Why can't I see progress logs in my notebook during parallel runs?

Jupyter typically suppresses stdout from worker threads. This skill uses the 'as_completed' pattern to return results to the main thread, ensuring logs are visible in the notebook interface.

Why does my Jupyter kernel crash during parallel GPU tasks?

Crashes are often caused by GPU thread explosion where multiple processes attempt to access the same GPU simultaneously, leading to Out of Memory (OOM) errors. This skill uses a queue system to ensure tasks are mapped 1:1 to available GPUs.

How does this skill handle memory leaks between tasks?

It implements a verified cleanup pattern that clears CuPy's default and pinned memory pools and triggers Python's garbage collector immediately after a task completes, before the GPU is released for the next task.

Can I use this for multi-GPU setups?

Yes, it is specifically designed for multi-GPU environments, allowing you to define a pool of GPU IDs and dynamically allocate them to a parallel task queue.

GPU Parallel Scheduling & Memory Safety

Name: GPU Parallel Scheduling & Memory Safety
Author: smith6jt-cop

bysmith6jt-cop

0•

Data Science & ML

Implements queue-based GPU allocation and memory cleanup patterns to prevent OOM crashes and ensure reliable progress tracking in parallel workflows.

This skill provides specialized patterns for managing multi-GPU parallel processing, specifically optimized for high-performance computing (HPC) environments using CuPy and Jupyter notebooks. It addresses critical issues like GPU thread explosion, nested executor crashes, and Jupyter's suppression of worker thread output by enforcing a strict 'one task per GPU' rule through dynamic queue-based allocation. By integrating automated memory cleanup and main-thread progress reporting, it ensures stable execution of complex data pipelines without the common pitfalls of cumulative memory pressure or silent kernel failures.

Key Features

01Jupyter-safe progress output via main-thread result synchronization

02Automated GPU memory cleanup patterns for CuPy and garbage collection

03Prevention strategies for nested ThreadPoolExecutor 'thread explosions'

04Queue-based GPU allocation for 1:1 resource-to-task mapping

05Dynamic GPU device assignment and release workflows

060 GitHub stars

Use Cases

01Debugging and fixing silent kernel crashes in parallel HPC workflows

02Building OOM-resilient CuPy pipelines for large-scale data analysis and scientific computing

03Processing multiple image channels in parallel across multiple GPUs in Jupyter notebooks

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add smith6jt-cop/skills_registry gpu-parallel-scheduling

For use in Claude.ai and ChatGPT

Download Skill