The PyTorch Developer Podcast is a place for the PyTorch dev team to do bite sized (10-20 min) topics about all sorts of internal development topics in PyTorch.
The podcast PyTorch Developer Podcast is created by Edward Yang, Team PyTorch. The podcast and the artwork on this page are embedded on this page using the public podcast feed (RSS).
Compiler collectives are a PT2 feature where by compiler instances across multiple ranks use NCCL collectives to communicate information to other instances. This is used to ensure we consistently decide if inputs or static or dynamic across all ranks. See also PR at https://github.com/pytorch/pytorch/pull/130935
Inductor IR is an intermediate representation that lives between ATen FX graphs and the final Triton code generated by Inductor. It was designed to faithfully represent PyTorch semantics and accordingly models views, mutation and striding. When you write a lowering from ATen operators to Inductor IR, you get a TensorBox for each Tensor argument which contains a reference to the underlying IR (via StorageBox, and then a Buffer/ComputedBuffer) that says how the Tensor was computed. The inner computation is represented via define-by-run, which allows for compact definition of IR representation, while still allowing you to extract an FX graph out if you desire. Scheduling then takes buffers of inductor IR and decides what can be fused. Inductor IR may have too many nodes, this would be a good thing to refactor in the future.
I talk about VariableTracker in Dynamo. VariableTracker is Dynamo's representation of the Python. I talk about some recent changes, namely eager guards and mutable VT. I also tell you how to find the functionality you care about in VariableTracker (https://docs.google.com/document/d/1XDPNK3iNNShg07jRXDOrMk2V_i66u1hEbPltcsxE-3E/edit#heading=h.i6v7gqw5byv6).
This podcast goes over the basics of unbacked SymInts. You might want to listen to this one before listening to https://pytorch-dev-podcast.simplecast.com/episodes/zero-one-specialization Some questions we answer (h/t from Gregory Chanan):
- Are unbacked symints only for export? Because otherwise I could just break / wait for the actual size. But maybe I can save some retracing / graph breaks perf if I have them too? So the correct statement is "primarily" for export?
- Why am I looking into the broadcasting code at all? Naively, I would expect the export graph to be just a list of ATen ops strung together. Why do I recurse that far down? Why can't I annotate DONT_TRACE_ME_BRO?
- How does 0/1 specialization fit into this? I understand we may want to 0/1 specialize in a dynamic shape regime in "eager" mode (is there a better term?), but that doesn't seem to matter for export?
- So far we've mainly been talking about how to handle our own library code. There is a worry about pushing complicated constraints downstream, similar to torchscript. What constraints does this actually push?
Mikey Dagistes joins me to ask some questions about the recent recent composability sync https://www.youtube.com/watch?v=NJV7YFbtoR4 where we discussed 0/1 specialization and its implications on export in PT2. What's the fuss all about? What do I need to understand about PT2 to understand why 0/1 specialization is a thing?
What is torchdynamo? From a bird's eye view, what exactly does it do? What are some important things to know about it? How does it differ from other graph capture mechanisms?
For more reading, check out https://docs.google.com/document/d/13K03JN4gkbr40UMiW4nbZYtsw8NngQwrTRnL3knetGM/edit#
Join me with Richard Zou to talk about the history of functorch. What was the thought process behind the creation of functorch? How did it get started? JAX’s API and model is fairly different from PyTorch’s, how did we validate that it would work in PyTorch? Where did functorch go after the early user studies? Where is it going next?
What’s a learning rate? Why might you want to schedule it? How does the LR scheduler API in PyTorch work? What the heck is up with the formula implementation? Why is everything terrible?
What are they good for? (Caches. Private fields.) C++ side support, how it’s implemented / release resources. Python side support, how it’s implemented. Weak ref tensor hazard due to resurrection. Downsides of weak references in C++. Scott Wolchok’s release resources optimization.
Other episodes to listen to first: https://pytorch-dev-podcast.simplecast.com/episodes/reference-counting https://pytorch-dev-podcast.simplecast.com/episodes/pyobject-preservation
Mike Ruberry has an RFC about stride-agnostic operator semantics (https://github.com/pytorch/pytorch/issues/78050), so let's talk about strides. What are they? How are they used to implement views and memory format? How do you handle them properly when writing kernels? In what sense are strides overspecified, and therefore, not worth slavishly reimplementing in a system like PrimTorch? What does Edward think we should do about them?
My blog post that covers strides along with other topics can be found at http://blog.ezyang.com/2019/05/pytorch-internals/
AOTAutograd is a cool new feature in functorch for capturing both forward and backward traces of PyTorch operators, letting you run them through a compiler and then drop the compiled kernels back into a normal PyTorch eager program. Today, Horace joins me to tell me how it works, what it is good to use for, and what our future plans for it are.
Sherlock recently joined the PyTorch team, having previously worked on ONNX Runtime at Microsoft, and Sherlock’s going to ask me some questions about the dispatcher, and I’m going to answer them. We talked about the history of the dispatcher, how to override dispatching order, multiple dispatch, how to organize various dispatch keys and torch function mode. The companion video is at https://youtu.be/_qB2Ho1O3u4
PyTorch recently moved all of its CI from CircleCI to GitHub Actions. There were a lot of improvements in the process, making my old podcast about CI obsolete! Today, Eli Uriegas joins me to talk about why we moved to GitHub Actions, how the new CI system is put together, and what some cool features about our new CI.
C++ has exceptions, Python has exceptions. But they’re not the same thing! How do exceptions work in CPython, how do we translate exceptions from C++ to Python (hint: it’s different for direct bindings versus pybind11), and what do warnings (which we also translate from C++ to Python) have in common with this infrastructure?
PyTorch’s torch API is the Python API everyone knows and loves, but there’s also another API, the ATen API, which most of PyTorch’s internal subsystems are built on. How to tell them apart? What implications do these have on our graph mode IR design? Also, a plug for PrimTorch, a new set of operators, not designed for eager mode, that is supposed to be even lower level than ATen.
PyTorch is in the business of shipping numerical software that can run fast on your CUDA-enabled NVIDIA GPU, but it turns out there is a lot of heterogeneity in NVIDIA’s physical GPU offering and when it comes to what is fast and what is slow, what specific GPU you have on hand matters quite a bit. Yet there are literally hundreds of distinct NVIDIA GPU models on the market, how do you make sense of the madness? Today, Natalia Gimelshein joins me to talk about everything that’s going on in the NVIDIA GPU market, and what, as a framework developer, you have to care about to make sense of it all.
Further reading.
A lot of recent work going in PyTorch is all about adding new and interesting Tensor subclasses, and this all leads up to the question of, what exactly is OK to make a tensor subclass? One answer to this question comes from an old principle from Barbara Liskov called the Liskov substitution principle, which informally can be stated as S is a subtype of T if anywhere you have T, it can be replaced with S without altering "desirable" properties of this program. In this podcast I'll talk about LSP and how it relates to the design of Tensor subclasses and a hypothetical "abstract Tensor specification" which really doesn't exist but which sort of implicitly exists in the corpus of existing PyTorch programs.
Further reading:
In this episode I talk about reduced precision floating point formats float16 (aka half precision) and bfloat16. I'll discuss what floating point numbers are, how these two formats vary, and some of the practical considerations that arise when you are working with numeric code in PyTorch that also needs to work in reduced precision. Did you know that we do all CUDA computations in float32, even if the source tensors are stored as float16? Now you know!
Further reading.
Today I'm going to talk about a famous issue in PyTorch, DataLoader with num_workers > 0 causes memory leak (https://github.com/pytorch/pytorch/issues/13246). This bug is a good opportunity to talk about DataSet/DataLoader design in PyTorch, fork and copy-on-write memory in Linux and Python reference counting; you have to know about all of these things to understand why this bug occurs, but once you do, it also explains why the workarounds help.
Further reading.
PyTorch operates on its input data in a batched manner, typically processing multiple batches of an input at once (rather than once at a time, as would be the case in typical programming). In this podcast, we talk a little about the implications of batching operations in this way, and then also about how PyTorch's API is structured for batching (hint: poorly) and how Numpy introduced a concept of ufunc/gufuncs to standardize over broadcasting and batching behavior. There is some overlap between this podcast and previous podcasts about TensorIterator and vmap; you may also be interested in those episodes.
Further reading.
Python is a single dispatch OO language, but there are some operations such as binary magic methods which implement a simple form of multiple dispatch. torch_function__ (through its Numpy predecessor __array_function) generalizes this mechanism so that invocations of torch.add with different subclasses work properly. This podcast describes how this mechanism works and how it can be used (in an unconventional way) to build composable subclasses ala JAX in functorch.
Further reading:
Writing multithreading code has always been a pain, and in PyTorch there are buckets and buckets of multithreading related issues you have to be aware about and deal with when writing code that makes use of it. We'll cover how you interface with multithreading in PyTorch, what goes into implementing those interfaces (thread pools!) and also some miscellaneous stuff like TLS, forks and data structure thread safety that is also relevant.
Further reading:
CUDA is asynchronous, CPU is synchronous. Making them play well together can be one of the more thorny and easy to get wrong aspects of the PyTorch API. I talk about why non_blocking is difficult to use correctly, a hypothetical "asynchronous CPU" device which would help smooth over some of the API problems and also why it used to be difficult to implement async CPU (but it's not hard anymore!) At the end, I also briefly talk about how async/sync impedance can also show up in unusual places, namely the CUDA caching allocator.
Further reading.
We talk about gradcheck, the property based testing mechanism that we use to verify the correctness of analytic gradient formulas in PyTorch. I'll talk a bit about testing in general, property based testing and why gradcheck is a particularly useful property based test. There will be some calculus, although I've tried to keep the math mostly to intuitions and pointers on what to read up on elsewhere.
Further reading.
torch.use_deterministic_algorithms lets you force PyTorch to use deterministic algorithms. It's very useful for debugging!
There are some errors in the recording: the feature is called torch.use_deterministic_algorithms, and there is not actually a capability to warn (this was in an old version of the PR but taken out), we just error if you hit nondeterministic code.
Reference counting is a common memory management technique in C++ but PyTorch does its reference counting in a slightly idiosyncratic way using intrusive_ptr. We'll talk about why intrusive_ptr exists, the reason why refcount bumps are slow in C++ (but not in Python), what's up with const Tensor& everywhere, why the const is a lie and how TensorRef lets you create a const Tensor& from a TensorImpl* without needing to bump your reference count.
Further reading.
Memory layout specifies how the logical multi-dimensional tensor maps its elements onto physical linear memory. Some layouts admit more efficient implementations, e.g., NCHW versus NHWC. Memory layout makes use of striding to allow users to conveniently represent their tensors with different physical layouts without having to explicitly tell every operator what to do.
Further reading.
pytorch-probot is a GitHub application that we use to automate common tasks in GitHub. I talk about what it does and some design philosophy for it. Repo is at: https://github.com/pytorch/pytorch-probot
Lexical and dynamic scoping are useful tools to reason about various API design choices in PyTorch, related to context managers, global flags, dynamic dispatch, and how to deal with BC-breaking changes. I'll walk through three case studies, one from Python itself (changing the meaning of division to true division), and two from PyTorch (device context managers, and torch function for factory functions).
Further reading.
__future__ in libraries https://stackoverflow.com/questions/66927362/way-to-opt-into-bc-breaking-changes-on-methods-within-a-single-moduleToday, Shen Li (mrshenli) joins me to talk about distributed computation in PyTorch. What is distributed? What kinds of things go into making distributed work in PyTorch? What's up with all of the optimizations people want to do here?
Further reading.
Double backwards is PyTorch's way of implementing higher order differentiation. Why might you want it? How does it work? What are some of the weird things that happen when you do this?
Further reading.
Functional modules are a proposed mechanism to take PyTorch's existing NN module API and transform it into a functional form, where all the parameters are explicit argument. Why would you want to do this? What does functorch have to do with it? How come PyTorch's existing APIs don't seem to need this? What are the design problems?
Further reading.
What are CUDA graphs? How are they implemented? What does it take to actually use them in PyTorch?
Further reading.
What do default arguments have to do with PyTorch design? Why are default arguments great for clients (call sites) but not for servers (implementation sites)? In what sense are default arguments a canonicalization to max arity? What problems does this canonicalization cause? Can you canonicalize to minimum arity? What are some lessons to take?
Further reading. https://github.com/pytorch/pytorch/issues/54613 stop serializing default arguments
What's a domain library? Why do they exist? What do they do for you? What should you know about developing in PyTorch main library versus in a domain library? How coupled are they with PyTorch as a whole? What's cool about working on domain libraries?
Further reading.
Line notes.
What's TensorAccessor? Why not just use a raw pointer? What's PackedTensorAccessor? What are some future directions for mixing statically typed and typed erase code inside PyTorch proper?
Further reading.
Why are RNGs important? What is the generator concept? How do PyTorch's CPU and CUDA RNGs differ? What are some of the reasons why Philox is a good RNG for CUDA? Why doesn't the generator class have virtual methods for getting random numbers? What's with the next normal double and what does it have to do with Box Muller transform? What's up with csprng?
Further reading.
What is vmap? How is it implemented? How does our implementation compare to JAX's? What is a good way of understanding what vmap does? What's up with random numbers? Why are there some issues with the vmap that PyTorch currently ships?
Further reading.
What's an expect test? Why should you use them? Why is inline expect test better than out of line? How to write a good expect test?
Further reading. expecttest source implementation https://github.com/pytorch/pytorch/blob/master/torch/testing/_internal/expecttest.py (only 311 lines!)
What's PyTorch XLA? Why should you care? How is it implemented? How does PyTorch XLA trade off functionality versus ease of performance debugging? What are some new developments in this space?
Further reading.
What is TH? Why might you care? What is so horrible about it? What the heck is the generic/ folder? Why are we porting everything to C++? What are some downsides of having ported all our TH code to C++?
Further reading.
There is a really good TorchScript overview at https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/OVERVIEW.md and in this 20min podcast, I want to give you some of the highlights from this document.
Why is PyTorch's build so g-dang complicated. How to avoid having to deal with cmake at all? And if you have to deal with cmake, what are the most important things to know? And if you were going to improve our cmake, how would you go about doing it...
Further reading.
Liner notes.
torchdeploy is a way of running multiple Python interpreters inside the same process. It can be used to deploy Python PyTorch programs in situations where the GIL is a problem, not the CPython interpreter. How does it work, and what kind of challenges does it pose for people who want to write code that calls from C++ to Python?
Further reading.
What's the C++ frontend? Why is avoiding templates so important? Why is Tensor a reference type? How do we simulate keyword arguments in C++? Where did the nn Module support in the C++ API come from? Why did we reimplement all modules in C++? How are modules implemented in C++? What are some performance challenges of writing Python in C++, and how are we working around them?
Further reading.
Given two separately refcounted objects, how can you arrange for each of them to stay live so long as the other is live? Why doesn't just having a strong-strong or strong-weak reference between the two objects work? What is object resurrection in CPython? What's a finalizer and why does it make things more complicated? How does Python GC work?
Further reading.
What is mobile selective build? Why are we so obsessed with reducing binary size? How does selective build work? Why doesn't static linking just work? Why can't you just read out the ops used in a TorchScript model to determine what operators you actually need? What are the tradeoffs of statically determining the operator dependency graph versus tracing? What's up with the SELECTIVE_NAME macro? How the heck does selective build work at all when you have multiple mobile apps in a single Buck build system? What takeaways should I have as a regular PyTorch developer?
Further reading:
Liner notes:
binary size is premium; ship only what you actually need
big idea:
get the ops your model needs
apply this to build of pytorch
common hiccups
What goes into the implementation of torch.nn? Why do NN modules exist in the first place? What's the function of Parameter? How do modules actually track all the parameters in question? What is all of the goop in the top level NN module class? What are some new developments in torch.nn modules? What are some open problems with our modules?
Further reading:
Liner notes:
__call__ to forward (extra instrumentation)Why does PyTorch use code generation as part of its build process? Why doesn't it use C++ templates? What things is code generation used for? What are the pros/consof using code generation? What are some other ways to do the same things we currently do with code generation?
Further reading.
Outline:
Why is autograd so complicated? What are the constraints and features that go into making it complicated? What's up with it being written in C++? What's with derivatives.yaml and code generation? What's going on with views and mutation? What's up with hooks and anomaly mode? What's reentrant execution? Why is it relevant to checkpointing? What's the distributed autograd engine?
Further reading.
What is __torch_function__? Why would I want to use it? What does it have to do with keeping extra metadata on Tensors or torch.fx? How is it implemented? Why is __torch_function__ a really popular way of extending functionality in PyTorch? What makes it different from the dispatcher extensibility mechanism? What are some downsides of it being written this way? What are we doing about it?
Further reading.
__torch_function__ RFC: https://github.com/pytorch/rfcs/blob/master/RFC-0001-torch-function-for-methods.md__torch_function__ https://pytorch.org/docs/stable/notes/extending.html#extending-torchYou walk into the whiteboard room to do a technical interview. The interviewer looks you straight in the eye and says, "OK, can you show me how to add the elements of two lists together?" Confused, you write down a simple for loop that iterates through each element and adds them together. Your interviewer rubs his hands together evilly and cackles, "OK, let's make it more complicated."
What does TensorIterator do? Why the heck is TensorIterator so complicated? What's going on with broadcasting? Type promotion? Overlap checks? Layout? Dimension coalescing? Parallelization? Vectorization?
Further reading.
What does native_functions.yaml have to do with the TorchScript compiler? What multiple use cases is native_functions.yaml trying to serve? What's up with the JIT schema type system? Why isn't it just Python types? What the heck is the (a!) thingy inside the schema? Why is it important that I actually annotate all of my functions accurately with this information? Why is my seemingly BC change to native_functions.yaml actually breaking people's code? Do I have to understand the entire compiler to understand how to work with these systems?
Further reading.
What is serialization? Why do I care about it? How is serialization done in general in Python? How does pickling work? How does PyTorch implement pickling for its objects? What are some pitfalls of pickling implementation? What does backwards compatibility and forwards compatibility mean in the context of serialization? What's the difference between directly pickling and using torch.save/load? So what the heck is up with JIT/TorchScript serialization? Why did we use zip files? What were some design principles for the serialization format? Why are there two implementations of serialization in PyTorch? Is the fact that PyTorch uses pickling for serialization mean that our serialization format is insecure?
Further reading.
__reduce_ex__ https://github.com/pytorch/pytorch/blob/de845020a0da39e621db984515bc1cce03f526ea/torch/_tensor.py#L97-L178How is our CI put together? What is the history of the CI? What constraints are under the CI? Why does the CI use Docker? Why are build and test split into two phases? Why are some parts of the CI so convoluted? How does the HUD work? What kinds of configurations is PyTorch tested under? How did we decide what configurations to test? What are some of the weird CI configurations? What's up with the XLA CI? What's going on with the Facebook internal builds?
Further reading.
What's a stacked diff? Why might you want to do it? What does the workflow for stacked diffs with ghstack look like? How to use interactive rebase to edit earlier diffs in my stack? How can you actually submit a stacked diff to PyTorch? What are some things to be aware of when using ghstack?
Further reading.
What is shared memory? How is it used in your operating system? How is it used in PyTorch? What's shared memory good for in deep learning? Why use multiple processes rather than one process on a single node? What's the point of PyTorch's shared memory manager? How are allocators for shared memory implemented? How does CUDA shared memory work? What is the difference between CUDA shared memory and CPU shared memory? How did we implement safer CUDA shared memory?
Further reading.
What is automatic mixed precision? How is it implemented? What does it have to do with mode dispatch keys, fallthrough kernels? What are AMP policies? How is its cast caching implemented? How does torchvision also support AMP? What's up with Intel's CPU autocast implementation?
Further reading.
What are complex numbers? What is conjugation? Why is conjugation so common in linear algebra? Why would we like conjugation to behave similarly to transposition (and why is matrix multiply with a transposed input so fast?) What is a conjugate view? How is it implemented? What's the relationship between views, laziness and call-by-name evaluation?
Further reading.
What historical constraints and design choices lead to the design of Tensor/Storage (and their Impl variants) as they are today? Why do we use intrusive refcounting? Why are we trying to get rid of virtual methods on TensorImpl? Why are there so many frickin' bitfields?
Further reading.
What's the general process by which a new operator is added to PyTorch? Why is this actually something of a rare occurrence? How do you integrate an operator with the rest of PyTorch's system so it can be run end-to-end? What should I expect if I'm writing a CPU and CUDA kernel? What tools are available to me to make the job easier? How can I debug my kernels? How do I test them?
Further reading.
What is a Variable? Why did it exist as a wrapper in the first place? Why did it get removed? How did we remove it? What are some of the lingering consequences of its removal?
Further reading:
What's the current state of backend extensibility? How did PyTorch evolve from being a CPU and CUDA only framework to also support AMD ROCm and XLA? What are some problems with adding an out-of-tree backend, and what's some work to make it better?
Further reading:
Structured kernels are a new way to write kernels in PyTorch. Why did they take so long? What finally convinced us that we should do them? Why did it end up taking me the better part of a year to only be half done with them?
Further reading:
Functionalization is the process by which we remove mutation from autograd graphs in PyTorch, leaving us with a purely functional graph that we can execute in the normal way. Why do we need to do functionalization? What makes it not so easy to do? How do we do it? And how does it compare to mutation removal that you might see in a compiler?
Further reading:
Ever wanted to learn about CUDA but not sure where to start? In this sixteen minute episode I try to jam in as much CUDA knowledge as could be reasonably expected in a podcast. You won't know how to write a kernel after this episode, but you'll know about what a GPU is, what the general CUDA programming model is, why asynchronous execution makes everything complicated, and some general principles PyTorch abides by when designing CUDA kernels.
Further reading:
What's inference mode? Why doesn't my code run fast if I use no_grad or make sure requires_grad=False? How come inference mode is safe but AutoNonVariableTypeMode is not?
Further reading:
What is vectorization? How do you use it in PyTorch? What are some of the traps and pitfalls of writing vectorized code in PyTorch?
Further reading:
Why is PyTorch split into so many libraries? What's the point of these splits? What do Windows, mobile and CUDA have to do with the library splits?
Further reading:
Why is the dispatcher the way it is today? How did evolve over time, and what constraints got added so that it is the kind of complicated piece it is today?
Further reading:
In this episode, we will discuss how to bind a C++ object in Python? We'll try to answer the following questions: How does pybind11 do it? What's different about how we implement it for Tensor? What are some downsides of the approach?
Note from the future: I recorded and then decided I didn't like my follow up episode about how to preserve PyObjects even when they go dead in Python. Maybe some day!
Further reading:
En liten tjänst av I'm With Friends. Finns även på engelska.