the yaml thing is not a switch statement though. torchgen takes native_functions.yaml and generates the bindings, autograd nodes and dispatch entries from that one file.
there are ~~3700 ops. try keeping that in sync by hand.
also dynamo never calls gcc. it just captures the graph from python bytecode. inductor is what writes triton/cpp and sends it to real compilers. they built their own compiler stack once (torchscript), it's dead now.
so this time they reused what exists.
spaghetti yes. 10 years of layers. you can see every era in the repo.
Thanks for the PyTorch internals OG link and for putting together your PyTorch one-pager, if we can call it that :)
Kudos.
Off-topic:
Some of your comments in this thread are dead, but I’ve vouched for some them as I don’t see anything wrong with what you’ve written, other than your writing here on HN sounds a little like you are using an LLM to compose each reply (or perhaps your regular use of LLMs is showing through in your comments?).
Not bad, but Claude is not quite good yet for writing these explainers. We should try again in a few months with newer models.
Can tell it's garbage just from the first figures caption
Lol, a yaml file for function signatures of which to generate a giant switch statement.
The "compiler" dynamo part is also a hack that calls real compilers like gcc and Triton.
It is a big bowl of spaghetti.
the yaml thing is not a switch statement though. torchgen takes native_functions.yaml and generates the bindings, autograd nodes and dispatch entries from that one file. there are ~~3700 ops. try keeping that in sync by hand.
also dynamo never calls gcc. it just captures the graph from python bytecode. inductor is what writes triton/cpp and sends it to real compilers. they built their own compiler stack once (torchscript), it's dead now. so this time they reused what exists.
spaghetti yes. 10 years of layers. you can see every era in the repo.
can someone give recommendations of other pytorch explanations like this one?
author here. honest answer: there isn't much that covers the whole thing end to end, that's partly why I started writing this.
but the ones I learned from: - Edward Yang's "pytorch internals" post: http://blog.ezyang.com/2019/05/pytorch-internals/ it's the classic, goes deep on the c++ side.
- the pytorch developer podcast, same author. short episodes, one internal topic each
- the pytorch repo's own CONTRIBUTING.md, the folder layout explained by the maintainers
- and if you want to build autograd instead of reading about it, Karpathy's micrograd / zero to hero
mine is planned as 12 parts, the list is at the end of the post. part 1 (chapter Tensor: storage, strides, views) is in progress now.
Thanks for the PyTorch internals OG link and for putting together your PyTorch one-pager, if we can call it that :)
Kudos.
Off-topic:
Some of your comments in this thread are dead, but I’ve vouched for some them as I don’t see anything wrong with what you’ve written, other than your writing here on HN sounds a little like you are using an LLM to compose each reply (or perhaps your regular use of LLMs is showing through in your comments?).
thanks for the vouch, appreciated! :)
and honest answer, English is not my first language, so yes, I rely on co-pilot tools when I write and it probably shows.
the PyTorch content itself I can defend line by line, that part is mine.
If you like these sort of AI generated explainers this one is pretty good too: https://www.longformexplainers.com/llm-pretraining
Solid write-up, thanks again