[karthiksenv]

Concurrent inference on one Llama instance shares mutable batch/context state and can abort in native code, as reported in #2392. The README currently does not explain this limitation.

Add a shared-instance locking example for embeddings and a streaming example that holds the lock through iteration and closes the generator before releasing it. Link the section from Embeddings and note the synchronous-worker constraint. This addresses the documentation part of #2392; it does not add internal locking or parallel inference.

Validation:

• Ruff check and format checks passed for llama_cpp and tests (29 files formatted).
• Both new Python examples passed AST parsing; git diff --check passed.
• python -m mkdocs build --clean passed, and the generated heading/navigation/embeddings links were checked. Existing relative-link, docstring and theme-deprecation warnings remain.
• Local docs environment used pymdown-extensions 11.0.2 because 12.3 caused a mkdocstrings Highlighter API mismatch. No dependency declarations changed.

Native model tests and independent reproduction of the reported crash were not run. Source inspection confirms shared _batch/context mutation and the generator returned by create_completion(stream=True). Keeping this draft for review of the documentation scope and remaining validation.

AI-assisted: Codex prepared the change and ran the checks above; the contributor authorized publication of this draft.