Windows wheels resolve unpinned torch dependency to CPU-only build: 'cannot use CUDA device 0: only 0 device(s) visible' #607

Summary

The Windows wheels don't constrain their torch dependency to the CUDA build, so a fresh install (or any --reinstall) that resolves torch from PyPI silently yields CPU-only torch, and the engine then fails with:

Backend supervisor: RuntimeError: cannot use CUDA device 0: only 0 device(s) visible (CUDA_VISIBLE_DEVICES=None)

Repro

uv pip install --python <ft venv python> --reinstall \
  https://github.com/FlashML-org/FreeToken-Web/releases/download/beta/freetoken-0.1.3%2Bgfa0e9537d-cp312-cp312-win_amd64.whl \
  https://github.com/FlashML-org/FreeToken-Web/releases/download/beta/freetoken_kernel_cache-0.1.3%2Bcu130.gfa0e9537d-py3-none-win_amd64.whl

uv resolves the wheel's Requires-Dist: torch (unpinned) from PyPI:

- torch==2.11.0+cu130
+ torch==2.11.0            # CPU build, no CUDA
+ torchvision==0.26.0      # was 0.26.0+cu130

torch.cuda.is_available() → False, torch.version.cuda → None. Engine can't see the GPU at all. The engine itself is fine — re-installing the CUDA build from the PyTorch index restores everything:

uv pip install --python <ft venv python> --reinstall-package torch --reinstall-package torchvision \
  "torch==2.11.0+cu130" "torchvision==0.26.0+cu130" --index-url https://download.pytorch.org/whl/cu130

Suggested fixes (any one would do)

1. Pin the CUDA build in the wheel metadata: torch==2.11.0+cu130 — pip/uv can fetch it given the documented extra index, and existing installs keep resolving correctly.
2. Or declare torch>=2.11,<2.12 plus a runtime self-check at ft serve startup: if torch.version.cuda is None, fail with "CPU-only torch installed — see install docs" instead of the cryptic "0 device(s) visible".
3. Or document the exact torch requirement next to the wheel URLs.

Seen on Windows 11, Python 3.12.10, uv 0.9.x, engine 0.1.3+gfa0e9537d.