Coverage for src / quber / __init__.py: 100%

1""" 

2Quber - Advanced document parsing platform. 

3 

4Two engine-agnostic seams: 

5 

6- `Parser` Protocol (`quber.core.parsers`) — produces a full DoclingDocument. 

7- `TableExtractor` Protocol (`quber.core.extractors`) — produces a list of 

8 ExtractedTable Pydantic models from table-only engines like Camelot+LLM. 

9 

10Composition of the two is intentionally deferred (see plan §11 Deferred). 

11""" 

12 

13__version__ = "0.3.0" 

14 

15# Two import-time initializers, each a named function (no inline side effects): 

16# - preload_cuda_runtime_libs (QUE-219): preload CUDA 12 runtime libs so 

17# onnxruntime-gpu's CUDAExecutionProvider can resolve them on CUDA-13 hosts. 

18# Must run before any docling/onnxruntime import; silent on CPU-only hosts. 

19# - init_s3_cache (QUE-223): register the persistent, TTL-swept S3 client as 

20# cloudpathlib's default for s3:// URIs. Offline-safe (boto3 resolves creds 

21# and network lazily), so it is fine to run at import on any host. 

22from quber.files.cache import init_s3_cache 

23from quber.utils.cuda_runtime import preload_cuda_runtime_libs 

24 

25preload_cuda_runtime_libs() 

26init_s3_cache() 

27 

28# No eager re-exports here. Importing the docling-backed parsers or the 

29# inference processor at package import would pull the heavy GPU/ML stack into 

30# every `import quber` -- including the CPU-only `table` cloud-job image, which 

31# imports none of it. Import them from their modules at the point of use, e.g. 

32# `from quber.core.parsers import parser_for_preset` or 

33# `from quber.processors import TableInferenceProcessor`.