Adds support for SymInts in the FakeTensor cache.
A couple notes:
1. When a SymInt is present in the input key for a FakeTensor operation we cache on the ShapeEnv instead of using the FakeTensorMode cache. This is necessary so we don't have to remember and check the guards. It reduces the cache hits but there's diminishing return on how much work we can do before the cache becomes more of a burden than a gain.
2. We need to be careful that when we cache an output SymInt that is a direct copy from the input that when we have a cache-hit we copy the SymNode from the input to the output. This is important because the fx-graph building code actually uses SymNode ids in the process of building the graph so constructing a same-content-but-different-id SymNode will fail.
3. In the cache key we store SymInts as a _PySymInputStub. These represent SymInt (and friends) but support `__hash__` and `__eq__` (which SymInt do not).
4. In the cache entry we store SymInts as a _SymIntOutputStub.
Perf example:
```
python benchmarks/dynamo/timm_models.py --ci --accuracy --timing
--explain --inductor --dynamic-shapes --dynamic-batch-only --device cuda
--training --amp --total-partitions 2 --partition-id 0 --output
/tmp/training_timm_models.csv --filter crossvit_9_240
```
fake tensor cache before:
```
INFO: FakeTensor cache stats:
INFO: cache_hits: 68137
INFO: cache_misses: 837
INFO: cache_bypasses:
INFO: symbolic shape: 48224
INFO: CompositeImplicitAutograd: 917
INFO: non-fake tensor: 70
INFO: non-FakeTensor output: 62
INFO: non-builtin: 8
INFO: dynamic output shape: 1
```
and after:
```
INFO: FakeTensor cache stats:
INFO: cache_hits: 88187
INFO: cache_misses: 14233
INFO: cache_bypasses:
INFO: CompositeImplicitAutograd: 1037
INFO: non-FakeTensor output: 602
INFO: non-fake tensor: 70
INFO: unsafe view: 36
INFO: non-builtin: 8
INFO: dynamic output shape: 1
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127596
Approved by: https://github.com/eellison
ghstack dependencies: #131014, #129780
When handling an input to dynamo that's a view of a subclass, dynamo does some handling to reconstruct the view. Part of this is to construct symints for the input parameters to the view.
Previously, the code would just call `create_symbol()` which by default specifies a _positive_ symint (>= 0); this fails in the case where you have an aten::view that was called with a -1.
Fix: just specify `positive=None` when calling `create_symbol()`, to avoid restricting the symint to >= 0 or <= 0.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128662
Approved by: https://github.com/jbschlosser
This adds dumps of MetaTensorDesc and MetaStorageDesc to structured logs
when they are triggered from Dynamo. The logs look like this:
```
V0522 08:13:25.267000 140224882566144 torch/_subclasses/meta_utils.py:195] {"describe_storage": {"id": 0, "describer_id": 0, "size": 32}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
V0522 08:13:25.267000 140224882566144 torch/_subclasses/meta_utils.py:220] {"describe_tensor": {"id": 0, "ndim": 1, "dtype": "torch.float32", "device": "device(type='cpu')", "size": [8], "is_leaf": true, "stride": [1], "storage": 0, "view_func": "<built-in method _view_func_unsafe of Tensor object at 0x7f882959e840>", "describer_id": 0}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
V0522 08:13:25.268000 140224882566144 torch/_subclasses/meta_utils.py:1594] {"describe_source": {"describer_id": 0, "id": 0, "source": "L['x']"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
```
The `describer_id` is used to disambiguate ids. We expect it to be
unique per frame id, but if there is a bug it possibly is not. Note you will get
redundant dumps when evaluation restarts.
tlparse can use this to give a visualization of input tensors to a
model, you could also use this to generate example inputs to run graphs
on.
Some care is taken to avoid redumping the tensor metadata multiple
times, which would happen ordinarily because AOTAutograd refakifies
everything after Dynamo, to deal with metadata mutation.
Partially fixes https://github.com/pytorch/pytorch/issues/126644
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126879
Approved by: https://github.com/jamesjwu
The `usort` config in `pyproject.toml` has no effect due to a typo. Fixing the typo make `usort` do more and generate the changes in the PR. Except `pyproject.toml`, all changes are generated by `lintrunner -a --take UFMT --all-files`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127122
Approved by: https://github.com/kit1980
Internal xref: https://fb.workplace.com/groups/6829516587176185/posts/7228787720582401/
There a few improvements here, which luckily fix some xfails:
* In generally, it can be unsafe to call operations on Tensors under a `no_dispatch()` mode that is purely trying to disable ambient modes, because this ALSO disables tensor subclass handling. So we test to see if there is a tensor subclass and don't propagate real tensors if that's the case. Another acceptable outcome might be to try to only disable the ambient fake tensor mode, this would help us propagate real tensors through more exotic tensor types, but I'm not going to do it until someone asks for it.
* We're graph breaking for wrapped tensors too late. Pull it up earlier so we do it before we try to muck around with the real tensor.
* I noticed that occasionally when I do `storage.copy_(real_storage)`, the sizes mismatch. Careful code reading suggests that I should just copy in the real data when the tensor was initially allocated, so that's what I do now, eliminating the need for a storage copy.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126281
Approved by: https://github.com/Skylion007
Internal xref:
https://fb.workplace.com/groups/6829516587176185/posts/7211398545654652/
Previously I did it in a crappy way using clone_input in the callback,
but this results in tensors that don't have quite the same
size/stride/storage offset and there was an internal test case where
not having completely accurate information was causing a downstream
problem in propagation. So now I make real tensors as similar to their
fake equivalents as much as possible. Though... I don't bother with
autograd lol.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126175
Approved by: https://github.com/albanD
More details further down, but first a more high-level description of "how do we functionalize storage resizing"
Today, dynamo converts `param.untyped_storage().resize_(x)` calls that it sees from fsdp into a custom op, `ops.inductor.resize_storage_bytes_(x)`
So given this setup, there are 3 main cases that I think we want to handle:
(1) graph input starts with a real storage size, gets resized down to zero in the graph
(2) graph input starts with 0 storage size, gets resized up in the graph
(3) graph input starts with 0 storage size, gets resized up and used in some compute, then resized back down to 0
For case (1) we need to emit a `resize_storage_bytes_` at the end of the graph, similar to how we emit `copy_()` for data mutations.
For case (2), we need to emit a `resize_storage_bytes_` in the graph, and we **also** need to emit a `copy_()` (the input had its storage resized up, and filled in with data, which is we need to reflect as an input mutation)
For case (3), the net effect is that the input had no data on entry and exit of the function, so we don't need to emit any mutable ops in the end of the graph.
The main thing to call out is that: we need to write a functionalization rule for `resize_storage_byte_`, (`FunctionalTensorWrapper::storage_resize_()`) and this rule actually does very little. We would like to **not** emit any new ops in the graph (like say, a functional resize op). Instead, we should expect / rely on the fact that any resize up will be immediately followed by a `copy_()`/`foreach_copy_`/`out=` op, that will fill in the data of the tensor. So `FunctionalTensor` can temporarily live in a state where its data is invalid, until the `x.copy_(y)` "updates" its data with the new tensor.
So effectively, all that this rule does is:
(1) it stores metadata on the storage, indicating that the tensor was resized, as well as the updated storage size. We need this info in AOTAutograd, so it knows whether to emit a mutable resize_() op in the graph epilogue
(2) There is also a corner case: if we are resizing down to zero, but our tensor had **previously** had a zero size storage, then we update `value_` to point to the original value of the tensor. The reason this seems safe is because if we have a zero storage sized tensor `x`, and we resize it up, use it in some compute, resize it back down to zero, and use it somewhere, we would want the functional version of this code to use the original `x` after the second resize. For FSDP, this is important because we end up saving parameters (graph inputs) for backward, and we want to make sure that the thing we save (and the output to the forward graph) is the original, zero-storage-sized parameter, and not the "version 2" of the parameter after the first resize_()
I think a good order to look at changes in this PR would be:
(1) `test_aotdispatch.py` shows the 3 main cases I focused on as well as the expected functionalized graphs
(2) In `FunctionalStorageImpl.h/cpp`, I had to add a notion of "original base", and "original/curr_size". The first is so I can re-use the zero-size tensor after multiple resizes, and the second is so I can tell in AOTAutograd whether any resizes canceled each other out into a no-op
(3) FunctionalTensorWrapper.h/cpp has the new resize functionalizion rule + some extra utils
(4) `_functorch/_autograd`: the main changes in this folder were around adding the logic at trace-time to detect when we need to put a resize_() in the graph. I also have some assertions to check that any inputs that experience storage resizing will **always be in the graph** and not the opaque epilogue, and I also limited the resize_() mutation case so that you can only ever start with zero storage, or end with zero storage (you can't do e.g. `torch.ones(2).storage().resize_(3)`), and banned it on tensor subclasses
(5) `fake_tensor.py`/`meta_utils.py`: we now need to be able to fakeify tensors with zero storage, so I added a quick version of it in meta_utils.py. This also.. has ramifications for fake tensor caching that I need to fix (include the storage size on the cache key, maybe?)
------------------
This PR subsumes https://github.com/pytorch/pytorch/pull/120971.
This PR is enough to **almost** get a simple ppFSDP forward pass tracing with a functionalized resize_() properly. It also attempts to do the updated version from @jansel, where we don't have any notion of `resize_()` in the graph at all, post functionalization. It would probably be good to test it with @yf225 's FSDP changes, and see how many of the FX passes it allows us to remove. I think that in theory, it should allow us to remove all FX passes that affect the forward graph / partitioner, **except** the one that forces views to be recomputed in the backward (more details below).
There are a few things worth calling out:
(1) failed attempt at functionalizing `aten.copy_()`. I originally wanted to get a version takes these operations:
```
param.storage().resize_(all_gather_size)
param.copy_(all_gather_buffer)
out = aten.matmul(param, param)
```
and functionalizes them into:
```
out = aten.matmul(all_gather_buffer, all_gather_buffer)
```
This would involve getting functionalization to turn `x.copy_(y)` into a giant no-op that just returns `y`. Unfortunately, we can't actually do this in a reasonable way within functionalization (instead, there's a functional `aten.copy` in the graph - see the test case graph expecttest for details). Why? In order for that transformation to be safe, `x` and `y` need to have the same metadata. However, it's possible for `x` and `y` to be subclasses of different types. This is not something we can easily tell from within functionalization, and would be a layering violation. So for now I'm leaving it to downstream code to optimize away the `aten.copy` (this is already the case today, so I think inductor can handle this)
(2) The forward doesn't **actually** run successfully in this PR (see the `assertRaisesRegex` in the test). Why?
The final forward graph looks like this:
```
def forward(self, primals_1, primals_2):
_foreach_copy = torch.ops.aten._foreach_copy.default([primals_1], [primals_2]); primals_2 = None
getitem = _foreach_copy[0]; _foreach_copy = None
mm = torch.ops.aten.mm.default(getitem, getitem); getitem = None
t_1 = torch.ops.aten.t.default(primals_1); primals_1 = None
return [mm, t_1]
```
Where `primals_1` starts out as a secretly-zero-storage-size parameter, and gets resized up and back down within the forward (these are functionalized away).
Importantly, the matmul happy on the result of the `foreach_copy`, **but** the activation that we save for backward (`t_1`) is the result of transposing the **original parameter** (the zero-storage-size param). This is exactly the optimization in fsdp that allows us to have good peak memory usage.
The problem is that the min-cut partitioner decides to save `t_1` for backward. Running this code in eager breaks, because the kernel for `aten.permute(x)` is not happy when `x` has secretly-zero-sized-storage.
The real problem here is that in eager mode the `permute` kernel runs during the backward, after backward hooks have properly resized the saved activation. Here, we are running the transpose in the forward.
One option would be to turn off the checks in our view kernels and allow them to work on zero-storage-sized tensors, which feels pretty bad. Another option is to tweak the partitioner (or use one of Will's FX passes) to force the partitioner to not save views for backward, and allow the views to be recomputed in the backward. This seems kind of silly, but is also probably harmless.
(3) The backward is still broken. To be fair, this issue is pretty separable from "functionalizing storage resize calls", and can be fixed later (either by a real fix to our tracing infra, or via another hacky FX pass). More description of this problem is described at issue (8) of my PR description in https://github.com/pytorch/pytorch/pull/120971
(4) I only added support for "full graph" resizing: basically, the limited case where a param starts with zero storage size, and gets resized up and back down. I think we can add support for the graph break case, but I think we can keep that add-on separate from this PR unless we need it immediately. I also added asserts so we should fail loudly when we hit this case
(5) I have a change to FakeTensor creation when inputs have zero storage size that.. is probably ok. But I also removed FakeTensor caching on view ops, which I probably need to fix before I can land this PR
(6) I added a notion of "original_base" to `FunctionalStorageImpl`. More details are in the comments, but my rational for this was that we basically need it to ensure that autograd saves the **original**, zero-storage-sized param for backward, after resizing up and back down
(7) I had to update our eager kernels for `aten.copy` and `aten._foreach_copy`, to handle the case where the `self` argument has secretly-zero-storage. Inductor can probably generate correct code for this case, but we need these ops to work properly in this situation for the `aot_eager` backend to do the right thing
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122434
Approved by: https://github.com/jansel
Fixes https://github.com/pytorch/pytorch/issues/123298
I was also seeing some crashes in torchtrain due to dynamic shapes, even when I set `compile(dynamic=False)` (cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @wanchaol). This doesn't fix the underlying dynamic shape issues with compile + DTensor, but it does prevent dynamic shapes from leaking in.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123348
Approved by: https://github.com/ezyang
ghstack dependencies: #122502, #122751
When fakifying a grad tracking tensor, if the level is -2 (sentinel
value) we can just unwrap the grad tensor and return a fake version of
it. In this PR, we update the `assert_metadata_eq` to not compare if
the grad tensor and the unwrapped ones are leafs or not, as this may
not be always true.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122728
Approved by: https://github.com/zou3519
At a high level, the goal of this refactor was to make it so that `MetaConverter.__call__` has a straightforward code structure in three steps: (1) check if we support doing meta conversion, (2) describe the tensor into MetaTensorDesc, (3) call `meta_tensor` on MetaTensorDesc. However, this is not so easy to do, because there is a big pile of special cases for functional tensor inside `__call__`.
The primarily complication is handling the ambient functionalization state: specifically, the functorch dynamic layer stack and the Python functionalization dispatch. The old code demands that meta tensor conversion happen with this state disabled. But I discovered that when I reconstruct functorch tensors it demands that the functorch layers be active; in fact a batch tensor will have a pointer to the internal functorch layer.
I had some discussion with Richard Zou about what code structure here makes sense. In particular, one of the goals of the refactor here is that I can inflate MetaTensorDesc from an entirely different process, which may not have all of the functorch layers activated at the time we do reconstruction. So it seems to me that we should make it explicit in MetaTensorDesc that there was some functorch layer active at the time the functorch tensor was serialized, so that we could potentially know we need to reconstruct these layers on the other side. This is NOT implemented yet, but there's some notes about how potentially it could proceed. But the important thing here is we SHOULD disable everything when we run `meta_tensor`, and internally be responsible for restoring the stack. Actually, the necessary infra bits in functorch don't exist to do this, so I added some simple implementations in pyfunctorch.py.
The rest is splitting up the manipulations on tensor (we do things like sync the real tensor before describing it; Describer is responsible for this now) and I also tried to simplify the not supported condition, based on my best understanding of what the old thicket of conditions was doing. You may notice that the internal meta_tensor handling of functional tensor is inconsistent with surrounding code: this is because I *exactly* replicated the old reconstruction behavior; a further refactor would be to rationalize this.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122202
Approved by: https://github.com/zou3519
Fixes https://github.com/pytorch/pytorch/issues/121085
This PR pretty involved so pay attention to this description. At a high
level, the refactor is intended to be mechanical: anywhere in
MetaConverter where previously we took a Tensor as argument, we now take
a MetaTensorDesc, which contains all of the information that we would
have queried off of the Tensor, but placed into a separate data
structure which we can serialize or use to recreate a fake tensor in
a separate fake tensor mode in exact fidelity to the original.
However, this transformation is not always entirely mechanical. Here
is what you need to pay attention to:
- The memo table from real Tensor -> meta/fake Tensor is now broken
into two memo tables: real Tensor -> stable int id -> meta/fake
Tensor. The stable int id is needed so that when we do serialization,
we know when tensors/storages alias each other and can ensure we preserve
this aliasing upon deserialization.
The way I have implemented changes the weak reference behavior.
Previously, when either the real Tensor OR the meta/fake Tensor went
dead, we would remove the entry from the memo table. Now, this only
removes entries from one of the two memo tables. This semantically
makes sense, because the user may have held on to the stable int id
out of band, and may expect a real Tensor to continue to be numbered
consistently / expect to be able to lookup a meta/fake tensor from
this id. If this is unacceptable, it may be possible to rejigger
the memo tables so that we have real Tensor -> stable int id
and real Tensor -> meta/fake Tensor, but TBH I find the new
implementation a lot simpler, and arranging the memo tables in this
way means that I have to muck around with the real tensor to save
to the memo table; in the current implementation, I never pass the
Tensor to meta_tensor function AT ALL, which means it is impossible
to accidentally depend on it.
- When I fill in the fields of MetaTensorDesc in describe_tensor, I need
to be careful not to poke fields when they are not valid. Previously,
preconditions were implicitly checked via the conditional structure
("is this sparse? is this nested?") that is tested before we start
reading attributes. This structure has to be replicated in
describe_tensor, and I have almost assuredly gotten it wrong on my
first try (I'll be grinding through it on CI; a careful audit will
help too, by auditing that I've tested all the same conditionals that
the original access was guarded by.)
- I originally submitted https://github.com/pytorch/pytorch/pull/121821
for the symbolic shapes change, but it turned out the way I did it
there didn't actually work so well for this PR. I ended up just
inlining the symbolic shapes allocation logic into MetaConverter
(look for calls to maybe_specialize_sym_int_with_hint), maybe there
is a better way to structure it, but what I really want is to
just read sizes/strides/offset directly off of MetaTensorDesc; I
don't want another intermediate data structure.
- Some fields aren't serializable. These are documented as "NOT
serializable". ctx/type should morally be serializable and I just
need to setup a contract with subclasses to let them be serialized.
The fake_mode is used solely to test if we are refakefying with
a pre-existing ShapeEnv and we want to reuse the SymInt
directly--serializing this case is hopeless but I am kind of hoping
after this refactor we do not need this at all. view_func is not
serializable because it's a bound C implemented method. Joel has
promised me that this is not too difficult to actually expose as a
true data structure, but this is the edgiest of edge cases and there
is no reason to deal with it right now.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122044
Approved by: https://github.com/eellison
List of changes:
- Replace JVP_NESTING by torch._C._functorch.maybe_current_level()
- Remove all increment nesting functions from wrap_fx_proxy_cls
- fwAD.make_dual receives the dual_level as keyword argument
- Add jvp_increment_nesting, set_fwd_grad_enabled and dual_level context managers to dynamo
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119926
Approved by: https://github.com/zou3519
Fixes https://github.com/pytorch/pytorch/issues/121085
This PR pretty involved so pay attention to this description. At a high
level, the refactor is intended to be mechanical: anywhere in
MetaConverter where previously we took a Tensor as argument, we now take
a MetaTensorDesc, which contains all of the information that we would
have queried off of the Tensor, but placed into a separate data
structure which we can serialize or use to recreate a fake tensor in
a separate fake tensor mode in exact fidelity to the original.
However, this transformation is not always entirely mechanical. Here
is what you need to pay attention to:
- The memo table from real Tensor -> meta/fake Tensor is now broken
into two memo tables: real Tensor -> stable int id -> meta/fake
Tensor. The stable int id is needed so that when we do serialization,
we know when tensors/storages alias each other and can ensure we preserve
this aliasing upon deserialization.
The way I have implemented changes the weak reference behavior.
Previously, when either the real Tensor OR the meta/fake Tensor went
dead, we would remove the entry from the memo table. Now, this only
removes entries from one of the two memo tables. This semantically
makes sense, because the user may have held on to the stable int id
out of band, and may expect a real Tensor to continue to be numbered
consistently / expect to be able to lookup a meta/fake tensor from
this id. If this is unacceptable, it may be possible to rejigger
the memo tables so that we have real Tensor -> stable int id
and real Tensor -> meta/fake Tensor, but TBH I find the new
implementation a lot simpler, and arranging the memo tables in this
way means that I have to muck around with the real tensor to save
to the memo table; in the current implementation, I never pass the
Tensor to meta_tensor function AT ALL, which means it is impossible
to accidentally depend on it.
- When I fill in the fields of MetaTensorDesc in describe_tensor, I need
to be careful not to poke fields when they are not valid. Previously,
preconditions were implicitly checked via the conditional structure
("is this sparse? is this nested?") that is tested before we start
reading attributes. This structure has to be replicated in
describe_tensor, and I have almost assuredly gotten it wrong on my
first try (I'll be grinding through it on CI; a careful audit will
help too, by auditing that I've tested all the same conditionals that
the original access was guarded by.)
- I originally submitted https://github.com/pytorch/pytorch/pull/121821
for the symbolic shapes change, but it turned out the way I did it
there didn't actually work so well for this PR. I ended up just
inlining the symbolic shapes allocation logic into MetaConverter
(look for calls to maybe_specialize_sym_int_with_hint), maybe there
is a better way to structure it, but what I really want is to
just read sizes/strides/offset directly off of MetaTensorDesc; I
don't want another intermediate data structure.
- Some fields aren't serializable. These are documented as "NOT
serializable". ctx/type should morally be serializable and I just
need to setup a contract with subclasses to let them be serialized.
The fake_mode is used solely to test if we are refakefying with
a pre-existing ShapeEnv and we want to reuse the SymInt
directly--serializing this case is hopeless but I am kind of hoping
after this refactor we do not need this at all. view_func is not
serializable because it's a bound C implemented method. Joel has
promised me that this is not too difficult to actually expose as a
true data structure, but this is the edgiest of edge cases and there
is no reason to deal with it right now.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122044
Approved by: https://github.com/eellison
ghstack dependencies: #122018
List of changes:
- Replace JVP_NESTING by torch._C._functorch.maybe_current_level()
- Remove all increment nesting functions from wrap_fx_proxy_cls
- fwAD.make_dual receives the dual_level as keyword argument
- Add jvp_increment_nesting, set_fwd_grad_enabled and dual_level context managers to dynamo
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119926
Approved by: https://github.com/zou3519
List of changes:
- Replace JVP_NESTING by torch._C._functorch.maybe_current_level()
- Remove all increment nesting functions from wrap_fx_proxy_cls
- fwAD.make_dual receives the dual_level as keyword argument
- Add jvp_increment_nesting, set_fwd_grad_enabled and dual_level context managers to dynamo
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119926
Approved by: https://github.com/zou3519
This PR:
* Uses reified ViewFuncs to swap in fake tensors / symbolic SymInts for view replay during subclass view fake-ification
* Enables automatic dynamic on view bases -> fakeifies according to the resultant symbolic context instead of the old "all-static" approach
* Covers the following view types:
* subclass -> dense
* dense -> subclass
* subclass -> subclass
* Dense -> dense views are handled the old way via an `as_strided()` call, as it's likely there is no view func available
Differential Revision: [D54269082](https://our.internmc.facebook.com/intern/diff/D54269082)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118405
Approved by: https://github.com/ezyang
Decorates all NT tests with `@markDynamoStrictTest` to ensure we get the correct signal. Adds xfails where needed to get things passing.
Includes a fix in meta_utils.py for a bug that was breaking several python 3.11 tests. In particular, a dense tensor graph input that is a view of a strided NT would slip past Dynamo's check and break in meta-ification.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116111
Approved by: https://github.com/soulitzer, https://github.com/zou3519
ghstack dependencies: #115192
Continuation of #112185, following the design in this [doc](https://docs.google.com/document/d/1ipSxcTzEMMOAPvxP-YJlD5JBZZmIGgh8Q34ixtOUCRo).
Summary:
* Introduce `SubclassSymbolicPolicy` containing separate dynamic dim / constraint policies for the outer and inner tensors
* Expand the automatic dynamic algorithm to recurse into inner tensors and produce one of these for a subclass instance
* Maintain legacy behavior for subclasses by recursively calling `mark_dynamic()` on inner tensors *of the same dim as outer* when `mark_dynamic(outer, ...)` is called
* Addresses this: 6a86cf00ad/torch/_dynamo/variables/builder.py (L1750)
* Add `outer_size` and `outer_stride` arguments to `__tensor_unflatten__()` so that you can find out what symbols were allocated for the outer size / stride (you are expected to return a tensor that compares equal to the outer symbols)
* Signatures now:
```python
# attrs is a list of inner tensor attributes on x; inner_tensor = getattr(x, attr)
# ctx is anything useful for rebuilding the class we want to guard on
attrs, ctx = x.__tensor_flatten__()
...
# inner_tensors is a dict of {attr -> tensor}
# ctx is taken unmodified from flattening and (eventually) guarded on
# outer_size is the expected size of the output; possibly symbolic
# outer_stride is the expected strides of the output; possibly symbolic
y = MySubclass.__tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride)
# at the __tensor_unflatten__() call-site in PT2, we assert y.shape == outer_size and y.stride() == outer_stride
# the assert simplifies symbols when there are relationships between outer and inner symbols
```
* Size info needed for `NestedTensor` at least, stride info needed for `DTensor` at least
* Punting on `outer_storage_offset` because storage_offset handling is horribly broken in PT2 right now
* ~~Add new `__tensor_mark_dynamic__()` to allow overriding the behavior of mark_dynamic on a per-subclass basis~~ (booted to future work)
* ~~Add guards for tensor subclasses by calling `__tensor_flatten__()` in the guard to test equality on `ctx`~~
* Now handled in #114469
* Next PR: add TENSOR_MATCH guards on inner tensors
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114311
Approved by: https://github.com/ezyang, https://github.com/drisspg, https://github.com/voznesenskym, https://github.com/bdhirsh
Summary:
The primary problem we are setting out to solve here is fake tensor freshness. Before this PR, fake tensors after dynamo represented fake tensors *at the end* of trace, so subsequent retraces like aot_autograd would start off with fake tensors in the wrong (end result) state, rather than their expected fresh state. The solution here is to start a fresh fake mode, and re-fakify the tensors. The nuance comes from ensuring that symbols are uniformly created for the symbolic sizes and strides of the tensor.
This PR is the result of *a lot* of back and forth with ezyang and eellison. Initially, the first pass at this was not super different from what we have in the PR - the broad strokes were the same:
1) We cache source->symbol in shape_env
2) We pass policy objects around, stored at dynamo fakificaiton time, and reused for later fakification
3) We create a new fake mode for backends
(from https://github.com/pytorch/pytorch/pull/113605/files)
This is ugly, and has some layering violations. We detoured our decision making through a few other alternatives. Immutable/mutable fake tensor mode was the most interesting alternative, https://github.com/pytorch/pytorch/pull/113653, and was struck down on concerns of complexity in fake mode combined with it not covering all edge cases. We also detoured on what to do about tensor memoization returning back potentially different tensors than requested, and if that was an anti pattern (it is) we want to hack in with the symbol cache (we don't).
We went back to the drawing board here, but with a few concessions:
1) the cache for source->symbol must live outside of shape_env, for both lifecycle, and layering reasons
2) A good amount of work needs to be done to pipe policy around fake_mode and meta_utils correctly, to cover all the cases (ezyang did this)
cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng
imported-using-ghimport
Test Plan: Imported from OSS
Reviewed By: huydhn, Chillee
Differential Revision: D51566250
Pulled By: voznesenskym
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114526
Approved by: https://github.com/Chillee, https://github.com/huydhn
The primary problem we are setting out to solve here is fake tensor freshness. Before this PR, fake tensors after dynamo represented fake tensors *at the end* of trace, so subsequent retraces like aot_autograd would start off with fake tensors in the wrong (end result) state, rather than their expected fresh state. The solution here is to start a fresh fake mode, and re-fakify the tensors. The nuance comes from ensuring that symbols are uniformly created for the symbolic sizes and strides of the tensor.
This PR is the result of *a lot* of back and forth with @ezyang and @eellison. Initially, the first pass at this was not super different from what we have in the PR - the broad strokes were the same:
1) We cache source->symbol in shape_env
2) We pass policy objects around, stored at dynamo fakificaiton time, and reused for later fakification
3) We create a new fake mode for backends
(from https://github.com/pytorch/pytorch/pull/113605/files)
This is ugly, and has some layering violations. We detoured our decision making through a few other alternatives. Immutable/mutable fake tensor mode was the most interesting alternative, https://github.com/pytorch/pytorch/pull/113653, and was struck down on concerns of complexity in fake mode combined with it not covering all edge cases. We also detoured on what to do about tensor memoization returning back potentially different tensors than requested, and if that was an anti pattern (it is) we want to hack in with the symbol cache (we don't).
We went back to the drawing board here, but with a few concessions:
1) the cache for source->symbol must live outside of shape_env, for both lifecycle, and layering reasons
2) A good amount of work needs to be done to pipe policy around fake_mode and meta_utils correctly, to cover all the cases (@ezyang did this)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113926
Approved by: https://github.com/ezyang, https://github.com/eellison
Subsumes half of https://github.com/pytorch/pytorch/pull/113605
We support fakeifying an already fake tensor, which will give you a new fake tensor mirroring the same structure as the original fake tensor, which is what is needed by https://github.com/pytorch/pytorch/issues/113643 . However, when this refakeification happens, we will naively reallocate all new sizes for all of the fake tensor. This is the right thing to do if you are re-fakeifying on a fresh ShapeEnv (because you're reparametrizing the sizes or something), but if you have two fake tensor modes which are sharing a shape environment, you would actually rather just reuse the original sizes/strides/offset from the original fake tensor. This ends up being pretty simple. I recommend viewing with whitespace diff turned off.
There's some fuzz around jagged tensor handling; that code is probably not quite right, but I fixed it for this particular case in the most straightforward way.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113651
Approved by: https://github.com/albanD, https://github.com/eellison, https://github.com/bdhirsh
Currently meta_utils relies on as_strided when handling the view case (recursively meta-ify the base, and then do as_strided to simulate the view), but NestedTensor does not support as_strided today (though maybe it could?), so what we want to do instead is call Tensor. _view_func. Conveniently, _view_func IS always available for nested tensors.
A detail to note is that _view_func actually incurs a guard because it needs to perform some metadata checks to make sure the view is still valid. This PR adds Tensor._unsafe_view_func which can avoid that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112205
Approved by: https://github.com/jbschlosser
In this PR:
- Adds support for strides for jagged tensor (design doc for this coming soon)
- NestedTensor skips automatic dynamic
- Make use of @bdhirsh's subclass fakification logic by adding the __tensor_{un,}flatten__ functions.
- Additional logic for fakification: since existing subclass fakification logic does not handle the case where the outer tensor has an additional dimension. We insert one-off logic to (1) insert an extra SingletonSymInt onto the fakified NestedTensor. (2) make sure we call track_symint on both the sizes on the inner and outer tensor during guard creation.
Remaining things that are weird:
- Still need to skip some logic in meta utils for some reason (I was going to write this up more, but decided not to since we're not able to do this anyway for a immediate reason: we cannot arbitrarily compare singleton ints. For now I'm just following Brian's advise from [here](https://github.com/pytorch/pytorch/pull/109171#discussion_r1328137070) )
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109171
Approved by: https://github.com/ezyang, https://github.com/bdhirsh
Reland - the previous PR was reverted by internal with this error:
```
File "/data/sandcastle/boxes/eden-trunk-hg-fbcode-fbsource/buck-out/v2/gen/fbcode/363cd7e240f5d021/caffe2/torch/fb/trainer/data_modules/tests/__test_dataloader__/test_dataloader#link-tree/torch/__init__.py", line 29, in <module>
from ._utils_internal import _functionalize_sync as _sync
ImportError: cannot import name '_functionalize_sync' from 'torch._utils_internal'
```
I couldn't figure out why internal was unhappy with the import. One potential reason is that I see a build rule for *another* `_utils_internal.py` in the fb folder here ([link](https://www.internalfb.com/code/fbsource/[30ed85cd88409af98b7490be137aaa5dfd7afd01]/fbcode/caffe2/TARGETS?lines=444))
Rather than burn more time investigating, I confirmed internally that the error goes away if I move the util from `torch/_utils_internal.py` to `torch/_utils.py`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109518
Approved by: https://github.com/albanD
