gpu-memory-cleanup

Jupyter cells hang after GPU processing because CuPy's memory pool retains allocated GPU memory even after computation completes. Explicitly freeing the memory pool prevents the kernel from blocking indefinitely and allows cells to finish cleanly.

Call CuPy's memory cleanup methods—cp.get_default_memory_pool().free_all_blocks() and cp.get_default_pinned_memory_pool().free_all_blocks()—followed by Python garbage collection to release GPU memory after processing loops, deconvolution, or multi-GPU pipelines in Jupyter cells.

Yes. This approach applies specifically to CuPy-based GPU workflows in Jupyter notebooks and interactive environments, with graceful handling when CuPy is unavailable, making it safe to add to existing cells without breaking execution.

Add explicit CuPy memory pool cleanup to your processing cell to prevent accumulation across iterative tasks like BaSiC illumination corrections or EDF processing, ensuring responsive kernel behavior and avoiding cell hangs.

No. The same CuPy memory cleanup pattern—calling free_all_blocks() on both default and pinned memory pools—works uniformly across single-GPU and multi-GPU pipelines, adapting to your processing scale automatically.