[Accelerate notebook_launcher] "CUDA has been initialized before the `notebook_launcher` could create a forked subprocess"

Cause (Documented platform behavior): Forked workers cannot re-initialize CUDA once the parent has initialized it. Fix status: documented_behavior Other error fragments: - the `Accelerator` should only be initialized inside your training function. Restart your notebook and make sure no cells initializes an `Accelerator` Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/huggingface/accelerate/421af8bee5bda705d3c763f491b7a794f471818e/src/accelerate/launchers.py (official_docs, unknown, documented_behavior): notebook_launcher raises for CUDA initialised before fork and for pre-created Accelerator. - https://raw.githubusercontent.com/huggingface/accelerate/421af8bee5bda705d3c763f491b7a794f471818e/docs/source/basic_tutorials/notebook.md (official_docs, unknown, documented_behavior): Notebook tutorial for launching multi-GPU training. Search phrasings: notebook_launcher CUDA has been initialized before forked subprocess; accelerate notebook multi gpu Cannot re-initialize CUDA Evidence basis (self-declared by the contributing chat client): public_source.