-
Notifications
You must be signed in to change notification settings - Fork 21.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[FSDP][8/N] Replace _FSDPPolicy.policy
with _Policy._run_policy
#104969
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/104969
Note: Links to docs will display an error until the docs builds have been completed. ✅ 1 Unrelated FailureAs of commit ad2e8ce: UNSTABLE - The following job failed but was likely due to flakiness present on trunk and has been marked as unstable:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
ghstack-source-id: c41660b85e42e6de90e78d12b859a9961ddd703e Pull Request resolved: #104969
[ghstack-poisoned]
ghstack-source-id: 53a5622bc1a3603308b3d266e5840a730b6247ae Pull Request resolved: #104969
[ghstack-poisoned]
ghstack-source-id: 783d09ccc43c84621f91e801159db877b508b541 Pull Request resolved: #104969
[ghstack-poisoned]
_FSDPPolicy.policy
with run_policy
_FSDPPolicy.policy
with _run_policy
ghstack-source-id: a5fcf40d6e6d5f7db475bf0b60d3cfa3a42ee3d3 Pull Request resolved: pytorch#104969
This is a code organization improvement. It formalizes the contract that an FSDP policy should return something like `target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]` that maps each module to wrap to its FSDP kwargs. It does so by requiring a `_run_policy` abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call `_run_policy()` to generate the dict and do any validation or post-processing. This PR is technically BC-breaking because it removes the public `ModuleWrapPolicy.policy`. However, I do not think anyone was using that anyway, so this is a pretty safe breakage. [ghstack-poisoned]
ghstack-source-id: 86c9da70764481465d83b24d69224561d1f83fc3 Pull Request resolved: pytorch#104969
This is a code organization improvement. - It renames `_FSDPPolicy` to `_WrapPolicy` to show that it is not only for FSDP but for any module-level API. - It formalizes the contract that such a policy should return something like `target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]` that maps each module to wrap to its kwargs. It does so by requiring a `_run_policy` abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call `_run_policy()` to generate the dict and do any validation or post-processing. This PR is technically BC-breaking because it removes the public `ModuleWrapPolicy.policy`. However, I do not think anyone was using that anyway, so this is a pretty safe breakage. [ghstack-poisoned]
ghstack-source-id: 8e8b698c212c8c5e8269641ec5c19c0498488878 Pull Request resolved: pytorch#104969
ghstack-source-id: 8e8b698c212c8c5e8269641ec5c19c0498488878 Pull Request resolved: pytorch#104969
This does some code organization improvement. - It renames `_FSDPPolicy` to `_WrapPolicy` to show that it is not only for FSDP but for any module-level API. - It formalizes the contract that such a policy should return something like `target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]` that maps each module to wrap to its kwargs. It does so by requiring a `_run_policy` abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call `_run_policy()` to generate the dict and do any validation or post-processing. This PR is technically BC-breaking because it removes the public `ModuleWrapPolicy.policy`. However, I do not think anyone was using that anyway, so this is a pretty safe breakage. [ghstack-poisoned]
continue | ||
elif isinstance(module, module_classes): | ||
# Shallow copy to avoid coupling changes across modules | ||
target_module_to_kwargs[module] = copy.copy(root_kwargs) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
why do we need to copy the args? Do we expect root_kwargs to change?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We must copy the kwargs or else if a function overrides it (e.g. the mixed precision overrides), then it will override it for every target module. We have to avoid them sharing a reference to the same kwarg dict.
ghstack-source-id: 3c5d71e48d07274d6ed3e3f791d300427849d2da Pull Request resolved: pytorch#104969
ghstack-source-id: 3c5d71e48d07274d6ed3e3f791d300427849d2da Pull Request resolved: pytorch#104969
ghstack-source-id: 3c5d71e48d07274d6ed3e3f791d300427849d2da Pull Request resolved: pytorch#104969
ghstack-source-id: 3c5d71e48d07274d6ed3e3f791d300427849d2da Pull Request resolved: pytorch#104969
_FSDPPolicy.policy
with _run_policy
_FSDPPolicy.policy
with _Policy._run_policy
…n_policy`" This does some code organization improvement. - It renames `_FSDPPolicy` to `_Policy` to show that it is not only for FSDP but for any module-level API. - It formalizes the contract that such a policy should return something like `target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]` that maps each module to wrap to its kwargs. It does so by requiring a `_run_policy` abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call `_run_policy()` to generate the dict and do any validation or post-processing. This PR is technically BC-breaking because it removes the public `ModuleWrapPolicy.policy`. However, I do not think anyone was using that anyway, so this is a pretty safe breakage. [ghstack-poisoned]
…n_policy`" This does some code organization improvement. - It renames `_FSDPPolicy` to `_Policy` to show that it is not only for FSDP but for any module-level API. - It formalizes the contract that such a policy should return something like `target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]` that maps each module to wrap to its kwargs. It does so by requiring a `_run_policy` abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call `_run_policy()` to generate the dict and do any validation or post-processing. This PR is technically BC-breaking because it removes the public `ModuleWrapPolicy.policy`. However, I do not think anyone was using that anyway, so this is a pretty safe breakage. [ghstack-poisoned]
ghstack-source-id: f0778e5f19edfb0a16e282f9cb133829badb591d Pull Request resolved: pytorch#104969
…n_policy`" This does some code organization improvement. - It renames `_FSDPPolicy` to `_Policy` to show that it is not only for FSDP but for any module-level API. - It formalizes the contract that such a policy should return something like `target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]` that maps each module to wrap to its kwargs. It does so by requiring a `_run_policy` abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call `_run_policy()` to generate the dict and do any validation or post-processing. This PR is technically BC-breaking because it removes the public `ModuleWrapPolicy.policy`. However, I do not think anyone was using that anyway, so this is a pretty safe breakage. [ghstack-poisoned]
Unstable test and unrelated: |
@pytorchbot merge -f "unrelated failure" |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
This PR adds a new `CustomPolicy` that acts like the existing `lambda_auto_wrap_policy` except it (1) leverages the new auto wrapping infrastructure and (2) allows overriding FSDP kwargs for particular instances. (1) gives it access to the validation checks (like for frozen parameters), and (2) makes it as expressive as manual wrapping. This should allow us to effectively deprecate manual wrapping if desired. The API is as follows: ``` def lambda_fn(module: nn.Module) -> Union[bool, Dict[str, Any]]: ... policy = CustomPolicy(lambda_fn) ``` The `lambda_fn` can return: - `False` or `{}` to indicate no wrapping - `True` to indicate wrapping while inheriting the root's FSDP kwargs - Non-empty `dict` to indicate wrapping while overriding the specified FSDP kwargs and inheriting the rest from the root --- After this PR, the follow-up work items for auto wrapping are: 1. Add shared parameter validation 2. (Longer-term / exploratory) Add a policy that provides a reasonable auto wrapping with "minimal" user input Pull Request resolved: #104986 Approved by: https://github.com/ezyang ghstack dependencies: #104427, #104967, #104999, #104969
Stack from ghstack (oldest at bottom):
CustomPolicy
#104986_FSDPPolicy.policy
with_Policy._run_policy
#104969ModuleWrapPolicy
to takeIterable
#104999ModuleWrapPolicy
#104427This does some code organization improvement.
_FSDPPolicy
to_Policy
to show that it is not only for FSDP but for any module-level API.target_module_to_kwargs: Dict[nn.Module, Dict[str, Any]]
that maps each module to wrap to its kwargs. It does so by requiring a_run_policy
abstract method (this time private since users do not need to care about it). Then, our auto wrapping can just call_run_policy()
to generate the dict and do any validation or post-processing.This PR is technically BC-breaking because it removes the public
ModuleWrapPolicy.policy
. However, I do not think anyone was using that anyway, so this is a pretty safe breakage.