fix: enable multi-GPU DDP training in Jupyter notebooks#928

Open

mfazrinizar wants to merge 15 commits intoroboflow:developfrom

mfazrinizar:fix/ddp-notebook-cuda-init

Commits on Apr 5, 2026

fix: defer CUDA init to enable DDP training in notebooks
mfazrinizar
committed

Commits on Apr 6, 2026

fix: skip CUDA bf16 probe for ddp_notebook strategy
mfazrinizar
committed
fix: eliminate all CUDA driver context leaks before DDP fork
mfazrinizar
committed
fix: use overridden num_workers in all dataloaders for ddp_notebook
mfazrinizar
committed
fix: possible thread-state corruption from fork()
mfazrinizar
committed
revert: remove torch.set_num_threads that crashes forked DDP children
mfazrinizar
committed
fix: use spawn-based DDP for ddp_notebook to avoid OpenMP SIGABRT
mfazrinizar
committed
fix: adding logger for ddp_notebook strategy
mfazrinizar
committed
fix: use spawn-based DDP for ddp_notebook to avoid OpenMP SIGABRT
mfazrinizar
committed
fix: remove unnecessary num_workers=0 override for ddp_notebook
mfazrinizar
committed
fix: use standard precision probing for DDP and guard auto-batch
mfazrinizar
committed
Merge branch 'develop' into fix/ddp-notebook-cuda-init
mfazrinizar
authored
fix(pre-commit): 🎨 auto format pre-commit hooks
pre-commit-ci[bot]
committed
style: fix ruff E402 imports and codespell in DDP tests
mfazrinizar
committed

Commits on Apr 7, 2026

fix: handle None from torch.accelerator on CPU-only environments
mfazrinizar
committed