The design of export API expects constraints to be specified on dynamic dimensions, while assuming all other dimensions are static by default. However a user who wishes to export a model may not be fully familiar with the code to plan what to specify.
This diff provides support for discovering constraints to specify. The basic idea is to take the set of generated shape guards and convert them into appropriate constraints. However, we usually generate a LOT of shape guards, and there is often a LOT of redundancy in them. Thus, we also need to simplify the guards so that our suggested constraints are concise yet capture the information content in the guards.
The algorithm for simplification uses `sympy` under the hood, but very surgically to avoid any risk of blowing up. See comments inline for a full description. Briefly,
1. We consider only univariate inequalities, and among them, solve for equalities first.
2. We substitute these exact solutions to convert multivariate inequalities progressively into univariate.
3. Remaining univariate inequalities are solved using `sympy.solvers.inequalities.reduce_inequalities`.
4. As pre-processing, we also eliminate all `//` and `%` operations to generate a set of linear congruence guards, and solve these using `sympy.ntheory.modular.solve_congruence`.
The results are quite dramatic. For example, an internal model produced several hundreds of guards with `dynamic_shapes=True`, which were pretty much inscrutable for humans. The summary contains around 30 dimensions that were specialized and 3 constraints on dynamic dimensions. The output format looks like this:
```
The following dimensions have been specialized and CANNOT be dynamic.
NOTE: Specializations will happen by default with `assume_static_by_default=True`.
L['foo']['bar'].size()[0] == 4
...
L['baz']['qux'].size()[3] == 96
The following dimensions CAN be dynamic.
You can use the following code to specify the constraints they must satisfy:
constraints=[
dynamic_dim(L['blah']['bleh'], 1) == dynamic_dim(L['blah']['bloh'], 1),
...,
2 <= dynamic_dim(L['blah']['bloh'], 1),
]
```
Differential Revision: [D44731747](https://our.internmc.facebook.com/intern/diff/D44731747/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98463
Approved by: https://github.com/voznesenskym, https://github.com/ezyang
Months ago, in order to get dynamic shapes working through to Dynamo backends, we changed the calling convention to pass fake tensors rather than real tensors as example inputs to backends. The motivation at the time was, well, backends shouldn't really be peeking at the real tensors when they are doing compilation, and so it would make more sense to hide the real tensors from backends. But there were a bunch of problems:
* This interacted poorly with our accuracy minifier design: accuracy minifier needs access to the real inputs in order to run the model and figure out what happens!
* The TensorRT backend required real inputs and we never figured out how to fix it.
* In practice, all the backends needed to detect if they were passed real tensors, and fakeify them anyway (certainly AOTAutograd does this)
* Parameters and inputs are treated non-uniformly: parameters had to be passed as real tensors, because CUDA graphs requires knowing what the actual tensors are
Furthermore, there were some more problems discovered after the fact:
* Backends may want to optimize on aspects of tensors which you cannot tell without having real tensors; e.g., alignment of the data pointer
So, this PR decides that changing the calling convention was a bad idea, and switches back to passing real tensors. There is a problem though: AOTAutograd will perform fakeification, which means that in practice backends are still going to end up with fake tensors in the end anyway. I want to change this, but this will require some work with bdhirsh's upcoming AOTAutograd export refactor.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99320
Approved by: https://github.com/voznesenskym
Months ago, in order to get dynamic shapes working through to Dynamo backends, we changed the calling convention to pass fake tensors rather than real tensors as example inputs to backends. The motivation at the time was, well, backends shouldn't really be peeking at the real tensors when they are doing compilation, and so it would make more sense to hide the real tensors from backends. But there were a bunch of problems:
* This interacted poorly with our accuracy minifier design: accuracy minifier needs access to the real inputs in order to run the model and figure out what happens!
* The TensorRT backend required real inputs and we never figured out how to fix it.
* In practice, all the backends needed to detect if they were passed real tensors, and fakeify them anyway (certainly AOTAutograd does this)
* Parameters and inputs are treated non-uniformly: parameters had to be passed as real tensors, because CUDA graphs requires knowing what the actual tensors are
Furthermore, there were some more problems discovered after the fact:
* Backends may want to optimize on aspects of tensors which you cannot tell without having real tensors; e.g., alignment of the data pointer
So, this PR decides that changing the calling convention was a bad idea, and switches back to passing real tensors. There is a problem though: AOTAutograd will perform fakeification, which means that in practice backends are still going to end up with fake tensors in the end anyway. I want to change this, but this will require some work with bdhirsh's upcoming AOTAutograd export refactor.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99320
Approved by: https://github.com/voznesenskym
Billing of changes:
* Get rid of `print_guards`; instead, you control this with `TORCH_LOGS=torch.fx.experimental.symbolic_shapes`, debug logging toggles stack traces
* Don't incorrectly report the tracing context frame when we're compiling; we just don't have this info anymore! (TODO: use the saved frames instead). This is via a new TracingContext.clear_frame context manager
* Add TracingContext.extract_stack() which gives you the tracing context stack.
* Add ShapeEnvLoggingAdapter to report which ShapeEnv any given operation is from (this is helpful for debugging situations when there are too many ShapeEnvs floating around)
* Tweak create_symbol log message to also report Source
* Add a debug log whenever duck sizing occurs
* Report an excerpt of both the user and system backtrace whenever a guard is added in INFO mode. I found this is a good balance of "where did the guard come from" without full backtrace verbosity.
Example log output with the new output:
```
[2023-04-12 08:25:49,003] torch.fx.experimental.symbolic_shapes: [INFO] 0: create_env
[2023-04-12 08:25:49,021] torch.fx.experimental.symbolic_shapes: [INFO] 0: create_symbol s0 = 32 for L['x'].size()[0]
[2023-04-12 08:25:50,154] torch.fx.experimental.symbolic_shapes: [INFO] 0: evaluate_expr s0 < 128 [guard added] at w.py:11 in forward2 (_dynamo/variables/tensor.py:476 in evaluate_expr)
[2023-04-12 08:25:52,057] torch.fx.experimental.symbolic_shapes: [INFO] 0: evaluate_expr Eq(Mod(s0, 16), 0) [guard added] (_inductor/codegen/triton.py:77 in is_aligned)
```
from running
```
import torch
import torch._dynamo
def f(x, y):
return x + y
def forward(x, y):
return forward2(x, y)
def forward2(x, y):
if x.size(0) < 128:
x = x * 2
else:
x = x * 3
r = f(x, y)
r = r * y
return r
def woof():
fn_compiled = torch.compile(forward, dynamic=True)
x = torch.randn(32, device='cuda')
y = torch.randn(32, device='cuda')
print(fn_compiled(x, y))
woof()
```
(To induce the Triton guard, I synthetically reverted https://github.com/pytorch/pytorch/pull/98471)
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98941
Approved by: https://github.com/wconstab
Summary:
Replace _dynamo.config with an object instead of module
Current usage patterns of setting and reading fields on config will work
unchanged.
Only changes needed going forward:
1. import torch._dynamo.config will not work. However, just doing
import torch._dynamo is sufficient to access dynamo config
as torch._dynamo.config.
2. Files inside of _dynamo folder need to access config via
from torch._dynamo.config_util import config instead of
from torch._dynamo import config. Because _dynamo/__init__.py
imports some of the files so it would be circular import.
Test Plan:
Reviewers:
Subscribers:
Tasks:
Tags:
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96455
Approved by: https://github.com/williamwen42
This PR makes basic nnmodule forward hooks work by default, without any overhead. But it leaves silent correctness issues if users modify/remove their hooks later, thus also emits a warning.
- the usual case is to not use hooks, so avoid guard overhead here
- registering any hook before compile will trigger a warning about hook support
- registering a hook later (or removing one) requires user knowledge and opting in,
currently this isn't warnable (but maybe we can observe compiled nnmodules to make it
warnable).
Why skip hook guards by default instead of not tracing __call__/hooks by default?
- avoid having a mode flag that alters dynamo tracing behavior (harder to test both codepaths
in CI with full coverage)
- the most basic hook usecase (registering a hook before compile, and never removing it)
will work by default with this PR, while it would require enablement and incur overhead
in the 'not tracing __call__' proposal.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98371
Approved by: https://github.com/jansel
Symbolic shapes compile time on full CI with inductor is horribly long (even though our aot_eager local runs seemed to suggest that the added latency was only 10s per model.) To patch over the problem for now, run the benchmark suite with dynamic batch only. This should absolve a lot of sins.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97912
Approved by: https://github.com/janeyx99, https://github.com/desertfire
This lets users that are sure they won't use hooks avoid overhead
related to dynamo guards on (assumedly) empty hook dicts on all
nn modules.
Only enable this flag if you are sure you won't change hook-behavior
after compiling. It is ok to register a hook and then compile, if
you promise never to remove/alter the hook. It is also ok to
not register a hook and compile, if you never register a hook later.
Note- this is not the best we can do, and hopefully in the future
we can avoid the need for this option following some of these paths
- make guards fast enough to not be an issue when guarding on hook
dicts
- make a mode where dynamo actually skips tracing __call__ so
hooks are consistently ignored by compiled programs
- use nnmodule versioning so hook changes can be guarded without
explicit hook dict guards
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97830
Approved by: https://github.com/jansel
Summary:
Adds NNC-like logging that is configured through an env var `TORCH_COMPILE_LOGS`
Examples:
`TORCH_LOGS="dynamo,guards" python script.py` - prints dynamo logs at level INFO with guards of all functions that are compiled
`TORCH_LOGS="+dynamo,guards,graph" python script.py` - prints dynamo logs at level DEBUG with guards and graphs (in tabular) format of all graphs that are compiled
[More examples with full output](https://gist.github.com/mlazos/b17f474457308ce15e88c91721ac1cce)
Implementation:
The implementation parses the log settings from the environment, finds any components (aot, dynamo, inductor) or other loggable objects (guards, graph, etc.) and generates a log_state object. This object contains all of the enabled artifacts, and a qualified log name -> level mapping. _init_logs then adds handlers to the highest level logs (the registered logs), and sets any artifact loggers to level DEBUG if the artifact is enabled.
Note: set_logs is an alternative for manipulating the log_state, but if the environment contains TORCH_LOGS, the environment settings will be prioritized.
Adding a new log:
To add a new log, a dev should add their log name to torch._logging._registrations (there are examples there already).
Adding a new artifact:
To add a new artifact, a dev should add their artifact name to torch._logging._registrations as well.
Additionally, wherever the artifact is logged, `torch._logging.getArtifactLogger(__name__, <artifact_name>)` should be used instead of the standard logging implementation.
[design doc](https://docs.google.com/document/d/1ZRfTWKa8eaPq1AxaiHrq4ASTPouzzlPiuquSBEJYwS8/edit#)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94858
Approved by: https://github.com/ezyang
Summary: Makes the debug dir location configurable with TORCH_COMPILE_DEBUG_DIR env var
Test Plan: TORCH_COMPILE_DEBUG_DIR=”.” buck2 run mode/dev-nosan //caffe2/test/inductor:minifier_smoke
Reviewed By: bertmaher
Differential Revision: D43639955
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96089
Approved by: https://github.com/bertmaher
Adds a profiler start and end callback to dynamo's C eval_frame impl, which can be used to profile a region providing a name for visualization. Currently only hooks up one usage to profile cache lookup (primarily covering guards and linear search through linked list).
Example profile taken from toy model:
`python benchmarks/dynamo/distributed.py --toy_model --profile --dynamo aot_eager`
<img width="1342" alt="image" src="https://user-images.githubusercontent.com/4984825/223225931-b2f6c5a7-505a-4c90-9a03-34982f6dc033.png">
Planning to measure overhead in CI, and probably can't afford to check this in enabled by default. Will have to evaluate UX options such as `config.profile_dynamo_cache = True` or some other way.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96119
Approved by: https://github.com/jansel
OK, so this PR used to be about reducing the number of constants we specialize on, but it turns out that unspecialization was ~essentially never used (because we still constant specialized way too aggressively) and I ended up having to fix a bunch of issues to actually get tests to pass. So this PR is now "make int unspecialization actually work". As part of this, I have to turn off unspecialization by default, as there are still latent bugs in inductor.
The general strategy is that an unspecialized int is represented as a SymInt. Representing it as a 0d tensor (which is what the code used to do) is untenable: (1) we often need unspecialized ints to participate in size computations, but we have no way of propagating sympy expressions through tensor compute, and (2) a lot of APIs work when passed SymInt, but not when passed a Tensor. However, I continue to represent Numpy scalars as Tensors, as they are rarely used for size computation and they have an explicit dtype, so they are more accurately modeled as 0d tensors.
* I folded in the changes from https://github.com/pytorch/pytorch/pull/95099 as I cannot represent unspecialized ints as SymInts without also turning on dynamic shapes. This also eliminates the necessity for test_unspec.py, as toggling specialization without dynamic shapes doesn't do anything. As dynamic shapes defaults to unspecializing, I just deleted this entirely; for the specialization case, I rely on regular static shape tests to catch it. (Hypothetically, we could also rerun all the tests with dynamic shapes, but WITH int/float specialization, but this seems... not that useful? I mean, I guess export wants it, but I'd kind of like our Source heuristic to improve enough that export doesn't have to toggle this either.)
* Only 0/1 integers get specialized by default now
* A hodgepodge of fixes. I'll comment on the PR about them.
Fixes https://github.com/pytorch/pytorch/issues/95469
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95621
Approved by: https://github.com/jansel, https://github.com/Chillee
This takes the strategy described in https://docs.google.com/document/d/1lFRYAJo5nrfxRhwIzGnfi2pbLpU6T4ytSRSuLJ5qebI/edit#
It is essentially https://github.com/pytorch/pytorch/pull/95222 but squashed and with changes that are unnecessary given that we assume nonzero returns > 1.
What's in the PR:
* nonzero now supports meta propagation. When `capture_dynamic_output_shape_ops`, it will return a tensor with an unbacked SymInt representing the size in question.
* The unbacked SymInt is UNSOUNDLY assumed to be not equal to 0/1. We will still error if you guard otherwise.
* PrimTorch pointwise operators are updated to use empty_permuted, to avoid guarding on unbacked SymInt from empty_strided (tested in `test_dynamic_pointwise_scalar`)
* Convolution is updated to skip backend selection if batch is unbacked, to avoid guarding on unbacked SymInt (tested in `test_unbacked_batch_resnet`)
* I kept the helper utilities like `definitely_true` for working with possibly unbacked SymInts. They're not used right now but maybe someone will find them useful.
* Added `constrain_unify` to let you specify two unbacked SymInts must have the same value
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95387
Approved by: https://github.com/voznesenskym
Summary:
There are a few races/permission errors in file creation, fixing
OSS:
1. caffe2/torch/_dynamo/utils.py, get_debug_dir: multiple process may conflict on it even it's using us. Adding pid to it
2. caffe2/torch/_dynamo/config.py: may not be a right assumption that we have permission to cwd
Test Plan: sandcastle
Differential Revision: D42905908
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93407
Approved by: https://github.com/soumith, https://github.com/mlazos
Previously, Dynamo faked support for item() when `capture_scalar_outputs` was True by representing it internally as a Tensor. With dynamic shapes, this is no longer necessary; we can represent it directly as a SymInt/SymFloat. Do so. Doing this requires you to use dynamic shapes; in principle we could support scalar outputs WITHOUT dynamic shapes but I won't do this unless someone hollers for it.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Differential Revision: [D42885775](https://our.internmc.facebook.com/intern/diff/D42885775)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93150
Approved by: https://github.com/voznesenskym
@mlazos: skips `item()` calls if compiling with dynamo, by defining a helper function `_get_value` which either returns the result of `.item()` or the scalar cpu tensor if compiling with dynamo. This was done because removing `item()` calls significantly regresses eager perf. Additionally, `_dispatch_sqrt` calls the appropriate sqrt function (math.sqrt, or torch.sqrt).
Fixes https://github.com/pytorch/torchdynamo/issues/1083
This PR will no longer be needed once symint support is default.
This PR closes all remaining graph breaks in the optimizers (!!)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88173
Approved by: https://github.com/albanD
- Adds `log_level` to aot's config
- Outputs log to `<graph_name>_<log_level>.log` in aot_torchinductor subfolder of the debug directory
- Modifies the Inductor debug context to use the graph name when naming the folder instead of the os pid
- Adds `TORCH_COMPILE_DEBUG` flag to enable it, (as well as separate dynamo and inductor logs)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88987
Approved by: https://github.com/Chillee
The old code didn't actually fakeify traceable tensor subclasses at the
time they are added as a GraphArg to the module; now we do, by ignoring
the subclass during fakeification and relying on Dynamo to simulate
the subclass on top. See comments for more details.
BTW, this codepath is super broken, see filed issues linked on the
inside.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90009
Approved by: https://github.com/wconstab, https://github.com/voznesenskym
Performance benchmarks on 6 popular models from 1-64 GPUs compiled with
torchinductor show performance gains or parity with eager, and showed
regressions without DDPOptimizer. *Note: resnet50 with small batch size shows a regression with optimizer, in part due to failing to compile one subgraph due to input mutation, which will be fixed.
(hf_Bert, hf_T5_large, hf_T5, hf_GPT2_large, timm_vision_transformer, resnet50)
Correctness checks are implemented in CI (test_dynamo_distributed.py),
via single-gpu benchmark scripts iterating over many models
(benchmarks/dynamo/torchbench.py/timm_models.py/huggingface.py),
and via (multi-gpu benchmark scripts in torchbench)[https://github.com/pytorch/benchmark/tree/main/userbenchmark/ddp_experiments].
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88523
Approved by: https://github.com/davidberard98
I noticed that a lot of bugs are being suppressed by torchdynamo's default
error suppression, and worse yet, there's no way to unsuppress them. After
discussion with voz and soumith, we decided that we will unify error suppression
into a single option (suppress_errors) and default suppression to False.
If your model used to work and no longer works, try TORCHDYNAMO_SUPPRESS_ERRORS=1
to bring back the old suppression behavior.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
cc @jansel @lezcano @fdrocha @mlazos @soumith @voznesenskym @yanboliang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87440
Approved by: https://github.com/voznesenskym, https://github.com/albanD
