 “ONNX Runtime minimal dependencies.”

---

TL;DR

Regarding #250, this change lets users install ONNX-only builds of llm-guard while preserving full functionality for PyTorch users.

1. Optional-extras matrix (onnxruntime, onnxruntime-gpu, pytorch, extended, full).
2. Lazy imports for heavy libs (torch, transformers, …).
3. Smarter device/provider detection that doesn’t require PyTorch.
4. No breaking API changes; legacy imports still work.

---

✨ Key changes

Area	Details
Dependencies	pyproject.toml splits heavy libs into extras; base wheel is slim.
Scanners	New get_scanner_class() + refactored get_scanner_by_name() use lazy imports.
util.device()	No longer hard-imports torch; falls back to env/OS inspection when absent.
Model helpers	ONNX provider chosen via onnxruntime.get_available_providers().
Back-compat	Public re-exports unchanged.

---

🔍 Motivation

• Smaller memory footprint
• Faster cold-starts
• User choice — install only what you need.

---

🛠️ Install matrix

Scenario	Command
ONNX / CPU	pip install llm_guard[onnxruntime]
ONNX / GPU	pip install llm_guard[onnxruntime-gpu]
Full PyTorch	pip install llm_guard[pytorch]
Everything	pip install llm_guard[full]

---

⚠️ Breaking changes?

By default now it will install the minimal dependencies

---

📌 Next steps before “Ready for Review”

• Revisit if this breaks something, I have only tested with CompetitorsCheck

---

References

• Issue ONNX Runtime minimal dependencies #250 – “ONNX Runtime minimal dependencies” (opened by @JuanFKurucz)

---

Thanks for reviewing! Feedback is welcome before moving this draft to a full PR.