This PR is on the way to getting compiled autograd's initial capture to
stop specializing on Tensor metadata.
This PR changes compiled autograd's initial capture to proxy an opaque
(w.r.t. Dynamo) function into the graph for all built-in codegen'ed
autograd nodes and validate_outputs.
We changed each codegen'ed apply_with_saved (e.g.
MulBackward0::apply_with_saved) to call into Python to proxy a function
(compiled_autograd.ops.MulBackward0) into the graph. Then, we use the
node's InputMetadata to "guess" at the properties of the output Tensors
to create some new FakeTensors.
Some details:
- MulBackward0::apply_with_saved lives in libtorch_cpu, but needs to be
call to Python via libtorch_python. There is an indirection
(PyCompilerInterface) to do this.
- MulBackward0::apply_with_saved passes a C++ function to Python. To make
our lives easier, every codegen'ed apply_with_saved passes a C++
function with the same signature
`(variable_list, ivalue_list) -> variable_list`.
- We define how to pack arbitrary C++ types into IValue via a helper
IValuePacker struct and codegen functional variants of each builtin
C++ autograd node (e.g. MulBackward0_apply_functional_ivalue).
MulBackward0 before this PR:
https://gist.github.com/zou3519/a80381d5fa38e970e413fcd91b0530de
MulBackward0 after this PR:
https://gist.github.com/zou3519/0c2eee8b3d8d96232b51ef430b53c5b0
Test Plan:
- existing tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143296
Approved by: https://github.com/jansel
This PR refactors all builtin autograd nodes (e.g. MulBackward0) from
having a single MulBackward0::apply into having:
- a "pure function variant" `MulBackward0_apply_functional`
- a stateful variant MulBackward0::apply that ends up calling
`MulBackward0_apply_functional`.
In order to do this we left the stateful pieces in MulBackward0::apply
(like unpacking of saved vars, determining which gradients actually need
computing).
The motivation is that this will be useful for compiled autograd in a
future PR. We might refactor this more later, but I wanted to get
something reviewed, shipped, and tested in-tree because the entire stack
is going to be big and this change by itself might have subtle perf issues.
The new codegen looks like the following:
- https://gist.github.com/zou3519/84721cfbef71bb640ddf1a64ef8583a3
Here's the old codegen for comparison:
- https://gist.github.com/zou3519/73f925fe6aca6dd3ceb0a6e6fcf5f77d
Test Plan:
- existing tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141278
Approved by: https://github.com/soulitzer
This branch:
1) converts the autograd tape into an FX graph
2) caches that conversion using a "shadow" graph
3) compiles and runs the generated FX graph instead of the normal autograd
What works currently:
1) Caching, capture, and initial integration
2) Backwards hooks
3) Inlining AotAutograd generated subgraphs
4) torch.compiling the generated FX graph
5) Auto-detecting dynamic shapes based on changes
Future work
1) Larger scale testing
1) Boxed calling convention, so memory can be freed incrementally
1) Support hooks on SavedTensor
1) Additional testing by running eager autograd tests under compiled_autograd.enable()
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103822
Approved by: https://github.com/ezyang, https://github.com/albanD
The strategy is that we will heap allocate a LargeNegativeIntSymNodeImpl whenever we have a large negative int, so that we can keep the old `is_symbolic` test (now called `is_heap_allocated`) on SymInt. Whenever we need to do something with these ints, though, we convert them back into a plain `int64_t` (and then, e.g., wrap it in whatever user specificed SymNodeImpl they need.) We cannot wrap directly in the user specified SymNodeImpl as we generally do not know what the "tracing context" is from C++. We expect large negative ints to be rare, so we don't apply optimizations like singleton-ifying INT_MIN. Here's the order to review:
* c10/core/SymInt.h and cpp
* `is_symbolic` renamed to `is_heap_allocated` as I needed to audit all use sites: the old `is_symbolic` test would return true for large negative int, but it would be wrong to then try to dispatch on the LargeNegativeIntSymNodeImpl which supports very few operations. In this file, I had to update expect_int,
* If you pass in a large negative integer, we instead heap allocate it in `promote_to_negative`. The function is written in a funny way to keep compact constructor code for SymInt (the heap allocation happens out of line)
* clone is now moved out-of-line
* New method maybe_as_int which will give you a constant int if it is possible, either because it's stored inline or in LargeNegativeIntSymNodeImpl. This is the preferred replacement for previous use of is_symbolic() and then as_int_unchecked().
* Rename toSymNodeImpl to toSymNode, which is more correct (since it returns a SymNode)
* Complete rewrite of `normalize_symints.cpp` to use new `maybe_as_int`. Cannot easily use the old code structure, so it's now done doing a macro and typing out each case manually (it's actually not that bad.)
* Reimplementations of all the unary operators by hand to use `maybe_as_int`, relatively simple.
* c10/core/LargeNegativeIntSymNodeImpl.h - Just stores a int64_t value, but it has to be big and negative. Most methods are not implemented, since we will rewrap the large negative int in the real SymNodeImpl subclass before doing operations with it
* The rest of the files are just rewriting code to use `maybe_as_int`. There is a nontrivial comment in c10/core/SymIntArrayRef.h
Very minor test adjustment in c10/test/core/SymInt_test.cpp . Plan to exercise this properly in next PR.
Companion XLA PR: https://github.com/pytorch/xla/pull/4882
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99157
Approved by: https://github.com/albanD
This PR:
- registers all of the codegened Nodes to the torch._C._functions module, this is where special nodes like AccumulateGrad are already registered.
- creates a autograd.graph.Node abstract base class that all of the newly registered nodes subclass from. We make the subclassing happen by implementing the ``__subclasshook__`` method
- enables static type checking to work and also enables Sphinx to generate documentation for the Node and its methods
- handles both the custom Function and codegened cases
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91475
Approved by: https://github.com/albanD
This refactor was prompted by challenges handling mixed int/float
operations in C++. A previous version of this patch
added overloads for each permutation of int/float and was unwieldy
https://github.com/pytorch/pytorch/pull/87722/ This PR takes a different
approach.
The general outline of the patch is to combine the C++ types SymIntNode
and SymFloatNode into a single type, SymNode. This is type erased; we
no longer know statically at C++ if we have an int/float and have to test
it with the is_int()/is_float() virtual methods. This has a number of
knock on effects.
- We no longer have C++ classes to bind to Python. Instead, we take an
entirely new approach to our Python API, where we have a SymInt/SymFloat
class defined entirely in Python, which hold a SymNode (which corresponds
to the C++ SymNode). However, SymNode is not pybind11-bound; instead,
it lives as-is in Python, and is wrapped into C++ SymNode using PythonSymNode
when it goes into C++. This implies a userland rename.
In principle, it is also possible for the canonical implementation of SymNode
to be written in C++, and then bound to Python with pybind11 (we have
this code, although it is commented out.) However, I did not implement
this as we currently have no C++ implementations of SymNode.
Because we do return SymInt/SymFloat from C++ bindings, the C++ binding
code needs to know how to find these classes. Currently, this is done
just by manually importing torch and getting the attributes.
- Because SymInt/SymFloat are easy Python wrappers, __sym_dispatch__ now
takes SymInt/SymFloat, rather than SymNode, bringing it in line with how
__torch_dispatch__ works.
Some miscellaneous improvements:
- SymInt now has a constructor that takes SymNode. Note that this
constructor is ambiguous if you pass in a subclass of SymNode,
so an explicit downcast is necessary. This means toSymFloat/toSymInt
are no more. This is a mild optimization as it means rvalue reference
works automatically.
- We uniformly use the caster for c10::SymInt/SymFloat, rather than
going the long way via the SymIntNode/SymFloatNode.
- Removed some unnecessary toSymInt/toSymFloat calls in normalize_*
functions, pretty sure this doesn't do anything.
- guard_int is now a free function, since to guard on an int you cannot
assume the method exists. A function can handle both int and SymInt
inputs.
- We clean up the magic method definition code for SymInt/SymFloat/SymNode.
ONLY the user classes (SymInt/SymFloat) get magic methods; SymNode gets
plain methods; this is to help avoid confusion between the two types.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
cc @jansel @mlazos @soumith @voznesenskym @yanboliang @penguinwu @anijain2305
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87817
Approved by: https://github.com/albanD, https://github.com/anjali411
Partially fixes: #66328
This PR:
- adds support for `ITensorList` to the dispatcher for:
- computing the dispatch key
- boxing and unboxing `ITensorList`
- modified the codegen for structured kernels:
- codegen APIs use `ITensorList` instead of `ArrayRef<Tensor>`
**Changes summary:**
- Signature changes due to the different APIs:
- dispatcher API (e.g. `BatchingRegistrations.cpp`)
- C++ API (e.g. `TensorShape.cpp`)
- Miscelaneous functions used by codegen'd functions (e.g. `FunctionalTensorWrapper.*`)
- Dispatcher changes for handling `ITensorList` correctly (e.g. `DispatchKeyExtractor.h`)
- Signature changes of `at::cat` due to the need of `const` inside `TensorBody.h`
- Forward declarations of `ITensorList` (e.g. `MethodOperators.h`)
- Codegen changes, special casing structured kernels (e.g. `gen.py`)
**Short description of structured kernels special casing:**
I introduced, mainly, 5 types of changes to the codegen for generating code depending on
whether the kernel is structured or not:
1. Added a `structured_type_override` flag to the `argument_type` function definition of
the affected APIs (mainly the dispatcher and C++ APIs).
- `api/cpp.py`, `api/dispatcher.py`, `api/native.py`
2. Added a `structured_type_override` member to the signature
classes (e.g. `CppSignature`), since `FunctionSchema` doesn't really know whether the
function is structured or not
- `api/types.py`
3. Added a `part_of_structured_group` to `NativeFunction` class, which is just a
convenient function to forward to `structured_type_override` wherever needed
- `model.py`
4. Appropriately changed the rest of the codegen, whenever it used either the signature
classes or the `arguments` function directly
5. Added a check for `const ITensorList&` type wherever there was a check for `TensorList`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73350
Approved by: https://github.com/bdhirsh
Summary:
Like it says in the title. Currently, this will return output like this:
In Buck1, that's OK because Buck1's caching doesn't really care too much about
However, in Buck2, this is a disaster, because caching is based exclusively
on inputs and outputs and
The diff here proposes making the path relative to the codegen script itself,
which should carry about as much info, but avoid cache misses.
Concretely, this:
```
// generated from /dev/shm/uid-34135/cfbc5712-seed-nspid4026533424_cgpid2794673-ns-4026533443/tools/autograd/templates/python_functions.h
```
Becomes, this:
```
// generated from ../tools/autograd/templates/python_functions.h
```
So, we keep the useful part, and we get caching. This matters because those
headers are used in actions like:
```
fbcode//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops -- action (cxx_compile gen_embedding_backward_adam_split_unweighted_cuda.cu (pic))
```
Those actions take upwards of 5 minutes to finish, so by allowing a cache hit,
we are a) saving our users a lot of time and b) saving some RE capacity as
well.
This actually matters a lot because right now those targets are produced by
`//caffe2:generate-code`, which itself doesn't get cache hits from RE because
`generate_code.par` is non-deterministic (this is, unfortunately, true of PARs
in general), so that rule introduces non-determinism that the codegen
propagates and we get zero caching.
This diff doesn't fix `//caffe2:generate-code`'s inputs being
non-deterministic, but it does fix its *outputs* being non-deterministic, which
means the non-determinism stops there, and we get back to cache hits.
Test Plan:
- CI
```
buck2 build fbcode//caffe2:generate-code
buck2 build fbcode//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops
```
Reviewed By: ndmitchell
Differential Revision: D39348565
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84695
Approved by: https://github.com/soulitzer
### Introduction
<!-- What did you change and why was it needed? -->
Removing unnecessary weight gradient calculation is very important for applications that need high-order derivatives during training. However, this is not supported by the current Autograd engine.
For more detail: The backward function of a `matmul` operator (e.g., `linear` `addmm` `mm`), has two matmuls, one for `input gradient` and another for `weight gradient`. For a typical neural network (nn) with a few linear layers and activation functions, if the user calls `torch.autograd.grad()` to calculate the derivative of the nn output `y` w.r.t the nn input `x`, only the `input gradient` of the `matmul` operator is needed, and the `weight gradient` is discarded. However, the current PyTorch autograd engine will always calculate the `weight gradient` if `weight` requires gradient (the calculation of the high-order derivative is performed during training).
The figure attached shows the autograd graph of the following code snippet:
```py
y = torch.nn.functional.linear(x, weight, bias)
y = y.pow(2)
# first order derivative
y__x, = torch.autograd.grad(y, x, grad_outputs=grad_outputs, create_graph=True)
# first order derivative
y__x__x, = torch.autograd.grad(y__x, x, grad_outputs=grad_outputs, create_graph=True)
```
The path with ❌ is not needed when calculating derivatives.
<img width="50%" alt="image" src="https://user-images.githubusercontent.com/9999318/182018117-719c5a23-bcc6-4a63-8e8d-1bca3ebda2e3.png">
### Issue
<!-- Link to Issue ticket or RFP -->
Related issue: https://github.com/pytorch/pytorch/issues/56500
### Method
When calling `torch.autograd.grad`, `exec_info_` is created for each GraphTask, which allows filtering paths on the graph that are not needed. However, when the GraphTask calls into the node, the node still does not know whether the edges are needed or not. In the case of matmul, `weight.requires_grad is True` so the weight gradient is always calculated.
Following https://github.com/pytorch/pytorch/issues/56500#issuecomment-825694656, this PR passes the graph task's thread_local `exec_info_` into the node, so it could trim unnecessary edges during `torch.autograd.grad` calls.
### Benchmark
Benchmark script: https://gist.github.com/yueyericardo/24158433a2021c51eeef9c3e2722df99
Benchmark result:
6 hidden layers, batch size 10000, on A100
FP32 result
| hessian benchmark | FP32 (before) | FP32 (After) | FP32 (Functorch v0.1.1) |
| ----------------------------- | ------------- | ----------------- | ----------------------- |
| Linear + ReLU (no backward) | 55.658 ms | 29.392 ms (1.90X) | 29.547 ms (1.90X) |
| Linear + ReLU (with backward) | 81.173 ms | 54.917 ms (1.47X) | 68.988 ms (1.18X) |
TF32 result
| hessian benchmark | TF32 (before) | TF32 (after) | TF32 (Functorch v0.1.1) |
| ----------------------------- | ------------- | ----------------- | ----------------------- |
| Linear + ReLU (no backward) | 19.801 ms | 11.259 ms (1.76X) | 10.754 ms (1.84X) |
| Linear + ReLU (with backward) | 29.167 ms | 20.466 ms (1.42X) | 22.784 ms (1.28X) |
For FP32 result, we could get 1.9X speed up for hessian calculation, and 1.47X speed up during training, which is even faster than functorch `vmap(jacfwd(jacrev` implementation. (functorch has performance regression on v0.2.0, https://github.com/pytorch/functorch/issues/989, so we are using v0.1.1 for benchmark)
@zou3519 does functorch also includes similar optimizations during hessian calculation? If not, what do we need to do so the functorch could also benefit from this PR?
### Testing
<!-- How did you test your change? -->
- [x] we need to figure out a way for unittest
### Thanks
Thanks for the great blog: [How Computational Graphs are Executed in PyTorch | PyTorch](https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/)
cc @zasdfgbnm @albanD
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82544
Approved by: https://github.com/soulitzer
`derivatives.yaml` can now take a `dispatch` entry which registers per-autograd dispatch key derivatives such as
```
name: foo(Tensor self, Tensor y) -> Tensor
dispatch:
Default:
x: grad
y: grad.expand(y.sizes())
AutogradNestedTensor:
x: grad
y: NestedTensor_foo_backward(grad, y)
output_differentiabilty: [True]
```
However the old schema where there is no `dispatch` entry is still supported.
Would greatly appreciate feedback on *how to improve the testing strategy* of this PR, currently have registered an aten test op in TestOps.cpp with dummy gradients in derivatives.yaml and have some tests in test_autograd.py:TestAutogradMultipleDispatch but I am not sure whether these are sufficiently rigorous.
Additionally, this PR also makes the assumption that sets like [VIEW_FUNCTIONS](ff5399e528/tools/autograd/gen_inplace_or_view_type.py (L60)) are per-native-function and not per-native-function-and-dispatch-key. I'm not sure whether this is necessarily the case, *would there ever be a situation where (e.g. a nested_tensor op is a view op but the aten function is not or vice versa?)*
* __->__ #82801
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82801
Approved by: https://github.com/bhosmer, https://github.com/albanD
Done via
```
git grep -l 'SymbolicIntNode' | xargs sed -i 's/SymbolicIntNode/SymIntNodeImpl/g'
```
Reasoning for the change:
* Sym is shorter than Symbolic, and consistent with SymInt
* You usually will deal in shared_ptr<...>, so we're going to
reserve the shorter name (SymIntNode) for the shared pointer.
But I don't want to update the Python name, so afterwards I ran
```
git grep -l _C.SymIntNodeImpl | xargs sed -i 's/_C.SymIntNodeImpl/_C.SymIntNode/'
```
and manually fixed up the binding code
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82350
Approved by: https://github.com/Krovatkin
With ufmt in place https://github.com/pytorch/pytorch/pull/81157, we can now use it to gradually format all files. I'm breaking this down into multiple smaller batches to avoid too many merge conflicts later on.
This batch (as copied from the current BLACK linter config):
* `tools/**/*.py`
Upcoming batchs:
* `torchgen/**/*.py`
* `torch/package/**/*.py`
* `torch/onnx/**/*.py`
* `torch/_refs/**/*.py`
* `torch/_prims/**/*.py`
* `torch/_meta_registrations.py`
* `torch/_decomp/**/*.py`
* `test/onnx/**/*.py`
Once they are all formatted, BLACK linter will be removed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81285
Approved by: https://github.com/suo
