aps-ai-beamline-controller 0.0.71

AI-driven Beamline Controller

AI-Beamline-Controller

Ai-driven autonomous beamline controller

pip install aps-ai-beamline-controller detects the installing machine's NVIDIA driver and pulls the matching torch / torchvision / torchaudio build, plus ax-platform. No --index-url needed by the user.

How it works

Dependencies are declared dynamic (PEP 643), so they are computed when the user's pip builds the wheel locally from the sdist. At that point setup.py:

1. runs nvidia-smi to read the host's max CUDA version,
2. maps it to a (torch release, CUDA build) tier,
3. queries the PyTorch index and finds the exact wheel URLs matching the running interpreter's tags,
4. emits them as PEP 508 direct-reference requirements (torch @ https://...).

pip then downloads those exact wheels.

Tier map (edit TIERS in setup.py to change):

driver CUDA	torch	build
≥ 13.0	2.12.1	cu130
12.8–12.9	2.11.0	cu128
12.6–12.7	2.12.1	cu126
12.4–12.5	2.6.0	cu124
11.8–12.3	2.6.0	cu118
none/older	2.12.1	cpu

Publishing — sdist ONLY

This is the load-bearing rule. If a wheel is on PyPI, pip installs it and the detection never runs. Build and upload only the sdist:

python -m build --sdist        # produces dist/*.tar.gz only
twine upload dist/*.tar.gz     # do NOT upload any .whl

The uploaded sdist's metadata says dynamic = ["dependencies"] (no direct URLs), so PyPI accepts it. The direct URLs exist only in the wheel each user builds locally, which PyPI never sees.