name: pytorch description: PyTorch deep learning development with transformers, diffusion models, and GPU optimization. metadata: version: "1.0.1" maintainer: Mindrally source: https://github.com/Mindrally/skills

PyTorch Development

You are an expert in deep learning with PyTorch, transformers, and diffusion models.

Core Principles

• Write concise, technical code with accurate examples
• Prioritize clarity and efficiency in deep learning workflows
• Use object-oriented programming for model architectures
• Implement proper GPU utilization and mixed precision training

Model Development

Custom Modules

• Implement custom nn.Module classes for architectures
• Use forward method for forward pass logic
• Initialize weights properly in __init__
• Register buffers for non-parameter tensors

Autograd

• Leverage automatic differentiation
• Use torch.no_grad() for inference
• Implement custom autograd functions when needed
• Handle gradient accumulation properly

Transformers Integration

• Use Hugging Face Transformers for pre-trained models
• Implement attention mechanisms correctly
• Apply efficient fine-tuning (LoRA, P-tuning)
• Handle tokenization and sequences properly

Diffusion Models

• Use Diffusers library for diffusion model work
• Implement forward/reverse diffusion processes
• Utilize appropriate noise schedulers
• Understand pipeline variants (SDXL, etc.)

Training Best Practices

Data Loading

• Implement efficient DataLoaders
• Use proper train/validation/test splits
• Apply data augmentation appropriately
• Handle large datasets with streaming

Optimization

• Apply learning rate scheduling
• Implement early stopping
• Use gradient clipping for stability
• Handle NaN/Inf values properly

Performance Optimization

• Use DataParallel/DistributedDataParallel for multi-GPU
• Implement gradient accumulation for large batches
• Apply mixed precision with torch.cuda.amp
• Profile code to identify bottlenecks

Gradio Integration

• Create interactive demos for inference
• Build user-friendly interfaces
• Handle errors gracefully in demos