Changes decomposition behavior of `aten.to` to respect the aliasing/non-aliasing behavior in eager, and to specialize to the input/conversion dtype & device.
Before change: we always decompose `aten.to` into `_to_copy`, regardless of aliasing behavior. This leads us to ban mutations on the result of `_to_copy` when aliased, since we can't guarantee correct program semantics. This meant users had to explicitly call `.clone()` before mutating. In the special cases where we don’t ban mutations (e.g. dtype conversion), we add runtime assertions on the input & conversion dtype/devices in the decomposed program (see https://github.com/pytorch/pytorch/pull/142420).
After change: we decompose to the aliasing/non-aliasing behavior that matches eager, allowing mutations in all cases. We also add dtype/device assertions for all `aten.to` ops, starting in the pre-dispatch graph, basically specializing the program to the dtype/devices.
Differential Revision: D71229547
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149235
Approved by: https://github.com/tugsbayasgalan
Summary: There are ~260 tests for all the corner cases of export from test_export.py. utitlizing to test sigmoid in the OSS setting.
Test Plan: buck test mode/opt caffe2/test:test_export -- -r _sigmoid
Differential Revision: D69937387
Pull Request resolved: https://github.com/pytorch/pytorch/pull/147535
Approved by: https://github.com/yiming0416
Previously, in non-strict path, we always error when trying to inplace update a constant tensor because those constant tensors are not actually wrapped by functional tensors. This is correct behaviour in torch.compile, because dynamo makes all constant tensors into buffers and AOTDispatcher just lifts them and wraps them in functional tensors. However, in non-strict, there is no such step that registers constants as buffers so AOTDispatcher panics when it sees these dangling constant tensors when functioanalizing.
Due to recent change in the IR, this is no longer an issue in non-strict path because we don't call AOTDispatcher at training IR level, but now it is a problem for both strict and non-strict when we lower to inference. (lowering to inference is very similar to non-strict tracing) As a result, we have at least one external (https://github.com/pytorch/pytorch/issues/141336) and internal issues reported due to this difference.
To fix this, there are two ways:
1. Make functionalization be aware of constant tensors and map them to functional tensors on the fly. This makes functionalization invariant uglier and could potentially open up a gate for more nasty bugs.
2. Special handle this in export. This seems more aligned with what dynamo does today so i think we should do it this way. I think the current state could benefit from more refactors to make the run_deocmpositions to be more similar to strict export (because both of them now handle this constant registerinig logic) but it is bit complicated to do it now because strict export version of this logic is also not complete because it doesn't take into account of export graph renaming pass etc). I will follow up with more refactors after this PR (T213466691) to unblock users faster.
For future reference:
Why are we not doing "turning constants into non-persistent buffers and never de-register"? The reason is because in some internal models, they rely on module.to to reliably work to move params/buffers to correct device. As a result, buffers are moved while constants are not. In composibility meeting, we agreed that export won't do device agnostic tracing going forward (it will provide a way to specify FakeTensor in CPU that can be configured to be run on GPU), so after that is done, we can always turn constants into non-persistent buffers which will simplify export's constant handling.
Differential Revision: [D68610739](https://our.internmc.facebook.com/intern/diff/D68610739)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145593
Approved by: https://github.com/avikchaudhuri
If a user passes in a namedtuple as an input, currently the input TreeSpec looks like: `TreeSpec(type=namedtuple, context=”class_fqn”, children_spec=[*, *])`
The user then saves the program containing this input TreeSpec. But what happens if they load it in a new environment where `class_fqn` now contains an additional field?
This means that the exported program is now expected to take in another input. But since those fields were not used in the original program, users should be able just drop those additional fields and the program will run successfully. This is needed/used in APS where they use unflattener's adapter to adapt the inputs based on the previously saved treespecs.
There are a couple of [solutions](https://docs.google.com/document/d/1V4ZSdy-8PUISWc8RqvGu3DU01BVegJhHHPWqa1Io7Eg/edit?tab=t.0) for how we can address this, but eventually we settled on saving a side table mapping namedtuple types to their list of field names, which can then be accessed by the adapter.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145956
Approved by: https://github.com/zhxchen17
Summary:
Add experimental support for torch.nn.Module as input types.
Before this change, we don't support module inputs but recently we saw some interesting use cases like gpt-fast https://github.com/pytorch-labs/gpt-fast/blob/main/generate.py#L68 where we directly pass in a module input for different variants of the same models.
Since we don't really care about non-param or non-buffer states in non strict mode, we don't care about those either and pretend they are like plain constants during tracing. We treat any module input like a nested container of tensor, and each time we will automatically register a pytree handler for these module types to flatten its state dict into a group of tensors. We will just inline any module method call during tracing like we did for `self` module in export_for_training. This will make input modules' behavior very similar to the training module in typical case, except that we don't record the inputs as parameter or buffers but rather just plain user inputs.
Test Plan: buck run mode/opt caffe2/test:test_export -- -r test_module_input
Differential Revision: D67680827
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143925
Approved by: https://github.com/tugsbayasgalan
Summary:
A reland of https://github.com/pytorch/pytorch/pull/142426.
Copying the description over here:
For torch.export (strict and non-strict), we don't do functional decomposition. Instead, we preserve the custom triton ops as custom ops. This is because we want the exported program to be high-level and serializable.
The alternative:
If we decompose the custom op to a functional hop and make it a node in exported program, we need to figure out ways of serializing the hop and its arguments, which can be triton.jited python functions and triton dtypes. This is undesireble because:
it can be tedious to maintain layer that serialize the jited function (e.g. with a string) and dtypes.
changes to triton or the serialization logic for triton arguments can be BC breaking
exported program will expose the implementation detail (i.e. triton source code) for a specific backend (GPU) to users, which mixes levels of abstraction.
Future plans:
After this PR, in the short term, we expect users to have a seperate aot_compile stage that compiles the exported program into a Cubin file on the same machine that users call export, which does autotuning and removes triton dependency and serve the model with Cubin. This guarantees that triton changes won't break BC.
In the long term, we may export multiple cubins for the triton op directly.
Test Plan: see new tests.
Differential Revision: D67879685
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144284
Approved by: https://github.com/zou3519
Over time, a large number of the existing type ignores have become irrelevant/unused/dead as a result of improvements in annotations and type checking.
Having these `# type: ignore` linger around is not ideal for two reasons:
- They are verbose/ugly syntatically.
- They could hide genuine bugs in the future, if a refactoring would actually introduce a bug but it gets hidden by the ignore.
I'm counting over 1500 unused ignores already. This is a first PR that removes some of them. Note that I haven't touched type ignores that looked "conditional" like the import challenge mentioned in https://github.com/pytorch/pytorch/pull/60006#issuecomment-2480604728. I will address these at a later point, and eventually would enable `warn_unused_ignores = True` in the mypy configuration as discussed in that comment to prevent accumulating more dead ignores going forward.
This PR should have no effect on runtime at all.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/142325
Approved by: https://github.com/Skylion007, https://github.com/janeyx99
Differential Revision: [D65362160](https://our.internmc.facebook.com/intern/diff/D65362160)
State after this IR:
1. For the tests that require inference IR, they are replaced with ep.run_decomp({}) so export_for_training_run_decomp is sort of redundant but i guess it is still nice that multiple round of retracing still working. In general, we need some auditing to reduce our redundant testing coverages.
2. After this PR landed and not get reverted for a week or so, i will replace the export_for_training calls with export as they are the same thing now.
3. Added more tests to also cover now "deprecated" old IR by patching export to use old export. For reviewers, please look at the internal version.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139511
Approved by: https://github.com/ydwu4, https://github.com/angelayi, https://github.com/avikchaudhuri
During export, we nub out most CIA ops to return NotImplemented to avoid decomposing them during tracing. To recover the existing shape propagation behavior, we register these CIA decomps directly as FakeTensorMode rules as well. The reason we have to do is because when we return NotImplemented, FakeTensor would fallback to running these CIAs with Meta backend causing device branching CIA ops to fail. (because now the device is Meta. One example is sdpa). If we register a kernel directly to FakeTensorMode, we won't fallback to Meta backend.
Differential Revision: [D65716260](https://our.internmc.facebook.com/intern/diff/D65716260/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140465
Approved by: https://github.com/bdhirsh
# Why?
I want the following code to work.
minimal repro:
```
class M(torch.nn.Module):
def forward(self, dilate_flag):
return dilate_flag.item()
input1 = (torch.tensor([1], dtype=torch.bool, device="cuda"),)
model = M().cuda()
ep = torch.export.export(model, input1, strict=True)
path = torch._inductor.aot_compile(ep.module(), input1)
aot_model = torch._export.aot_load(path, device="cuda")
actual_output = aot_model(*input1)
```
error: AssertionError: Encountered an unsupported object of type <class 'torch.SymBool'> while writing the metadata for exported program
second error will be handled by https://github.com/pytorch/pytorch/pull/138760
# Motivation
I could technically bypass it with a torch.int tensor. However, it doesn't work with torch.cond. I want the following to work. It would also require https://github.com/pytorch/pytorch/pull/138760 for aot compile to work.
```
class M(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
self.dilate_flag = 0
def forward(self, dilate_flag):
self.dilate_flag = dilate_flag.item()
def true_fn(dilate_flag):
return dilate_flag.clone()
def false_fn(dilate_flag):
return dilate_flag.clone()
torch.cond(
self.dilate_flag,
true_fn,
false_fn,
(dilate_flag,),
)
return self.dilate_flag
input1 = (torch.tensor([1], dtype=torch.bool, device="cuda"),)
input2 = (torch.tensor([0], dtype=torch.bool, device="cuda"),)
inputs = (input1, input2)
model = M().cuda()
for input in inputs:
expected_output = model(*input)
ep = torch.export.export(model, input, strict=False)
path = torch._inductor.aot_compile(ep.module(), input)
aot_model = torch._export.aot_load(path, device="cuda")
actual_output = aot_model(*input)
assert (
expected_output == actual_output
), f"henry they are not equal {expected_output} != {actual_output}"
```
Differential Revision: D64867504
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138765
Approved by: https://github.com/ydwu4
In this diff, i make test_torchbind.py tests to handle training IR. Today in the training IR, we don't see the effect token and HOP because this happens at the FunctionalTensorMode. Maybe in the future, we should move this logic up to the training IR so that writing passes etc on training Ir is safer. But for the migration purposes, i think it is ok for now. I also fixed two bugs:
1. ep.module() doesn't register all aliased constants in the module.
2. When we retrace, we need to fakify the original Torchbind object.
3. We don't run any DCE on training IR so we need to add some more torch ops to verifier.
Differential Revision: [D64853530](https://our.internmc.facebook.com/intern/diff/D64853530)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138658
Approved by: https://github.com/ydwu4, https://github.com/zhxchen17
As called out in https://github.com/pytorch/pytorch/pull/137999, preserving signatures of multiple calls when buffer mutations are present was NYI. The main problem was that intermediate values of buffers were not tracked, so couldn't be propagated statefully between multiple calls (i.e., they would need to be explicitly passed around, defeating the unlifting needed for preserving signatures).
This PR fixes this situation, by introducing module attributes that carry the necessary intermediate values of buffer mutations. In general, a buffer mutation can have several intermediate values it depends on recursively, even other buffers. So rather than tying an intermediate value with a particular buffer, we tie it with the submodules that create and read it. We install an attribute on all modules that create or read a particular intermediate value, sharing the same initial storage (i.e., initialized with the same empty tensor). For the module that creates this intermediate value, we copy the value into the corresponding attribute; and for the modules that read it, we read the corresponding attribute instead.
Another complication that needed to be addressed was that a `run_decompositions` following an `export_for_training` was not preserving module call graphs, which is needed for unflattening and, in particular, used when remapping inputs. Fortunately some existing metadata already tracks provenance of nodes, which we could use to update a module call graph after functionalization / decomposition.
Differential Revision: D64806175
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138669
Approved by: https://github.com/tugsbayasgalan
In this PR, we implement lazy dictionary for export decomp behaviour for following reasons:
1. Custom op loading can happen after import time, as a result, the decomp table might not be able to pick up the decomp. Therefore we try to delay materialization as late as possible.
I intentionally seperated out the core_aten_decomp to not have any custom CIA ops in this PR to mitigate the risk of getting reverted but in the future, core_aten_decomp under torch/_decomp will exist as an alias to official export table (torch.export.default_decompositions)
Differential Revision: [D64140807](https://our.internmc.facebook.com/intern/diff/D64140807)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/137650
Approved by: https://github.com/justinchuby, https://github.com/bdhirsh
Summary: Title
Test Plan: CI
This fixes some breaking tests in executorch. I think the root cause is when we have aten::matmul which we are not preserving, we register meta implementation from C++ side. It seems like the C++ kernel doesn't work well with mix of FakeTensor and real tensor. This PR sidesteps this problem by always preferring python CIA decomp over C++ Cia decomp
Differential Revision: D63297050
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136492
Approved by: https://github.com/bdhirsh
In this PR, we deprecate _preserve_ops feature in run_decomposition API. We can't kill this API completely because Executorch team depends on it. As the syncing between two repos is non-trivial, I just leave this argument as deprecated for now. In the next PR, i will immediately remove it.
After this PR, run_decompositions will only decompose what's inside the decomp table and preserve the rest by default. Note that this feature is only rolled out to OSS for now. Old code path is protected under IS_FBCODE flag.
Differential Revision: [D62163161](https://our.internmc.facebook.com/intern/diff/D62163161/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135080
Approved by: https://github.com/justinchuby, https://github.com/avikchaudhuri, https://github.com/bdhirsh
Summary: When we are placing nodes in the graph, we should also replace the references in module_call_graph.
Test Plan:
buck2 run 'fbcode//mode/opt' torchrec/fb/ir/tests:test_serializer -- --filter-regex test_serialize_deserialize_vlea
buck2 test 'fbcode//mode/opt' fbcode//torchrec/fb/ir/tests:test_serializer -- --exact 'torchrec/fb/ir/tests:test_serializer - torchrec.fb.ir.tests.test_serializer.TestSerializer: test_serialize_empty_value_vlea' --run-disabled
buck2 test 'fbcode//mode/opt' fbcode//torchrec/fb/ir/tests:test_serializer -- --exact 'torchrec/fb/ir/tests:test_serializer - torchrec.fb.ir.tests.test_serializer.TestSerializer: test_deserialized_device_vle' --run-disabled
Differential Revision: D62014035
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134830
Approved by: https://github.com/angelayi
Summary:
With training IR, we cannot rely on trapping `to()` in `FunctionalTensor` because the regular decomposition kicks it first, and that can cause it to be optimized away.
So instead we preserve it until we functionalize, and then replace it explicitly with `_to_copy()`.
Test Plan: expected test failures go away
Differential Revision: D61883878
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134622
Approved by: https://github.com/zhxchen17, https://github.com/tugsbayasgalan
Summary:
In export, we will generate many redundant getitem nodes branching from the same source, inserted by runtime assertions or any passes. This is causing issues with any downstream system relying on any value being uniquely defined by a single node.
I don't think it hurt to remove a bunch of getitem nodes only, so I just added to the ctor.
Test Plan:
rebase on D61256937
```
buck2 run scripts/bearzx:pt2_export_playground
```
Differential Revision: D61351578
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133618
Approved by: https://github.com/tugsbayasgalan
Summary:
A re-land of D60006710.
Fixed TrainingIRToRunDecomp failures for test_tensor_attribute_zero_args and also a few re-tracability failures because run_decomposition does a retracing.
edit: also remove the eliminate_dead_code() in _unlift because of one onnx test failure:
a constant tensor attr was lifted as constant_tensor input but it's not used in the graph after aot_autograd due to a short cut in its decomposition. This causes the setattr to be removed by eliminate_dead_code but the graph signature still contains the name of that buffer, which causes an inconsitency between the transformed graph and ep's original signature after _unlift. And it seems that this has happened a few times where some nodes are accidentally removed and we're in an inconsistent state.
The alternative of removing it would be: every time we call elimiate_dead_code, we verify the consistency of the graph with 1. the graph before transformation and 2. all the meta datas but i think this deserves a complete design
edit 2: Also fix the inconsistency of graph signatures when param_constant is marked as lifted_tensor_constants but it's registered as parameters in the output of ep.module().
Differential Revision: D60532628
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132307
Approved by: https://github.com/zhxchen17
Summary:
Dynamo doesn't track whether buffers are `persistent`. This led to some ugly code where we would mark buffers as always persistent when creating signatures, then later check whether the buffers were not in the state dict to infer whether they were non-persistent, and use this to fix up the signature.
This PR instead defines a utility to look up all the non-persistent buffers registered inside a module (this information is recorded in a private `_non_persistent_buffers_set` module attribute), and uses it to (a) correctly set the persistent flag on buffers when creating signatures (b) transfer this information to a Dynamo-traced graph module, which then causes non-persistent buffers to (correctly) not show up in the state dict.
Test Plan: existing tests + new case with non-persistent buffer in nested module
Differential Revision: D60224656
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131756
Approved by: https://github.com/zhxchen17, https://github.com/ydwu4
