When working with a highly-dynamic python code it's not always possible to express the static types. However if we consider the end-user experience for somebody who uses both pytorch and a static type checker (mypy, pyright), we should error on the side of being ergonomic and not technically correct.
The `nn.Module.__getattr__` is one of the such examples: on paper the return type is correct. In practice the community would benefit from having `Any` as a return type because it would avoid littering the idiomatic pytorch code with `cast`, `# type: ignore`, `assert`, `isinstance`, etc.
Some evidences:
- linked in the comment thread on pyright bug tracker https://github.com/microsoft/pyright/issues/4213
- `pyre` type checker steps outside of the normal type checking practices and special-cases `registrer_buffer()` in part to avoid this problem. https://pyre-check.org/docs/features/ This is not a very scalable solution since type-checkers generally aim at adhering to the spec (various typing PEPs).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104321
Approved by: https://github.com/kit1980, https://github.com/albanD
Fixes#64601 and #98906
Adds an `assign` argument to `load_state_dict` that loads params/buffers by assignment instead of doing `param.copy_(param_from_state_dict)`.
Primarily intended to remove the need for the `.to_empty()` in
```
with torch.device('meta'):
m = SomeModule()
m.to_empty()
state_dict = torch.load('...pth')
m.load_state_dict(state_dict)
```
so we can instead do
```
with torch.device('meta'):
m = SomeModule()
state_dict = torch.load('...pth')
m.load_state_dict(state_dict, assign=True)
```
**A problem with this PR for the case where the model is initialized on meta is what happens to nonpersistent buffers/params corresponding to keys missing from the state dict?**
What happens in the case where `load_state_dict(state_dict, strict=False, assign=True)` and the state_dict is missing some keys? The corresponding params missing from the `state_dict` and nonpersistent buffers would still be on `meta` and need to be manually initialized. However, I don't think we offer an API that would initialize these.
One solution would be to make these empty tensors but it might not be semantically correct...
Pull Request resolved: https://github.com/pytorch/pytorch/pull/102212
Approved by: https://github.com/albanD
Opening this so I can discuss with @albanD
I built a proof of concept of an in place API for an nn.Module that allows us to save and load a torch.compiled model with no issues https://github.com/msaroufim/mlsys-experiments/blob/main/save-compiled-model.py
So users can run` model.compile()` and then run `torch.save(model, "model.pt")` and `torch.load(model, "model.pt)` with no issues unlike the rather strange current suggestion we give to users which is `opt_mod = torch.compile(mod); torch.save(mod, "model.pt")`
Right now I'm trying to extend this to work for nn.modules more generally
TODO: Failing tests
* [x] torch.jit.load -> issue was because of aliasing `__call__` to `_call_impl`, _call_impl used to be skipped when now it lo longer is so expanded the skip check. I added an explicit `torch.jit.load()` test now which @davidberard98 suggested
* [x] functorch seems to be a flake - ran locally and it worked `pytest functorch/test_eager_transforms.py`
* [x] a test infra flake - `test_testing.py::TestImports::test_no_mutate_global_logging_on_import_path_functorch`
* [x] It seems like I broke inlining in dynamo though `python -m pytest test/dynamo/test_dynamic_shapes.py -k test_issue175` chatting with Voz about it but still not entirely sure how to fix - found a workaround after chatting with @yanboliang
* [x] `pytest test/dynamo/test_modules.py` and `test/dynamo/test_dynamic_shapes` `test/dynamo/test_misc.py` seem to be failing in CI but trying it out locally they all pass tests passed with 0 failures
* [x] `pytest test/profiler/test_profiler_tree.py ` these tests have ProfilerTrees explicitly printed and will now break if __call__ is not in tree - ran with `EXPECT_ACCEPT=1`
* [x] `pytest test/test_torch.py::TestTorch::test_typed_storage_deprecation_warning` a flake, ran this locally and it works fine
* [x] I reverted my changes to `_dynamo/nn_module.py` since it looks like @wconstab is now directly handling `_call_impl` there but this is triggering an infinite inlining which is crashing
* [x] Tried out to instead override `__call__`, python doesnt like this though https://github.com/pytorch/pytorch/pull/97565#issuecomment-1524570439
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97565
Approved by: https://github.com/aaronenyeshi, https://github.com/albanD, https://github.com/voznesenskym
Opening this so I can discuss with @albanD
I built a proof of concept of an in place API for an nn.Module that allows us to save and load a torch.compiled model with no issues https://github.com/msaroufim/mlsys-experiments/blob/main/save-compiled-model.py
So users can run` model.compile()` and then run `torch.save(model, "model.pt")` and `torch.load(model, "model.pt)` with no issues unlike the rather strange current suggestion we give to users which is `opt_mod = torch.compile(mod); torch.save(mod, "model.pt")`
Right now I'm trying to extend this to work for nn.modules more generally
TODO: Failing tests
* [x] torch.jit.load -> issue was because of aliasing `__call__` to `_call_impl`, _call_impl used to be skipped when now it lo longer is so expanded the skip check. I added an explicit `torch.jit.load()` test now which @davidberard98 suggested
* [x] functorch seems to be a flake - ran locally and it worked `pytest functorch/test_eager_transforms.py`
* [x] a test infra flake - `test_testing.py::TestImports::test_no_mutate_global_logging_on_import_path_functorch`
* [x] It seems like I broke inlining in dynamo though `python -m pytest test/dynamo/test_dynamic_shapes.py -k test_issue175` chatting with Voz about it but still not entirely sure how to fix - found a workaround after chatting with @yanboliang
* [x] `pytest test/dynamo/test_modules.py` and `test/dynamo/test_dynamic_shapes` `test/dynamo/test_misc.py` seem to be failing in CI but trying it out locally they all pass tests passed with 0 failures
* [x] `pytest test/profiler/test_profiler_tree.py ` these tests have ProfilerTrees explicitly printed and will now break if __call__ is not in tree - ran with `EXPECT_ACCEPT=1`
* [x] `pytest test/test_torch.py::TestTorch::test_typed_storage_deprecation_warning` a flake, ran this locally and it works fine
* [x] I reverted my changes to `_dynamo/nn_module.py` since it looks like @wconstab is now directly handling `_call_impl` there but this is triggering an infinite inlining which is crashing
* [x] Tried out to instead override `__call__`, python doesnt like this though https://github.com/pytorch/pytorch/pull/97565#issuecomment-1524570439
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97565
Approved by: https://github.com/aaronenyeshi, https://github.com/albanD
Some modules like lazyModule may override '_save_to_state_dict()', in this case, pre_state_dict hook will not be called. So move the pre_state_dict hook out from '_save_to_state_dict()' to make sure the pre hook could be called
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98964
Approved by: https://github.com/albanD
Reverting due to concerns over silent unsoundness (skipped hooks) if users have directly added hooks dicts without using official torch APIs.
This reverts commit 26045336ca323fd27cff2a7340fe896117d5fb6e.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96242
Approved by: https://github.com/albanD
This PR optimizes the guards overhead introduced by dynamo tracing module forward hooks.
It can and maybe should be followed by a wider change proposed by @voznesenskym to optimize specialized nnmodules by 'observing' any user mutations and directly invalidating the root guard, obviating the need to install other nnmodule guards. (But this observer change seems more involved...)
Idea: maintain a flag, and keep it up to date whenever adding or
removing hooks. Use the flag rather than dict checks to enter the call fast path.
- need to extend RemovableHandle to keep a ref to nnModule so it can update the flag on removal.
- also need to handle the flag in ScriptModule which still uses the python call impl when called from python.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95931
Approved by: https://github.com/ezyang, https://github.com/voznesenskym
Applies some more harmless pyupgrades. This one gets rid of deprecated aliases in unit_tests and more upgrades yield for loops into yield from generators which are more performance and propagates more information / exceptions from original generator. This is the modern recommended way of forwarding generators.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94309
Approved by: https://github.com/albanD
Attempts to fix#92656
BC-breaking! This changes the default of zero_grad in optim and in nn to default set grads to None instead of zero tensors. We are changing the default because there are proven perf wins and existing code has typically not regressed due to this change. (will probably have to flesh out this note more).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92731
Approved by: https://github.com/ngimel
Fixes#52664. Checks if the attribute is a property that defines a setter and uses fset in __setattr__ rather than registering an inaccessible module / parameter.
This is BC-breaking as the attribute setters on nn.Module properties used to be ignored and now will be called properly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92044
Approved by: https://github.com/albanD
Fixes#91654.
Currently, the `hook` parameters of `nn.Module.register_forward_pre_hook` and `nn.Module.register_forward_hook` are typed as `Callable[..., None]`, which 1) does not enable the validation of the signature of `hook` and 2) incorrectly restricts the return type of `hook`, which the docstrings of these methods themselves state can be non-`None`.
The typing of the first parameter of `hook` as `TypeVar("T", bound="Module")` allows the binding of `Callable` whose first parameter is a subclass of `Module`.
---
Here are some examples of:
1. forward hooks and pre-hook hooks being accepted by mypy according to the new type hints
2. mypy throwing errors d.t. incorrect `hook` signatures
3. false negatives of pre-hooks being accepted as forward hooks
4. false negatives of hooks with kwargs being accepted irrespective of the value provided for `with_kwargs`
```python
from typing import Any, Dict, Tuple
import torch
from torch import nn
def forward_pre_hook(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
) -> None:
...
def forward_pre_hook_return_input(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
) -> Tuple[torch.Tensor, ...]:
...
def forward_pre_hook_with_kwargs(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
kwargs: Dict[str, Any],
) -> None:
...
def forward_pre_hook_with_kwargs_return_input(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
kwargs: Dict[str, Any],
) -> Tuple[Tuple[torch.Tensor, ...], Dict[str, Any]]:
...
def forward_hook(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
output: torch.Tensor,
) -> None:
...
def forward_hook_return_output(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
output: torch.Tensor,
) -> torch.Tensor:
...
def forward_hook_with_kwargs(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
kwargs: Dict[str, Any],
output: torch.Tensor,
) -> None:
...
def forward_hook_with_kwargs_return_output(
module: nn.Linear,
args: Tuple[torch.Tensor, ...],
kwargs: Dict[str, Any],
output: torch.Tensor,
) -> torch.Tensor:
...
model = nn.Module()
# OK
model.register_forward_pre_hook(forward_pre_hook)
model.register_forward_pre_hook(forward_pre_hook_return_input)
model.register_forward_pre_hook(forward_pre_hook_with_kwargs, with_kwargs=True)
model.register_forward_pre_hook(forward_pre_hook_with_kwargs_return_input, with_kwargs=True)
model.register_forward_hook(forward_hook)
model.register_forward_hook(forward_hook_return_output)
model.register_forward_hook(forward_hook_with_kwargs, with_kwargs=True)
model.register_forward_hook(forward_hook_with_kwargs_return_output, with_kwargs=True)
# mypy(error): [arg-type]
model.register_forward_pre_hook(forward_hook)
model.register_forward_pre_hook(forward_hook_return_output)
model.register_forward_pre_hook(forward_hook_with_kwargs)
model.register_forward_pre_hook(forward_hook_with_kwargs_return_output)
model.register_forward_hook(forward_pre_hook)
model.register_forward_hook(forward_pre_hook_return_input)
# false negatives
model.register_forward_hook(forward_pre_hook_with_kwargs)
model.register_forward_hook(forward_pre_hook_with_kwargs_return_input)
model.register_forward_pre_hook(forward_pre_hook_with_kwargs, with_kwargs=False)
model.register_forward_pre_hook(forward_pre_hook_with_kwargs_return_input, with_kwargs=False)
...
```
---
Though it is not functional as of mypy 0.991, the ideal typing of these methods would use [`typing.Literal`](https://mypy.readthedocs.io/en/stable/literal_types.html#literal-types):
```python
T = TypeVar("T", bound="Module")
class Module:
@overload
def register_forward_hook(
self,
hook: Callable[[T, Tuple[Any, ...], Any], Optional[Any]],
*,
prepend: bool = ...,
with_kwargs: Literal[False] = ...,
) -> RemovableHandle:
...
@overload
def register_forward_hook(
self,
hook: Callable[[T, Tuple[Any, ...], Dict[str, Any], Any], Optional[Any]],
*,
prepend: bool = ...,
with_kwargs: Literal[True] = ...,
) -> RemovableHandle:
...
def register_forward_hook(...):
...
```
which would:
1. validate the signature of `hook` according to the corresponding literal value provided for `with_kwargs` (and fix the false negative examples above)
2. implicitly define the [fallback `bool` signature](https://github.com/python/mypy/issues/6113#issuecomment-1266186192) e.g. to handle if a non-literal is provided for `with_kwargs`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92061
Approved by: https://github.com/albanD
closes#35643
This PR is mostly borrowed from #82042. Thanks @Padarn for implementing
the first version and debugging into the errors.
Based on the discussion in #82042 this PR adds a with_kwargs
argument to register_forward_pre_hook and register_forward_hook
methods. When the arg is set to true, the provided hook must accept
kwargs args. Under the hook, this PR adds a
`_forward_pre_hooks_with_kwargs` and a `_forward_hook_with_kwargs`
set to keep track of which hooks accept kwargs.
Differential Revision: [D41431111](https://our.internmc.facebook.com/intern/diff/D41431111)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89389
Approved by: https://github.com/soulitzer
There is a bug in the implementation of the registration hooks introduced in https://github.com/pytorch/pytorch/pull/86148 whereby if the hook returns a tensor, then the short circuiting logic:
```
value = hook(self, name, value) or value
```
Raises an exception
```
RuntimeError: Boolean value of Tensor with more than one value is ambiguous
```
Fixing the logic so that it only checks to see if the value is `None` before overriding
Fixes#85837
CC: @albanD @jbschlosser
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87369
Approved by: https://github.com/albanD
As described in the issue, this PR adds hooks to be run when `register_parameter`, `register_buffer` and `register_module` are called.
Fixes#85837
cc @albanD @mruberry @jbschlosser @walterddr @kshitij12345 @saketh-are
Pull Request resolved: https://github.com/pytorch/pytorch/pull/86148
Approved by: https://github.com/albanD
Fixes#61398
The original implementation is very slow when the state_dict.keys() is long. This PR only passes relevant keys to the child module.
existing test passes: `pytest test/test_nn.py -k state_dict`
I couldn't figure out a good way to write a new test for this new behavior. Had a new snippet, but it will be flaky if integrated into the main CI because it's a timing based check.
But I can verify that the test took 30s to run, after this PR it only takes 0.5s.
```python
def test_load_state_dict_large(self):
# construct a module with 4 levels of module, 10 linear each, leads to 10k items in the dictionary
import copy
import time
base_module = nn.Linear(1,1)
model = base_module
for level in range(4):
model = nn.Sequential(*[copy.deepcopy(model) for _ in range(10)])
state_dict = model.state_dict()
self.assertEqual(len(state_dict), 20000)
st = time.time()
model.load_state_dict(state_dict, strict=True)
strict_load_time = time.time() - st
# it took 0.5 seconds to
self.assertLess(strict_load_time, 10)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85743
Approved by: https://github.com/albanD
This is a new version of #15648 based on the latest master branch.
Unlike the previous PR where I fixed a lot of the doctests in addition to integrating xdoctest, I'm going to reduce the scope here. I'm simply going to integrate xdoctest, and then I'm going to mark all of the failing tests as "SKIP". This will let xdoctest run on the dashboards, provide some value, and still let the dashboards pass. I'll leave fixing the doctests themselves to another PR.
In my initial commit, I do the bare minimum to get something running with failing dashboards. The few tests that I marked as skip are causing segfaults. Running xdoctest results in 293 failed, 201 passed tests. The next commits will be to disable those tests. (unfortunately I don't have a tool that will insert the `#xdoctest: +SKIP` directive over every failing test, so I'm going to do this mostly manually.)
Fixes https://github.com/pytorch/pytorch/issues/71105
@ezyang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82797
Approved by: https://github.com/ezyang
Module's overridden `__setattr__` has special handling for parameters, submodules, and buffers, resulting in reduces performance for its default attributes (which are not of those types). Setting them directly results in a significant improvement for module construction speed (~10x).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77098
Approved by: https://github.com/jbschlosser
