This only handles AttributeError, but in general, any exception coming from
here is a user exception. let me know if we prefer to catch all exceptions, and then reraise them as observed exceptions.
```
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/_dynamo/symbolic_convert.py", line 2200, in CALL_FUNCTION
self.call_function(fn, args, {})
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/_dynamo/symbolic_convert.py", line 1210, in call_function
self.push(fn.call_function(self, args, kwargs)) # type: ignore[arg-type]
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/_dynamo/variables/lazy.py", line 201, in realize_and_forward
return getattr(self.realize(), name)(*args, **kwargs)
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/_dynamo/variables/nn_module.py", line 472, in call_function
initialize_lazy_module(tx, mod, args, kwargs)
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/_dynamo/variables/nn_module.py", line 104, in initialize_lazy_module
mod._infer_parameters(mod, fake_args, fake_kwargs)
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/nn/modules/lazy.py", line 261, in _infer_parameters
module.initialize_parameters(*args, **kwargs)
...,
File "/packages/aps.ads.gmp/launcher_with_publish#link-tree/torch/nn/modules/module.py", line 1962, in __getattr__
raise AttributeError(
torch._dynamo.exc.InternalTorchDynamoError: AttributeError: '...' object has no attribute '...'
```
Note that we crash with a sligthly different exception trace in the other test I added. Let me know if we want this to not throw directly to the end user.
```
======================================================================
ERROR: test_lazy_module_bad_params (__main__.NNModuleTests.test_lazy_module_bad_params)
----------------------------------------------------------------------
Traceback (most recent call last):
File "/data/users/clr/pytorch/torch/testing/_internal/common_utils.py", line 3223, in wrapper
method(*args, **kwargs)
~~~~~~^^^^^^^^^^^^^^^^^
File "/data/users/clr/pytorch/test/dynamo/test_modules.py", line 1683, in test_lazy_module_bad_params
exp_res = opt_m(x, y)
File "/data/users/clr/pytorch/torch/_dynamo/eval_frame.py", line 411, in __call__
return super().__call__(*args, **kwargs)
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/data/users/clr/pytorch/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/data/users/clr/pytorch/torch/nn/modules/module.py", line 1784, in _call_impl
return forward_call(*args, **kwargs)
File "/data/users/clr/pytorch/torch/_dynamo/eval_frame.py", line 473, in _call_lazy_check
self._orig_mod._infer_parameters(self._orig_mod, args, kwargs)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/users/clr/pytorch/torch/nn/modules/lazy.py", line 261, in _infer_parameters
module.initialize_parameters(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/data/users/clr/pytorch/test/dynamo/test_modules.py", line 711, in initialize_parameters
self.foo += 1
^^^^^^^^
File "/data/users/clr/pytorch/torch/nn/modules/module.py", line 1962, in __getattr__
raise AttributeError(
f"'{type(self).__name__}' object has no attribute '{name}'"
)
AttributeError: 'LazyModuleBadInferParams' object has no attribute 'foo'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158501
Approved by: https://github.com/williamwen42, https://github.com/jansel
https://github.com/pytorch/pytorch/pull/152708 expanded support of `get_estimated_runtime` to many more types of `SchedulerNodes`. This caused an increase in compile time because we're always calling `get_estimated_runtime` to populate the metrics table. This PR adds a flag for this logging, which reduces the instruction count by 8%. Long term, we should probably merge metrics.py with TORCH_LOGS/tlparse (suggestion from @xmfan).
Update: added support for TORCH_LOGS for the metrics logging.
Test Plan:
mm_loop.py and many existing tests cover.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153506
Approved by: https://github.com/eellison
User defined NN module might have their own `__len__` or `__bool__`
methods which Dynamo needs to trace through, so that side effects and/or
reads to buffered writes are properly handled.
This patch removes the special `UnspecializedNNModuleVariable` branch in
Dynamo's branch handling, and lets these cases fall into the
`UserDefinedObjectVariable` branch, which handles the aforementioned
cases correctly.
Fixes#145284.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145786
Approved by: https://github.com/williamwen42
In hinsight, we never needed a DICT_SUBCLASS_GUARD_MANAGER, because Dynamo would inline through the overridden keys method. In this PR, we ensure that while creating guards and constructing variable trackers, we get the `d.keys()` value by using `dict.keys(d)`. This ensures that we do not call overridden keys method. Therefore, the C++ guard can use `PyDict_Next` directly to check the guards.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143722
Approved by: https://github.com/jansel
Changes:
1. Bump `ruff` from 0.7.4 to 0.8.4
2. Change `%`-formatted strings to f-string
3. Change arguments with the `__`-prefix to positional-only arguments with the `/` separator in function signature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143753
Approved by: https://github.com/Skylion007
In hinsight, we never needed a DICT_SUBCLASS_GUARD_MANAGER, because Dynamo would inline through the overridden keys method. In this PR, we ensure that while creating guards and constructing variable trackers, we get the `d.keys()` value by using `dict.keys(d)`. This ensures that we do not call overridden keys method. Therefore, the C++ guard can use `PyDict_Next` directly to check the guards.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143722
Approved by: https://github.com/jansel
Resolves issue #140464 by adding an option to not specialize int from nn.Modules (False by default to maintain existing behavior).
Test Plan: `buck2 test mode/opt caffe2/test/dynamo:test_dynamo -- test_modules.py::NNModuleTests::test_nn_module_unspec_int_attr`
Differential Revision: D66837042
Pull Request resolved: https://github.com/pytorch/pytorch/pull/142829
Approved by: https://github.com/ezyang, https://github.com/yanboliang
This is a bug on the main exposed by https://github.com/pytorch/pytorch/issues/139476
We have dict tag optimization where if the dict tag does not change, we
skip guards on all the items of the dict that are "immutable". We
considered tensors as immutable in such scenarios. This is critical for
guard eval performance, because generally users dont change their
parameters.
If I try to remove this optimization, we see slowdowns, e.g, 3.03x to
2.95x on conv_mixer TIMM benchamrk.
So, I am adding a flag which keeps the current state but allows the
users to remove this optimization. Not ideal, but given how serious guard eval perf has to be,
we are in the gray are of unsoundness vs performance tradeoff.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139560
Approved by: https://github.com/jansel
This is a bug on the main exposed by https://github.com/pytorch/pytorch/issues/139476
We have dict tag optimization where if the dict tag does not change, we
skip guards on all the items of the dict that are "immutable". We
considered tensors as immutable in such scenarios. This is critical for
guard eval performance, because generally users dont change their
parameters.
If I try to remove this optimization, we see slowdowns, e.g, 3.03x to
2.95x on conv_mixer TIMM benchamrk.
So, I am adding a flag which keeps the current state but allows the
users to remove this optimization. Not ideal, but given how serious guard eval perf has to be,
we are in the gray are of unsoundness vs performance tradeoff.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139560
Approved by: https://github.com/jansel
This patch
- removes the `is_lazy_module` check from `is_dynamic_nn_module`, and
adds a regression test.
- removes a series of dynamo expected failures on lazy modules. The few
ones I checked all were failing due to speculation log divergence,
similar to #138489.
Note that #100047 introduced the conditional removed in this patch, and
it was trying to fix#100001. But I've confirmed locally that #100001 no
longer repros after this patch.
Fixes#138489. See more context in the issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138639
Approved by: https://github.com/jansel
