hf-jobs-workflow

What problem does it solve?

This Skill provides a comprehensive operational framework for running training and fine-tuning jobs on Hugging Face Jobs, mitigating common failure modes and ensuring efficient, effective training runs.

Core Features & Use Cases

• Pre-Flight Checklist: A detailed checklist to ensure all prerequisites are met before submitting a job.
• GPU Hardware Sizing: Guidelines for selecting appropriate GPU hardware based on model size and training requirements.
• OOM Recovery: Strategies for handling out-of-memory errors without changing the training task.
• Sandbox-First Development: Encourages testing in a CPU/GPU sandbox to catch potential issues early.
• push_to_hub Discipline: Ensures that all job outputs are durable and retrievable.
• Training Logging: Best practices for logging that survives the logs tab.
• Submit-One-Then-Batch: Technique for submitting hyperparameter sweep jobs to avoid early failures.
• Dataset Format by Training Method: Guidance on dataset format requirements for different training methods.
• Prebuilt Kernels: Recommendations for using prebuilt attention kernels instead of compiling from source.
• Decision Table: When to use Hugging Face Jobs versus alternative platforms.

Quick Start

Use the hf-jobs-workflow skill to set up a training job on Hugging Face Jobs, ensuring all best practices are followed.