Summary:
**Review last commit only.** Stacked on top of #10949.
This commit fixes a number of issues connected to caching
differentiability status of graphs inside graph executors,
and changes the rules for optimization of differentiable subgraphs.
Previously every one of those was instantiated as a separate graph
executor, but now they are simply heavier-optimized graph regions,
and graph executors are only instantiated for their backward.
zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10977
Differential Revision: D9600626
Pulled By: apaszke
fbshipit-source-id: dad09a0f586e396afbd5406319c1cd54fbb8a3d3
Summary:
TODO: integrate into torch.onnx.export -- separate PR
*Problem:* We have a facility to trace PyTorch operations on Python code, but there are several failure modes where the trace is not representative of the actual underlying computation:
* The tracer encountered dynamic control flow
* Some computation escaped the tracer, and appeared as a Constant tensor node in the graph
* Some stateful function was traced, e.g. someone did an optimization in Python by memoizing function outputs
*Objective*: In an ideal world, this whole process would be automated and the user can trust that the system will magically capture the intended semantics from the program. Realistically speaking, we will likely have to settle with a human-in-the-loop error reporting system, allowing for the user to identify problems and modify the source code to allow for tracing.
*Stage 1* (this PR): Output-level checking & graph diff. torch.jit.trace gains a kwarg 'check_inputs', which is a list of tuples of input arguments. We will iterate through the list and trace the function again for each set of check inputs. We'll also interpret the original trace with these inputs and compare output values and graphs, printing a diff of the graph if there is a difference.
Examples:
```
torch.jit.trace(torch.rand(3, 4), check_inputs=[(torch.rand(4, 5),)])
def foo(x):
y = torch.arange(0, x.shape[0]).float()
return x + y.unsqueeze(1)
```
```
torch.jit.TracingCheckError: Tracing failed sanity checks!
ERROR: Graphs differed across invocations!
Graph diff:
graph(%0 : Dynamic) {
- %1 : Dynamic = prim::Constant[value= 0 1 2 [ CPULongType{3} ]]()
? ^
+ %1 : Dynamic = prim::Constant[value= 0 1 2 3 [ CPULongType{4} ]]()
? +++ ^
%2 : int = prim::Constant[value=0]()
%3 : Dynamic = aten::_cast_Float(%1, %2)
%4 : int = prim::Constant[value=1]()
%5 : Dynamic = aten::unsqueeze(%3, %4)
%6 : int = prim::Constant[value=1]()
%7 : Dynamic = aten::add(%0, %5, %6)
return (%7);
}
Node diff:
- %1 : Dynamic = prim::Constant[value= 0 1 2 [ CPULongType{3} ]]()
? ^
+ %1 : Dynamic = prim::Constant[value= 0 1 2 3 [ CPULongType{4} ]]()
? +++ ^
Trace source location:
dank.py(5): foo
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(402): wrapper
dank.py(3): <module>
Check source location:
dank.py(5): foo
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(281): check_trace
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(408): wrapper
dank.py(3): <module>
ERROR: Tensor-valued Constant nodes differed in value across invocations. This often indicates that the tracer has encountered untraceable code.
Node:
%1 : Dynamic = prim::Constant[value= 0 1 2 [ CPULongType{3} ]]()
Source Location:
dank.py(5): foo
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(402): wrapper
dank.py(3): <module>
Comparison exception:
Not equal to tolerance rtol=1e-07, atol=0
(shapes (3,), (4,) mismatch)
x: array([0, 1, 2])
y: array([0, 1, 2, 3])
```
==
```
torch.jit.trace(torch.rand(3, 4), check_inputs=[(torch.rand(3, 4),)])
def foo(x):
y = x.data
return x + y
```
```
torch.jit.TracingCheckError: Tracing failed sanity checks!
ERROR: Traced function outputs do not match the Python function outputs.
ERROR: Tensor-valued Constant nodes differed in value across invocations. This often indicates that the tracer has encountered untraceable code.
Node:
%1 : Dynamic = prim::Constant[value=<Tensor>]()
Source Location:
dank.py(6): foo
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(402): wrapper
dank.py(3): <module>
Comparison exception:
Not equal to tolerance rtol=1e-07, atol=0
(mismatch 100.0%)
x: array([0.397137, 0.956105, 0.169478, 0.560292, 0.392568, 0.108441,
0.97645 , 0.34412 , 0.951246, 0.793061, 0.557595, 0.770245],
dtype=float32)
y: array([0.243178, 0.315964, 0.972041, 0.0215 , 0.927751, 0.457512,
0.951092, 0.97883 , 0.048688, 0.118066, 0.779345, 0.271272],
dtype=float32)
```
==
```
import torch
torch.jit.trace(torch.rand(3, 4), check_inputs=[(torch.rand(4, 4),)])
def foo(x):
for _ in range(x.size(0)):
x = torch.neg(x)
return x
```
```
torch.jit.TracingCheckError: Tracing failed sanity checks!
ERROR: Traced function outputs do not match the Python function outputs.
ERROR: Graphs differed across invocations!
Graph diff:
graph(%0 : Dynamic) {
%1 : Dynamic = aten::neg(%0)
%2 : Dynamic = aten::neg(%1)
%3 : Dynamic = aten::neg(%2)
+ %4 : Dynamic = aten::neg(%3)
- return (%3);
? ^
+ return (%4);
? ^
}
```
==
```
import torch
def foo(x):
if not hasattr(foo, 'cache'):
foo.cache = torch.neg(x)
return x + foo.cache
traced = torch.jit.trace(torch.rand(3, 4), check_inputs=[(torch.rand(3, 4),)])(foo)
```
```
torch.jit.TracingCheckError: Tracing failed sanity checks!
ERROR: Traced function outputs do not match the Python function outputs.
ERROR: Graphs differed across invocations!
Graph diff:
graph(%0 : Dynamic) {
- %1 : Dynamic = aten::neg(%0)
+ %1 : Dynamic = prim::Constant[value=<Tensor>]()
%2 : int = prim::Constant[value=1]()
%3 : Dynamic = aten::add(%0, %1, %2)
return (%3);
}
Node diff:
- %1 : Dynamic = aten::neg(%0)
+ %1 : Dynamic = prim::Constant[value=<Tensor>]()
Trace source location:
test.py(5): foo
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(402): wrapper
test.py(8): <module>
Check source location:
test.py(6): foo
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(281): check_trace
/Users/jamesreed/onnx-fairseq/pytorch/torch/jit/__init__.py(408): wrapper
test.py(8): <module>
```
The following two examples show instances where program semantics are lost in the Python -> trace transformation, and repeated invocation does not give us useful debug information. Further design in underway for catching these scenarios.
```
import torch
torch.jit.trace(torch.rand(3, 4), check_inputs=[(torch.rand(3, 4),)])
def foo(x):
for i in range(3):
x[i, :] = torch.zeros(4)
return x
```
```
torch.jit.TracingCheckError: Tracing failed sanity checks!
ERROR: Traced function outputs do not match the Python function outputs.
Exception:
Not equal to tolerance rtol=1e-07, atol=0
(mismatch 100.0%)
x: array([0.830221, 0.915481, 0.940281, 0.555241], dtype=float32)
y: array([0., 0., 0., 0.], dtype=float32)
```
==
```
import torch
torch.jit.trace(torch.rand(3, 4), check_inputs=[(torch.rand(5, 6),)])
def foo(x):
x.view(-1).add_(-x.view(-1))
return x
```
```
torch.jit.TracingCheckError: Tracing failed sanity checks!
ERROR: Traced function outputs do not match the Python function outputs.
Exception:
Not equal to tolerance rtol=1e-07, atol=0
(mismatch 100.0%)
x: array([0.734441, 0.445327, 0.640592, 0.30076 , 0.891674, 0.124771],
dtype=float32)
y: array([0., 0., 0., 0., 0., 0.], dtype=float32)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10841
Differential Revision: D9499945
Pulled By: jamesr66a
fbshipit-source-id: 1f842a32d0b0645259cc43b29700b86d99c59a45
Summary:
Please review the expects carefully to make sure there are no regressions. I tried to go over them one by one when they changed, but it's sometimes easy to miss finer details.
Summary of changes:
- Renamed `TensorType` to `CompleteTensorType`. Added a new `TensorType` which records only the scalar type, number of dimensions, and device of a value. The argument behind the rename is to encourage people to use `CompleteTensorType` less, as most passes will only have limited information available. To make transition easier `complete_type->cast<TensorType>()` works, and makes our passes work with both kinds of specialization if they don't need extra the extra detail.
- Renamed `ArgumentSpec` to `CompleteArgumentSpec`. Added a new `ArgumentSpec`, which matches argument only at the level of the new `TensorType`.
- Shape analysis can process graphs with both `CompleteTensorType` and `TensorType`.
- Fuser was a part that heavily relied on full shape information being available. Now, we simply try to fuse the largest possible graphs, and have to do run-time checks to make sure they match the code we generate. If they don't, we fall back to regular interpretation. The shape checks are implementing using an optimized method exploiting algebraic properties of shapes with broadcasting, and the relations of broadcasting with pointwise ops. A full written proof of correctness of the shape checking algorithm is included in a comment in `graph_fuser.cpp`.
zdevito ezyang mruberry ngimel csarofeen
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10844
Differential Revision: D9498705
Pulled By: apaszke
fbshipit-source-id: 0c53c2fcebd871cc2a29c260f8d012276479cc61
Summary:
This is along the way of removing Tensor as a member of the tagged union in Scalar. This simplifies ordering dependencies, because currently Scalar and Tensor both depend on each other (so we introduce a TensorBase). Also, this API isn't particularly useful publicly: we can't autograd through Scalars, so you still need a Tensor overload basically everywhere anyway.
I'm undecided what the final API should be here. We could keep a Tensor constructor on Scalar, but have it generate a local scalar; this is convenient but given this API used to be non-synchronizing, it may not be the best.
For now, I'm just using _local_scalar, which is clear, although we should get rid of the prefix _ if that's the API we intend to promote.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10852
Reviewed By: ezyang
Differential Revision: D9496766
Pulled By: gchanan
fbshipit-source-id: 16f39b57536b9707132a5a4d915650c381bb57db
Summary:
This commit adds the ability to insert a node with inputs, using the schema to check the inputs are valid types, fill in any default values, and perform standard implicit conversions. Since it is schema based, it will discover and use the right overload.
Constructors to `NamedValue` enable it to be constructed using `IValue` constants so it is possible to use constant values in the input list as well:
```
g.insert(aten::add, {v, 3});
```
Keyword arguments are also supported:
```
g.insert(aten::add, {v}, {{"other", t}, {"scalar", 1}});
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10198
Differential Revision: D9307252
Pulled By: zdevito
fbshipit-source-id: 644620aa85047d1eae1288383a619d50fec44d9b
Summary:
* Changes `insertConstant(g, val)` to `g.insertConstant(val)`.
* Moves SourceRange to its own file to enable it.
* Cleans up dead attribute code in schema matching and graph.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10177
Differential Revision: D9137789
Pulled By: zdevito
fbshipit-source-id: 8a73cfb01a576f02e7e4dce019be9c0a0002989d
Summary:
This PR adds strings to the ast and implements them for print statements. Strings are lifted as attributes to the print node. They must be arguments to print itself, not as an argument for an object that is passed to print. If they are encountered elsewhere a NYI exception will be thrown.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9324
Reviewed By: jramseyer
Differential Revision: D8807128
Pulled By: eellison
fbshipit-source-id: 984401ff458ed18d473c6d1bd86750e56c77d078
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9740
- Remove implicit ArrayRef -> vector conversion
- Fix 4 call sites that accidentally did an implicit expensive vector conversion but wouldn't have needed to
- Remove explicit vector conversion from 4 call sites that also didn't need to do that
Reviewed By: ezyang
Differential Revision: D8961693
fbshipit-source-id: 980da9f988083c0072497f9dbcbbf6f516fa311c
Summary:
This should prevent slow startup times, and will not report as many
errors during static initialization time which are hard to debug
ezyang apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9801
Reviewed By: goldsborough
Differential Revision: D8986603
Pulled By: zdevito
fbshipit-source-id: 440d43ab5e8cffe0b15118cb5fda36391ed06dbc
Summary:
More clang tidy cleanups in `torch/csrc`. This time:
1. `hicpp-use-equals-default` recommends `= default` instead of `{}` for constructors/destructors. This is better practice because it expresses the intent better (https://stackoverflow.com/questions/6502828/what-does-default-mean-after-a-class-function-declaration)
2. `readability-inconsistent-declaration-parameter-name` enforces that parameter names in the declaration match parameter names in the definition. This is just generally useful and can prevent confusion and bugs.
Also updated my script a little bit.
apaszke ezyang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9737
Differential Revision: D9069069
Pulled By: goldsborough
fbshipit-source-id: f7b3f3a4eb4c9fadc30425a153566d3b613a41ae
Summary:
Based on top of #9763 (first 3 commits belong to that PR). The first commits from this PR are "Stop using attributes ..."
I tried to separate the changes into fairly meaningful commits. I can't split them up into smaller PRs, because everything starts working and all tests pass only after the whole sequence, but hopefully this will make reviewing somewhat easier.
Known issues/regressions/future tasks:
- `aten::lerp` and `aten::clamp` are no longer fusable
- `CreateAutodiffSubgraphs` needs a rewrite
- It is much more strict now, and will miss a lot of opportunities, especially when viewing ops are involved. Our previous approach was "ignore the assumption on shape availability in gradient formulas to determine differentiability, and hope that shape prop will be robust enough to actually deliver them before we differentiate", which obviously doesn't scale well to more complex cases. We should either work on reducing the size dependency of grad formulas (feasible e.g. for `view`/`reshape`, unfeasible for `squeeze`/`unsqueeze`), or make `CreateAutodiffSubgraphs` integrate some kind of "I could integrate this node into an AD subgraph, but will I be able to infer the shape of its input" reasoning (kind of like a limited shape prop, that doesn't infer anything, and only tells if it *could* infer something).
- It sometimes creates constant-only (or constants + one node) graphs, which is useless
- Broken `aten::add` in auto-batching, because it gained a non-tensor input. I changed the test for pointwise operations to use `aten::mul` instead, but I needed to disable the LSTM cell test. I'm not sure how scalar constants should be implemented in this case, because I don't fully understand our format. cc: ChunliF
- Graph import does some hacks to recover type of constants. This code should be removed once we'll gain the ability to export the IR along with value types.
- There's still a fair amount of dead code that can be removed. I didn't want to make this diff any bigger, and removing it is an easy task.
- Graph fuser could be improved to use signature matching (possibly using `OperatorSet`) instead of basing on node kinds.
- Manual constant propagation for the `ListConstruct` node in `torch/onnx/utils.py` should be replaced with a proper constant propagation pass (or we should ensure that the one we have handles at least this case before we remove this code).
zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9807
Reviewed By: ezyang
Differential Revision: D9004285
Pulled By: apaszke
fbshipit-source-id: fe88026a765f6b687354add034c86402362508b7
Summary:
Follow up task of #9584.
Commit 1:
- change expect/cast to return shared pointers instead of raw pointer
- isSubtypeOf accept TypePtr instead. Use `x->isSubtypeOf(NumberType::get())` rather than `x->isSubtypeOf(*NumberType::get())`
Commit 2:
- to address enable_shared_from_this pitfalls, we make the constructor private and expose the factory method to make sure user can only create it using our factory method.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9786
Reviewed By: zdevito
Differential Revision: D8980441
Pulled By: wanchaol
fbshipit-source-id: e5c923fc57a701014310e77cf29985b43bb25364
Summary:
I got some tensor->variable conversion exceptions from `torch/csrc/autograd/variable.h`, which used the `TORCH_ASSERTM` macros instead of `AT_CHECK`, so they didn't have backtraces. This was such a substantial loss for debugability that I decided to update the whole codebase to use the backtrace-enabled ATen macros instead of `TORCH_ASSERT` and `JIT_ASSERT`, the latter having been an alias of the former.
ezyang apaszke zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9575
Differential Revision: D8924566
Pulled By: goldsborough
fbshipit-source-id: 7a4013b13eec9dbf024cef94cf49fca72f61d441
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9718
This patch switches the interpreter to use IValue's primitive numbers rather than tensors for computing on integers and floats. In addition to preparing the interpreter for first-class support of other types, this cleans up the handling of primitive numbers, making it possible to just use the normal operator overloading dispatch to find the right implementation for numbers. As a result of this change, a lot of other functionality needed to be updated since it was the first time we use non-tensors in a lot of places in the code base.
Notes:
* Fixes code_template.py so that multi-line strings are indented correctly when used on a standalone line
* Cast operators (`int(x)`) now are functional. Some tests have addition conversions to integers because
we no longer allow implicit tensor -> integer conversions following the same convention as in python
* prim::ListConstruct/createList has been added to the interpreter for creating lists and this has
replaced aten::stack for integers lists
* gen_jit_dispatch.py has been refactored so that non-tensor types use operators on IValues to extract
the primitives
* IValue gains a .to<T> method that is the equivalent of tensor_as but for IValue instead of at::Tensor
* `constant_as<T>` is switched over to using IValues's `.to<T>` method, to make conversion from constant->IValue->C++ type
more consistent. This functionality combined with `toIValue(Value*)` replaces the `tensor_as` and `as_tensor` family of functions.
* conditional expressions (if, loop) and operators related to them are now computed on integers rather than tensors
* IValue gains constructors for constructing from at::Scalar and converting to it. However, IValue itself will always store
the scalars as a double or int64.
* To align with python 3 syntax, TK_INT, TK_FLOAT, and TK_BOOL have been removed from the parser, and int/float/bool are just treated as special identifiers in the compiler,
along with print. These are represented as special sugared values with a `call` method implemented. For int/float/bool this implements casting behavior.
* Dropped shared_from_this from Type/Module. They were not needed and they making debugging harder because they internally throw/catch exceptions.
* Shape propagation has been updated to support running nodes that include floating point primitive types, this required some refactoring of internal functions.
* TensorToNum and NumToTensor have actual implementations as operators now
* regster_prim_ops now contains implementations of math operators for float/int primitive types, and for mixed (prim <+> tensor) versions. This removes the need for special handling in compiler.cpp
* Primitive math is now entirely handled by letting the compiler choose the right overloads. This removes tons of special casing in the compiler.
* incorporates eellison's change to allow casting from return values. Due to the addition of primitive support, the code need slight modifications, so I just pre-merged it here.
* stack.h gains generic vararg versions of push/pop that know how to convert to/from C++ types:
```
at::Tensor a;
at::Scalar b;
pop(stack, a, b);
at::Tensor c = a + b;
push(stack, c);
```
apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9584
Reviewed By: apaszke
Differential Revision: D8910546
Pulled By: zdevito
fbshipit-source-id: 0f3e60d4d22217f196a8f606549430e43b7e7e30
Summary:
**REVIEW LAST COMMIT ONLY**
As discussed in our yesterday's meeting. Nodes can be now matched to particular overloads using the `matches(...)` function:
```cpp
n->matches("aten::type_as(Tensor self, Tensor other) -> Tensor")
```
This also changes the shape prop and peephole passes to use those functions for matching. This fixes a few bugs, makes them much more robust, and prepares us for removal of attributes.
zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9567
Reviewed By: zdevito
Differential Revision: D8938482
Pulled By: apaszke
fbshipit-source-id: eb2382eeeae99692aada2d78d5d0c87c8ef1545e
Summary:
This PR adds machinery to cache the schema in an IR node, and allows lookups of (possibly) constant inputs by their names (instead of position). The new methods are:
- `at::optional<T> get<T>(Symbol name)` - if the argument called name is a constant, then casts it to type `T` and returns it. If it's not constant returns `nullopt`. Raises an error if there's no argument with that name.
- `at::optional<IValue> get<T>(Symbol name)` - like above, but packs the result in an IValue
- `Value* getValue(Symbol name)` - retrieves a `Value*` for an argument (no need to know its position).
All above functions currently inspect the attributes as well, but that's only so that I could start using them in other places in the JIT without disrupting our current functionality. I wanted this diff to be a preparation that doesn't change the semantics too much, and so both the tracer and script create nodes with attributes. The next PR will put that to a stop, and hopefully the changes we need to make to other components will be simpler thanks to what I did here.
One more thing I'd like to do before actually stopping creating the non-attributed nodes is to have a convenient way of creating a schema programmatically, matching nodes against it, and creating them without having to pack inputs into flat argument lists (which is quite error prone).
zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9505
Reviewed By: ezyang
Differential Revision: D8915496
Pulled By: apaszke
fbshipit-source-id: 39d14fc9a9d73d8494f128367bf70357dbba83f5
Summary:
Closes https://github.com/pytorch/pytorch/pull/9057
Make the `_C` target depend on the `csrc-no-python` target. Also removes the `csrc` target and the with-python version of autogradpp (which is not used). Let me know if we should pick better names here.
I also ran into a nasty linker issue with only one symbol being undefined. It turns out had been given inline linkage in the `.cpp` file, which I believe is an error.
Reviewed By: orionr
Differential Revision: D8705750
fbshipit-source-id: 8de083e371dbf5e9f12c15572d88e1c595dfa087
```
JIT_ASSERT(v->setUnique(x)->uniqueName() == x);
```
This works by changing any other value in the graph with name x to a
different name. This mirrors llvm behavior and is useful when you
want to ensure some names have particular values.
The long-term fix is to remove the handling-creating pathways and
remove all the modes from PythonOp making it into an op that simply
calls a PyObject. Right now ONNX expects PythonOp to hold a
nn.Function, not a generic callable, so completely removing the legacy
pathway will also require changes to how ONNX symbolics are found.
* Allow tuples to be re-assigned
This commit improves our support of tuples by making them more first-class.
In particular, it allows tuples to be re-assigned across loops and ifs.
It does this by making them first-class values in the Graph IR, and then
removing the tuples in a LowerTuples pass.
An alternative approach would have added more support for desugaring tuples
in the Environment object as they were emitted. Instead,
the current approach was chosen anticipating a future when tuples are
fully supported (including the interpreter). In that future, the current
code can be completly reused with the LowerTuples pass just becoming
a optimization that removes unneeded tuple allocations.
* Namespaced symbols
- Our interned strings now have structure, "ns::symname" rather than just
"symname" before. We support efficient namespace testing for uniques
by encoding the namespace in one byte in the Symbol internal representation.
See torch/csrc/jit/interned_strings.h for a more in-depth implementation
discussion.
- All uses of ksymbol are now attr::symbol (or some appropriate namespace).
The valid namespaces are prim, attr, onnx and aten.
- Symbol is bound in Python as a qualified string "attr::symbol", EXCEPT for the
attribute setting/getting API, whose symbols must always be attr
symbols; they get special cased to assume strings are passed.
There's a little bit of naughtiness in the implementation, maybe you know
how to solve it.
- However, the g.op() convenience function assumes that you're generating
ONNX operators, unless you explicitly qualify.
- All ATen operators and nodes have built-in interned strings generated
for them, so you should never have to write a string literal ever again.
The tracing code is adjusted to use it.
- ONNX exporter now properly tests to see that all operators are in
onnx namespace before accepting the export. This is way more
robust than the previous exporter, which would be willing to
export capitalized operators which were not actually ONNX operators.
- A slight organizational change for symbolic.py; this module now ONLY
contains aten operators. In particular, the exporter for Constant
has moved into utils.py (along with Undefined, from the C++ side),
since primitive ops get "special treatment."
- The un-inplacing logic in recording is more robust, so that we don't
delete a trailing underscore from __and__. This never affected us
before because we didn't have any tests for it.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
This PR adds the possibility to build the C++ parts of autograd and jit, with no dependency on Python.
The goal is to allow taking a PyTorch IR representation (a tree s-expr) and running it with provided inputs.
Prerequisite: build PyTorch so that codegen runs once.
Instructions:
cd tools/cpp_build
bash build_all.sh
This will build libtorchjit and torchjit_test in tools/cpp_build/build/torchjit-build. The latter basically runs the code in test_jit.cpp for now.
While writing the PR, it turned out that a few of Python.h includes were redundant. They were removed here (PyTorch tests still pass on my machine, we'll see CI).
* Introduce Python-free builds of autograd and jit
* Remove NO_PYTHON ifdef in functions/special
* Add Python function calls to script
* Script compiler gains a `Resolver` object that runs when it does not understand a function call. This decouples the python resolution from the conversion to IR.
Support shape propagation with control-flow
* This allows us to enable optimization in the GraphExecutor for most
script tests.
* Changes Type to always be present (non-null) on a Value, removing `hasType()`
and `typeOption()`. A new type kind 'DynamicType' now represents when
a specific type has not been determined.
* If/Loop nodes propagate shapes/types in the simple cases where types of
outputs do not change depending on where control flows. In other
cases, we propagate DynamicType to indicate we do not know what
the shape will be.
* Remove the `cond` input to the body of Loop to simplify handling in
interpreter and shape propagation.
* Bugfix for zero-dim contiguousStridesOf
* Use stacks in the interpreter/aten_dispatch
Rather than have separate input/output lists,
the interpreter now works using a single stack.
Operators in the interpreter push/pop from the stack.
This allows ownership of tensors to transfer directly to an operator,
and an operator can drop the reference to a tensors as soon as it is
no longer needed. This is important for the GraphExecutor op,
which recursively runs the interpreter.
Once autograd is updated to pass variables to Function by value,
we will be able to ensure that we release ownership as soon as possible.
This commit also switches the interpreter to use a fake
tensor 'ContainerTensor' rather than at::Retainable to hold non-tensor
data in the interpreter. This allows us to use std::vector<at::Tensor>
for all registers, which is significantly less confusing than the
OwnedRetainables struct it was replacing.
* Add If and Loop to interpreter
* Preprocess loop to calculate where references to tensor should be dropped
* Add control instructions JumpZ/JumpNZ/Jump
* Switch from explicitly having stage structs to having a single list
of instructions with Store/Load instructions to take values off the
initial stack
* Make the interpreter tests executable rather than use expect files
* add a flag to interpreter code so that constants are variables
if the interpreter is running on variables.
* Add tensor_as to its own file
This commit is getting the IR ready for representing ONNX control flow.
It adds nested blocks to the IR.
* Each node now has blocks(), addBlock(), and eraseBlock() similar to a node's
output list.
* Blocks are a property of every node rather than an attribute because
to make it easier to manage the lifetime of the containing nodes and because
the behavior of cloning Blocks will likely be different from the way we clone other
attributes.
* A block itself has a list of nodes, as well as inputs and outputs.
The meaning of the nested input/output nodes are specific to the particular
node kind containing the block. It is safe to assume inputs to a block will be
in scope in the block.
* Each Block has an owningNode() and each node has an owningBlock().
The owningNode of the top-most block is null.
* Values are lexically scoped: nested blocks can use values from outer blocks
that have been defined in previous nodes. Lint has been updated with these
new scoping rules.
* This change preserves almost all of the pre-Block API. No attempt has been made
to make optimizations aware of Blocks. This will need to be done on a case-by-case
basis as we make optimizations capable of handling Blocks.
This adds the initial implementation of graph executor for the new JIT design. It includes a few python tests ensuring that nograd, backward, and double-backward cases work for simple examples and some corner cases. More work needs to be done to performance optimize as there are many extra copies and places where we hold onto variables longer than we should. These are noted in the comments.
This pass splits differentiable subgraphs into their own Node,
similar to a fusion group.
This initial implementation does not create optimal subgraphs, but
it works well in the case where most things are differentiable,
and has the building blocks (`mergeNodes`) to extend to the
better implementation.
Previous Symbol was just a uint32_t and we converts symbolToString and
stringToSymbol. Now Symbol is a struct with a toString method, and
constructors from either BuiltinSymbols enums (e.g. kParam) or strings.
Symbol is convertible to a uint32_t to ensure it can still be used in
switch statement BuiltinSymbol case branches.
cuModuleLoad is only valid for a single device so we need to
compile for the particular device that the fusion group will run on.
CompiledFunction already specializes different traces for tensors,
so we just need to have fusion_compiler produce the cuFunction on
the right device.
* Add interpreter support for Handles/PythonOp/CppOp
This treats Handles as a first-class type in the interpreter
since this turned out to be conceptually simpler than treating
them as a separate concept, which requires a second channel for
register allocating and moving data from one op to the next.
Notes:
* The refcounting nature of tensors is factored into its own base type
so that it can be shared with other refcounted types such as handle.
* Some methods redundant with TensorBase have been deleted from Tensor
* The interpreter uses raw refcounted handles. In addition to being
able to treat Tensors and Handles as the same base object, it removes
a lot of redundant refcounting as objects moved from tensors to input/
output lists.
* aten_dispatch has been updated to work directly on the raw refcounted
lists to avoid refcounting and duplicate lists.
* Removing jit_closure.cpp, The interpreter can now handle all pathways.
* Functions like `unsafeToTensorShare` describe how
ownership transfers in the interpreter. The `Steal` variants
take rvalue references as arguments, and invalidate those
arguments to prevent potential problems.
* Make TensorTemporary is not a subtype relationship because it is too easy to
do something horribly unsafe:
```
void foo(at::Tensor bar) {
// bar destructor call release on a temporary!
}
foo(TensorTemporary(retainable)); // structure slicing!
```
This commit adds a Value type similar to the one @ezyang suggested a while
ago for handling multi-return nodes.
Previously if we had a graph like:
a = op1(b)
c, d = op2(a)
Then its in-memory format would look like:
%0 = op1(b)
%1 = op2(%0)
%2 = select(%1, 0)
%2 = select(%1, 1)
Select nodes were used only to handle the multi-output case. In the
single-output case ops referred directly to their uses.
This required special handling for the single- and multi- output cases,
and was confusing when used with ONNX which distinguishes values (the
inputs/outputs of a node) from the nodes themselves (e.g. a Conv).
This commit adds the Node/Value distinction to the IR. In the example
above, `a`, `b`, `c`, and `d` are now Value objects, while `op1` and
`op2` are now Node objects. Inputs/Outputs to the graph are values.
* Nodes now always have multiple outputs, accessible through their `output()`
method.
* Methods exist for adding/removing outputs from a node.
* Nodes own their output Values, destroying a node destroys its outputs and it
is only valid to destroy a node when no uses of its outputs remain.
* Unlike select, Values do not appear in the nodes list.
* The method `node()` on `Value` retrieves its defining node. Calling it
is always valid. For inputs, its kind is "Param". Like "Return" there is a single Param
node representing all inputs.
* For single-output Nodes, the method `output()` retrieves the single
output Value, asserting that the node is in-fact single output.
* Functions are the same, but some functions like `type()` have moved to
Value.
* `replaceAllUsesWith` is now sanely defined for both Values and Nodes.
In the case of Nodes, it replaces all outputs of the node with the outputs
of the replacement node.
* stage is defined both on Node/Value. This is because Inputs require a stage.
* Apart from changing data types from Node->Value most passes remain the same.
Things that previously assumed single-output nodes now have to call output()
to get the node.
* This removes the uses = [...] field in the outputs because it was
getting confusing even before this commit when uses would refer to nodes,
but we print the names of Values. The lint pass validates the use list,
so printing it out seems less necessary.
Some knock on effects:
- at() is not supported on ArrayRef. I fixed this by adding a new
overload for input() to access a specific input. I also filed
https://github.com/zdevito/ATen/pull/152
- Need new overloads for fmap/filter, because template deduction won't
attempt an implicit constructor in attempt to match the argument.
- New overload in ir.cpp for printing ArrayRef.
- When we pybind11 an ArrayRef, we convert it into an iterator.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
This started off as a minor fix based on Adam's question, "why is printing
a graph not const" and snowballed into a giant yak shaving exercise.
- The Graph and Node APIs now uniformly enforce deep constness; e.g., if you
get a const Node* or const Graph*, it is not possible to get a non-const
Node*/Graph* somewhere else in the graph (even though the member variables
of these are non-const. Hooray for private access specifier.)
- A big pile of functions got const versions, most notably the printing
functions, and functions for accessing inputs().
- REALLY IMPORTANT, BC-BREAKING CHANGE: inputs() now returns a COPY of the
inputs, rather than a reference to the underlying. I was forced to do this
because there is no way to portably turn a std::vector<Node*> into a
std::vector<const Node*>, which is necessary to provide a const-correct
version of inputs() that enforces deep const-correctness. I then justified
this choice to myself with the observation that outputs() returned a
copy (by necessity), so this makes the API more uniform.
But making this change uncovered two very subtle bugs:
1. If you change functions from returning a reference to returning a copy,
the idiom node->inputs().begin() is no longer valid, because the memory
the iterator points to immediately becomes invalid. THIS SUCKS.
Honestly, we should add a lint rule rejecting calling begin()/end() on
temporaries because this is very dangerous. To excise this pattern from
the codebase, I added begin() and end() methods to Graph, so that we got
rid of the graph->nodes().begin() idiom, which happens to be sound,
despite not returning a reference, because graph_node_list is a
non-owning reference.
2. pybind11 doesn't handle std::vector<Node*> cast out of the box.
Fortunately, I found a simple fix in the GitHub issues tracker
that involved adding an extra type converter. And yes, this
does mean that outputs() in Python never worked correctly.
- New const_graph_node_list, which is a graph_node_list that gives you const
Node*
There are some more miscellaneous improvements:
- Applied CR comment fixes on export.cpp; using replaceInput, and renaming
variables for clarity.
- assertValidInput helper method added, and applied to replaceInput
- Use an explicit function to print THPObjectPtr, otherwise we get
the wrong overload.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* update fuser to match ATen-formatted JIT ops
* fix concat optimizations and add test
* allow onnx export to work with single-export functions
* fix onnx handling of multi-return nodes.
* nits, format, vision test update
* fix add constant
* fix driver init issues
* Add missing Neg symbolic.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
The pieces:
- I improved the lint / asserts to catch some bugs which I
committed while working on my export. There are two new
properties which the linter checks now:
(1) "Anticipated uses". If a node says that is used by
M, M better appear later in the topsort. Previously,
we only checked if it was in all_nodes.
(2) If you are a select node, you better be a multi-type node;
if you're not a select node, you better not be! And you
should never have an input that is multi-type.
- There is a new peephole optimization pass, for simple, local
transformations to graphs. Right now, it implements a simple
optimization: remove 'expand' invocations that are no-ops
(the size before matches the size after), but we can add other
things to it later. I needed this for ONNX because no-op expands
show up in the left-hand argument, which we don't support.
- There is now a broadcast fuser, which fuses ATen expand ops
into broadcastable ONNX ops (Add, Div, Mul, Pow, Sub, Gemm.)
It only fuses when the original size is a suffix of the new
size, as per the ONNX spec.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
