[MPS] compat: torch.accelerator module exists but is not exported at top level

[lvyufeng]

Summary

A recent upstream change added src/candle/accelerator.py, but the beginner tensor tutorial still fails because Candle does not expose accelerator on the top-level torch/candle module.

So the issue is no longer 'module missing on disk' — it is now specifically a top-level export / compat surface wiring gap.

Current state

Upstream now contains:

• src/candle/accelerator.py

But src/candle/__init__.py still does not expose accelerator, so this still fails:

import torch
print(torch.accelerator.is_available())

when USE_CANDLE=1 redirects torch to Candle.

Environment

• Repo: candle-org/candle
• Tested from latest upstream/main on 2026-03-27
• Commit tested: f37302f
• macOS Apple Silicon
• Python 3.11.15
• Candle installed editable from a fresh worktree
• Torch compat hook installed via python -m candle.torch_compat install
• Repro run with USE_CANDLE=1

Minimal reproduction

source ~/miniconda3/etc/profile.d/conda.sh && conda run -n candle311-tutorials \
  env USE_CANDLE=1 python -c "import torch; print(torch.accelerator.is_available())"

Observed

Traceback (most recent call last):
  File "<string>", line 1, in <module>
AttributeError: module 'candle' has no attribute 'accelerator'

Expected

If src/candle/accelerator.py is the intended compatibility implementation, it should be reachable through the top-level compat surface so that:

torch.accelerator.is_available()
torch.accelerator.current_accelerator()

work under USE_CANDLE=1.

Tutorial impact

This still blocks:

• beginner/basics/tensorqs_tutorial.py

Likely source

• src/candle/accelerator.py now exists
• src/candle/__init__.py does not re-export or otherwise attach accelerator

Related

• supersedes the originally closed gap report: [MPS] compat: torch.accelerator is missing and breaks basics tensor tutorial #272