Developer technical notes — durable, AI-accessible, and open to the OSS community.
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.
- 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
| 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 |
- Pick or create a topic directory under
content/— each folder automatically becomes a topic on the site - Add an
_index.mdwith atitleanddescriptionif creating a new topic - Write a Markdown file with a date-prefixed filename:
YYYY-MM-DD-title.md - Add images under
content/<topic>/images/ - Open a PR
See the post template for the recommended front matter and structure.