Replace use_cuda with device parameter in AnnLoader

[LiudengZhang]

Summary

• Replace use_cuda: bool with device: str = "cpu" in default_converter and AnnLoader.__init__, addressing the TODO comment in the code (# maybe replace use_cuda with explicit device option).
• use_cuda is kept for backward compatibility but emits a FutureWarning when explicitly passed. Passing both use_cuda and device raises a ValueError.
• pin_memory logic is now correctly guarded to only apply on CPU (it is a CPU-only operation).

Motivation

The existing use_cuda boolean is limiting — it only supports CPU or CUDA. The new device parameter accepts any PyTorch device string (e.g., "cpu", "cuda", "cuda:1", "mps"), making the API more flexible and consistent with PyTorch conventions.

Test plan

• test_annloader_default_device — default device="cpu" produces CPU tensors
• test_annloader_explicit_cpu_device — explicit device="cpu" works
• test_annloader_use_cuda_deprecation_warning — use_cuda emits FutureWarning
• test_annloader_use_cuda_and_device_conflict — both params raises ValueError

[codecov]

Codecov Report

❌ Patch coverage is 21.42857% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.35%. Comparing base (a0c4283) to head (a520fbc).
⚠️ Report is 27 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/anndata/experimental/pytorch/_annloader.py	21.42%	11 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2360      +/-   ##
==========================================
- Coverage   87.36%   85.35%   -2.01%     
==========================================
  Files          49       49              
  Lines        7785     7790       +5     
==========================================
- Hits         6801     6649     -152     
- Misses        984     1141     +157     

Files with missing lines	Coverage Δ
src/anndata/experimental/pytorch/_annloader.py	23.14% <21.42%> (-0.16%)	⬇️

... and 8 files with indirect coverage changes

[flying-sheep]

thanks! Some small notes from me

[flying-sheep]

we have a standard Empty.TOKEN object for this:

anndata/src/anndata/compat/__init__.py

Lines 101 to 102 in a0c4283

	class Empty(Enum):
	TOKEN = auto()

[LiudengZhang]

Good call — switched to Empty.TOKEN in 6f01e74.

[flying-sheep]

please use a Literal type enumerating all valid values. (maybe imported from torch if it exists there)

[LiudengZhang]

Done — used Literal["cpu", "cuda", "mps"] since torch doesn't export a device string literal. Fixed in 6f01e74.

[LiudengZhang]

Hi @flying-sheep, I've addressed both points (switched to Empty.TOKEN and used Literal type) in 6f01e74. Let me know if there's anything else!