Pytorch
PyTorch deep learning framework with dynamic graphs. Use for neural networks.From its SKILL.md
npx -y skills add G1Joshi/Agent-Skills --skill pytorchAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 12 stars12 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
1.2 KB, 287 tokens by cl100k_base, as published. Nobody here has run it
PyTorch
PyTorch is the dominant framework for research and production AI. v2.5 (2025) solidifies torch.compile and introduces FlexAttention.
When to Use
- Research: 99% of new papers (Arxiv) use PyTorch.
- Production: Recommended for almost all new DL projects.
- Performance:
torch.compileprovides C++ level speed with Python ease.
Core Concepts
torch.compile
Just-in-Time (JIT) compilation of your model.
model = torch.compile(model) -> 2x speedup.
Dynamic Graphs (Eager Mode)
Debug line-by-line (print(tensor.shape) works).
Fabric / Lightning
High-level wrappers to simplify training loops and multi-GPU setup.
Best Practices (2025)
Do:
- Use
torch.compile: It is now stable and essential for H100 performance. - Use
FlashAttention: Use the scaled dot product attention (SDPA) kernel for Transformers. - Use PyTorch 2.x: PyTorch 1.x is legacy.
Don't:
- Don't code
.cuda()manually: Use.to(device)or Fabric to handle device placement.
References
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.