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
Thanks @awgu for raising this issue and the small repro
From offline discussion with @albanD, in the case where a forward returns multiple outputs with different devices, we'd want to select the ready queue based on the device of the first one. Even though this is somewhat arbitrary, we prefer this over deciding which ready queue to push based on whichever input buffer's we happen to compute last, which can vary depending on more factors and thus be harder to reason about. This is in theory bc-breaking, but it seems unlikely that someone would depend on this behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135633
Approved by: https://github.com/albanD
Summary:
In some cases (especially those involving collective calls) - we would want to always kick off a collective call first before running going down another path.
For example:
```
tbe lookup -> a2a ->
overarch
dense ------------->
```
if the forward code is written as
a2a_out = a2a
dense = dense_net
out = overarch(a2a_out, dense)
out.backward()
The current default is running backwards in the opposite order the forward is called. However, there is no data dependency between a2a and dense, so in reality either of them could be run first. We would like the a2a to run first because it provides optimal (on average) overlap.
Changing the seq_nr of a2a_out to something large enough would allow autograd engine to kick it off first.
Test Plan: Tests incoming
Differential Revision: D51445261
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114120
Approved by: https://github.com/ezyang, https://github.com/albanD
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
How the old retains_grad hooks was implemented:
- retains_grad hooks are stored on the autograd_meta, as entries in a vector
- upon registration, a wrapper hook CppFunctionTensorPreHook is created to wrap that vector, and then that wrapper hook is registered to the grad_fn, i.e., by appending it to a vector of retains_grad hooks on the grad_fn
- upon in-place, for the old grad_fn we set the retains_grad hook to nullptr, so that even though the old grad_fn still references the vector, the vector contains a single nullptr. For the new grad_fn, we create a new wrapper hook around the vector (storing the single retains_grad hook) on autograd_meta.
The new retains_grad hook implementation:
- we store std::function by value, and we store it on the grad_fn rather than the autograd_meta
- a single grad_fn can have multiple outputs, so it can potentially hold multiple retains_grad hooks. We use an unordered_map (previously a vector).
- on in-place we remove the hook from the old grad_fn and put it in the new grad_fn (small implication of this change is that we we now need to have access to both the old grad_fn and new grad_fn, this isn't a problem)
Other details:
- CppFunctionTensorPreHook took a shared_ptr to vector of std::function. In our new implementation, we add a new wrapper hook CppFunctionSingleTensorPreHook, which takes a single std::function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92604
Approved by: https://github.com/albanD
This reverts commit e525f433e15de1f16966901604a8c4c662828a8a.
Original PR: #85849
Fixes #ISSUE_NUMBER
In addition to reverting the revert, this PR:
- defines the virtual destructor of FunctionPreHook in the header. Why? Presumably the internal build imports the header from somewhere, but does not have function_hooks.cpp (where the virtual destructor was previously defined) in the same compilation unit.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92559
Approved by: https://github.com/albanD
Addresses: https://github.com/pytorch/pytorch/issues/35802
Design doc: https://docs.google.com/document/d/19xSib7FFknRQ5f3ptGFUmiOt3BrgXSUlTQH2xMcZJYg/edit#
### Changes in this PR
#### Implementation
- We have now have 3 fields: pre_hooks, retains_grad_hooks, and tensor_pre_hooks so that we can more precisely define their ordering and when they are executed.
- Since retains grad uses an entirely new field, we cannot reuse the old retains grad, logic. We refactor retains grad to call directly into the variable.cpp logic. Other logic in variable.cpp that handle cpp hooks must also be updated.
#### Hooks ordering and execution:
- Defines pre-hooks registered on tensor to run before pre-hooks registered on grad_fn
- Updates pre-hooks registered on tensor to always run, even if they are the inputs= to .grad()
- Post hooks (and pre hooks) can now observe the modifications to gradient by the tensor pre hook
#### Retains grad hooks
- retains grad hooks always execute last, even if there are other tensor pre-hooks registered
#### Unchanged:
- pre_hooks registered to grad_fn aren't expected to execute if they are the inputs= to .grad()
Follow ups:
- simplify retains_grad field to not be a vector, since it always holds a single hook
- potentially merge capture hooks with tensor pre hooks, this would involve some additional refactoring since
- python hooks registered to tensor behavior on in-place is still wrong
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85849
Approved by: https://github.com/albanD
This PR removes the autograd.Function extension feature flag. This was
previously used for development of the functorch <> autograd.Function
interaction.
It's been in master for long enough with the feature flag defaulting to
True, so it's time to remove it.
Test Plan:
- existing tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92026
Approved by: https://github.com/soulitzer
This allows to know at any point during the backward pass what is running and where the Node currently running was created at:
```python
import torch
from torch.utils._python_dispatch import TorchDispatchMode
from torch.autograd import detect_anomaly
class MyMode(TorchDispatchMode):
def __torch_dispatch__(self, func, types, args, kwargs=None):
node = torch._C._current_autograd_node()
print(f"Running {func} from within {node}")
if node is not None:
print("The Node was created at:")
print("\n ".join(node.metadata["traceback_"]))
return func(*args, **kwargs or {})
with MyMode(), detect_anomaly():
print("FW")
a = torch.rand(10, requires_grad=True)
b = a.mul(2)
b = b.div(3)
b = b.sum()
print("BW")
b.backward()
```
Gives
```
$ python foo.py
foo.py:15: UserWarning: Anomaly Detection has been enabled. This mode will increase the runtime and should only be enabled for debugging.
with MyMode(), detect_anomaly():
FW
Running aten.rand.default from within None
Running aten.mul.Tensor from within None
Running aten.div.Tensor from within None
Running aten.sum.default from within None
BW
Running aten.ones_like.default from within None
Running aten.expand.default from within <SumBackward0 object at 0x7fa40c0c6dc0>
The Node was created at:
File "foo.py", line 20, in <module>
b = b.sum()
Running aten.isnan.default from within <SumBackward0 object at 0x7fa40c0c6500>
The Node was created at:
File "foo.py", line 20, in <module>
b = b.sum()
Running aten.any.default from within <SumBackward0 object at 0x7fa32b23a780>
The Node was created at:
File "foo.py", line 20, in <module>
b = b.sum()
Running aten._local_scalar_dense.default from within <SumBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 20, in <module>
b = b.sum()
Running aten.div.Tensor from within <DivBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 19, in <module>
b = b.div(3)
Running aten.isnan.default from within <DivBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 19, in <module>
b = b.div(3)
Running aten.any.default from within <DivBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 19, in <module>
b = b.div(3)
Running aten._local_scalar_dense.default from within <DivBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 19, in <module>
b = b.div(3)
Running aten.mul.Tensor from within <MulBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 18, in <module>
b = a.mul(2)
Running aten.isnan.default from within <MulBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 18, in <module>
b = a.mul(2)
Running aten.any.default from within <MulBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 18, in <module>
b = a.mul(2)
Running aten._local_scalar_dense.default from within <MulBackward0 object at 0x7fa40c0c9190>
The Node was created at:
File "foo.py", line 18, in <module>
b = a.mul(2)
Running aten.detach.default from within <AccumulateGrad object at 0x7fa40c0c9730>
The Node was created at:
File "foo.py", line 18, in <module>
b = a.mul(2)
Running aten.detach.default from within <AccumulateGrad object at 0x7fa40c0c94b0>
The Node was created at:
File "foo.py", line 18, in <module>
b = a.mul(2)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90867
Approved by: https://github.com/soulitzer
This PR adds a private runtime feature flag for the feature work we're going
to do with extending autograd.Function. The motivation of the feature flag
is:
- to guard the feature against unsuspecting users
- control the release of the feature to when we are ready to release it
We might not even need the feature flag (because we hope to have the
work done in the next month), but it is good practice and it does touch
currently public API (autograd.Function).
Concretely, "autograd.Function extension" refers to:
- adding an optional `setup_context` staticmethod to autograd.Function
- adding an optional `vmap` staticmethod to autograd.Function
- autograd.Function support for functorch
Test Plan:
- new test that the feature flag works
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89858
Approved by: https://github.com/soulitzer
In this PR:
- graph_task stores graph roots on construction so that we can later traverse through the graph
- before the nodes are returned, they needed to be converted from raw_ptr to shared_ptr, and this should be OK because the graph is guaranteed to be alive
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87507
Approved by: https://github.com/albanD
### 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
