Fixes#142397
Basic implementation is done. What's left:
- [x] Different dtype/device tensors in the TensorList
- [x] fast path for grouping the foreach kernel
- [x] Tests
Regarding tests, I found some tests in `test/test_torch.py` for GradScaler but I couldn't figure out what is the best way to enable the test for MPS device.
By removing `@onlyNativeDeviceTypes`, one enables the tests for MPS but also enables tests for all other devices which are not included in the native device types. If I put:
`instantiate_device_type_tests(TestTorchDeviceType, globals(), allow_mps=True)`
This enables lots of tests in that class for MPS which were not(?) being tested before? This part needs some clarification
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150255
Approved by: https://github.com/malfet
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Third PR in a series of PRs to broaden differentiable optimizer support w/ @janeyx99 (sorry for pinging over the holidays! I just wanted to put this one out but I am definitely not asking for review or anything like that rn)
This is also going to probably be my last PR before the holidays!
Note: This is a branch of #143710 -- I've never worked on a branch of a branch before so I wasn't sure about the protocol so I thought I'd just made the PR and wait until that one gets merged.
This is adding support for differentiable lr, weight_decay, and betas to Adam and AdamW (but after refactoring AdamW into an Adam subclass, it's really just changing code in torch/optim/adam.py)
I had one main thing I was wondering about, which is that adam already has a differentiable flag built in, so I have code like this
```py
if differentiable and isinstance(beta2, Tensor):
if beta2.requires_grad:
exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj().mul(1 - beta2))
else:
exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
else:
exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
```
That I could definitely simplify to just
```py
if differentiable and isinstance(beta2, Tensor):
exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj().mul(1 - beta2))
else:
exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
```
It would definitely be a little slower in the case that it's differentiable but doesn't need a grad for beta2, but the code would also be a lot more clear and I'm debating speed vs future code usability.
Also the line in the above example:
```py
exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj().mul(1 - beta2))
```
was concerning to me because it is considerably more expensive than `value=1 - beta2`, but I couldn't think of a better way to do it.
Further work on #141832
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143726
Approved by: https://github.com/janeyx99
Fixes#104899
Refactors AdamW into Adam by making AdamW a subclass of Adam. Additionally adds a test to assert that the added parameter `decoupled_weight_decay` is True in AdamW and also updates test_defaults_changed_to_foreach to account for the differences in module location for AdamW.
Heavily heavily inspired by #118857 by @tfsingh
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143710
Approved by: https://github.com/janeyx99
A proposal addressing Issue #1489: **Optimizer should track parameter names and not id.**
(also mentioned in here: [[RFC] Introducing FQNs/clarity eyeglasses to optim state_dict](https://dev-discuss.pytorch.org/t/rfc-introducing-fqns-clarity-to-optim-state-dict/1552)
## Summary
This PR introduces a backward-compatible enhancement where optimizers track parameter names instead of just their id.
Optimizers can be initialized with `named_parameters()` as:
```python
optimizer = optim.SGD(model.named_parameters(), lr=0.01, momentum=0.9)
```
This allows for greater clarity and ease when handling optimizers, as the parameters' names are preserved within the optimizer’s `state_dict` as:
```
state_dict =
{
'state': {
0: {'momentum_buffer': tensor(...), ...},
1: {'momentum_buffer': tensor(...), ...},
},
'param_groups': [
{
'lr': 0.01,
'weight_decay': 0,
...
'params': [0,1]
'param_names' ['layer.weight', 'layer.bias'] (optional)
}
]
}
```
Loading `state_dict` is not changed (backward-compatible) and the `param_names` key will be ignored.
## Key Features
#### Named Parameters in Optimizer Initialization:
Optimizers can accept the output of `model.named_parameters()` during initialization, allowing them to store parameter names directly.
#### Parameter Names in `state_dict`:
The parameter names are saved as a list in the optimizer’s `state_dict` with key `param_names`, alongside the `params` indices, ensuring seamless tracking of both names and parameters.
## Backward Compatibility
#### No Breaking Changes:
This change is fully backward-compatible. The added `param_names` key in the optimizer's `state_dict` is ignored when loading a state to the optimizer.
#### Customization with Hooks:
For more control, the loaded state_dict can be modified using a custom `register_load_state_dict_pre_hook`, providing flexibility for different design needs.
## Documentation Updates
Please refer to the documentation changes for more details on how this feature is implemented and how it can be used effectively.
## Solution Example:
A suggested solution to the problem mentioned in #1489, for the same parameters but in a different order.
The following `register_load_state_dict_pre_hook` should be added to the optimizer before loading to enable loading the state dict :
```python
def adapt_state_dict_ids(optimizer, state_dict):
# assuming a single param group.
current_state_group = optimizer.state_dict()['param_groups'][0]
loaded_state_group = state_dict['param_groups'][0]
# same number of params, same names, only different ordering
current_state_name_to_id_mapping = {} # mapping -- param_name: id
for i, name in enumerate(current_state_group['param_names']):
current_state_name_to_id_mapping[name] = current_state_group['params'][i]
# changing the ids of the loaded state dict to match the order of the given state dict.
for i, name in enumerate(current_state_group['param_names']):
loaded_state_group['params'][i] = current_state_name_to_id_mapping[name]
return state_dict
```
In this code, the loaded `state_dict` ids are adapted to match the order of the current optimizer `state_dict`.
Both the previous and the current optimizers are required to be initiated with `named_parameters()` to have the 'param_names' key in the dict.
### Note
This is my first contribution to PyTorch, and I wish to receive feedback or suggestions for improvement.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134107
Approved by: https://github.com/janeyx99
Co-authored-by: Jane (Yuan) Xu <31798555+janeyx99@users.noreply.github.com>
Use `typing_extensions.deprecated` for deprecation annotation if possible. Otherwise, add `category=FutureWarning` to `warnings.warn("message")` if the category is missing.
Note that only warnings that their messages contain `[Dd]eprecat(ed|ion)` are updated in this PR.
UPDATE: Use `FutureWarning` instead of `DeprecationWarning`.
Resolves#126888
- #126888
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126898
Approved by: https://github.com/albanD
This resolves a bug in eager where if an old state dict is loaded (without the capturable flag) but the original dict had the capturable flag, then state_steps would be on cuda but we would take the non-capturable path. We now fallback to eager if capturable=False.
Current design doc and discussion: https://docs.google.com/document/d/1DmmbiaSp16CDZtGw1qzXKHFTY_0gqc0xpnBdviXq0vk/edit#heading=h.871u7bvwz7ze
Note on the actual fallback logic - there was an issue with torchscript originally not handling *args, **kwargs properly, after rectifying that by using `functools.wraps`, there was an additional bug with scoping which required the single tensor implementation to be in the global scope at the time of the fallback closure being created. I pass in the single tensor function to the `_disable_dynamo_if_unsupported` decorator to workaround this bug.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123619
Approved by: https://github.com/janeyx99
Support fused_sgd_kernel support for CPU.
## Bench result:
32 core/sockets ICX
Test Scripts:
https://gist.github.com/zhuhaozhe/688763e17e93e4c5e12f25f676ec90d9https://gist.github.com/zhuhaozhe/ad9938694bc7fae8b66d376f4dffc6c9
```
Tensor Size: 262144, Num Tensor 4, Num Threads: 1
_single_tensor_sgd time: 0.2301 seconds
_fused_sgd time: 0.0925 seconds
Tensor Size: 4194304, Num Tensor 32, Num Threads: 32
_single_tensor_sgd time: 2.6195 seconds
_fused_sgd time: 1.7543 seconds
```
## Test Plan:
```
python test_optim.py -k test_fused_matches_forloop
python test_optim.py -k test_fused_large_tensor
python test_optim.py -k test_can_load_older_state_dict
python test_optim.py -k test_grad_scaling_autocast_fused_optimizers
python test_torch.py -k test_grad_scaling_autocast_fused
python test_torch.py -k test_params_invalidated_with_grads_invalidated_between_unscale_and_step
```
Looks like we already have some PRs under this issue https://github.com/pytorch/pytorch/issues/123451 to unified the UTs, I did not modified UT in this PR.
Co-authored-by: Jane Xu <janeyx@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123629
Approved by: https://github.com/jgong5, https://github.com/janeyx99
