PyTorch DevLog

Developer technical notes — durable, AI-accessible, and open to the OSS community.

Browse the devlog →

PyTorch developers regularly produce deep technical content about design decisions, performance analyses, and feature implementations. This repo makes that knowledge permanent, searchable, machine-indexable, and available to OSS users and AI agents — so it doesn't get buried in chat threads or ephemeral posts.

Why

• AI-accessible: LLM tools that index PyTorch automatically gain access to this technical context
• Open to OSS: Contributors get visibility into internal technical discussions
• Durable: Versioned Markdown files remain discoverable indefinitely

Topics

Directory	Scope
dynamic_shapes/	Unbacked shapes, guards, symbol semantics
dispatcher/	Dispatch keys, operator registry, extensibility
dynamo/	TorchDynamo graph capture
inductor/	TorchInductor codegen
distributed/	FSDP, DTensor, c10d
export/	torch.export, AOTInductor

Contributing

1. Pick or create a topic directory under content/ — each folder automatically becomes a topic on the site
2. Add an _index.md with a title and description if creating a new topic
3. Write a Markdown file with a date-prefixed filename: YYYY-MM-DD-title.md
4. Add images under content/<topic>/images/
5. Open a PR

See the post template for the recommended front matter and structure.