test_richdocument errors on Apple Silicon: docling layout pipeline fails on MPS (float64 unsupported) #1344

Problem

The four tests in test/stdlib/components/docs/test_richdocument.py that depend on the rd module fixture all error out on Apple Silicon with:

ERROR docling.pipeline.standard_pdf_pipeline:standard_pdf_pipeline.py:321
Stage layout failed run 1: Cannot convert a MPS Tensor to float64 dtype as float32

MPS does not support float64. Docling's layout pipeline automatically selects MPS on Apple Silicon and then attempts a float64 operation, causing the fixture to raise before any test body runs.

Affected tests

All four tests that share the rd fixture:

• test_richdocument_basics
• test_richdocument_markdown
• test_richdocument_save
• test_table

Why they aren't skipped

The rd fixture wraps RichDocument.from_document_file() in hf_skip(), which only catches HuggingFace Hub errors. The MPS tensor error originates inside docling's internal pipeline and propagates as an unhandled ERROR, taking all four downstream tests with it. There is no require_gpu, no MPS xfail, and no device override.

Options

A. Force CPU in the fixture on MPS — pass a docling pipeline config that pins the device to CPU when torch.backends.mps.is_available(). The tests would run correctly and slowly; no golden files needed.

B. xfail on MPS — add if torch.backends.mps.is_available(): pytest.xfail("docling layout pipeline does not support MPS float64") at the top of the fixture. Low effort, honest, and unblocks MPS developers until docling gains float64 MPS support.

C. require_gpu predicate — skip on non-CUDA hardware. This gives up on MPS support rather than fixing it, and is probably too broad since the tests run fine on CPU.

Option A is the cleanest long-term fix. Option B is a safe interim measure.

Related

• test_run_transformers fails on MPS due to CUDA-specific golden files; resource predicates missing #1342 — test_run_transformers MPS golden file mismatch (same theme, different test and cause)