[FSDP] Introduce ModuleWrapPolicy for simplicity
#88450
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
awgu
requested review from
H-Huang,
kwen2501,
mrshenli,
pritamdamania87,
rohan-varma,
yhcharles and
zhaojuanmao
as code owners
pytorch-bot
bot
added
the
release notes: distributed (fsdp)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/88450
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit cb44a77: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`). `ModuleWrapPolicy` inherits from an abstract base class `AutoWrapPolicy` that expects an `auto_wrap_policy` property. This decouples the construct of such `AutoWrapPolicy` classes and their actual `auto_wrap_policy`, which must abide by the `_recursive_wrap` interface.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `AutoWrapPolicy` that expects an `auto_wrap_policy` property. This decouples the construct of such `AutoWrapPolicy` classes and their actual `auto_wrap_policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `AutoWrapPolicy`, so this approach is fully backward compatible from a functionality perspective.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
added 3 commits
November 10, 2022 18:45
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
zhaojuanmao
approved these changes
Nov 11, 2022
| mixed_precision: Optional[MixedPrecision] = None, | ||
| cpu_offload: Optional[CPUOffload] = None, | ||
| auto_wrap_policy: Optional[Callable] = None, | ||
| policy: Optional[_FSDPPolicy] = None, |
Contributor
There was a problem hiding this comment.
let's still keep the 'auto_wrap_policy' name for now? feel policy is too general, also it is a big BC change
Collaborator
Author
There was a problem hiding this comment.
This is composable FSDP. In my understanding, we should be able to change the constructor?
I wanted policy to be general because we can configure FSDP this way. This can be an entry point for different flavors of FSDP. One option to enable tensor shape preservation may be via policy.
Collaborator
Author
There was a problem hiding this comment.
The wrapper FullyShardedDataParallel still calls it auto_wrap_policy in its constructor.
Contributor
There was a problem hiding this comment.
oh I see, that sounds good then.
Collaborator
Author
|
I checked internal code. There is no code passing |
Collaborator
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Collaborator
Merge failedReason: 1 additional jobs have failed, first few of them are: TorchBench CI (pytorch-linux-py3.8-cu116) Details for Dev Infra teamRaised by workflow job |
Collaborator
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Collaborator
Merge failedReason: 1 additional jobs have failed, first few of them are: TorchBench CI (pytorch-linux-py3.8-cu116) Details for Dev Infra teamRaised by workflow job |
Collaborator
Author
|
|
Collaborator
Author
|
@pytorchbot merge -f "TorchBench CI (pytorch-linux-py3.8-cu116) is skipped but being incorrectly treated as failed" |
Collaborator
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Collaborator
Merge failedReason: Command Details for Dev Infra teamRaised by workflow job |
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
[ghstack-poisoned]
Collaborator
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
awgu
pushed a commit
to awgu/pytorch
that referenced
this pull request
Nov 12, 2022
ghstack-source-id: 3142eae Pull Request resolved: pytorch#88450
This was referenced Nov 12, 2022
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Dec 10, 2022
**BC Breaking Change**
This renames `unwrapped_params` to `nonwrapped_numel`. I prefer `nonwrapped` over `unwrapped` because "unwrap" suggests that some wrapping has been undone. I prefer `numel` over `params` because that is unit of measurement; I think we should keep "params" to refer to `nn.Parameter`s themselves.
This only breaks anything that passes `unwrapped_params` as a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on our `pytorch` code).
In a follow-up, I want to rename `min_num_params` to `min_nonwrapped_numel` in `size_based_auto_wrap_policy`, which is also BC breaking. Again, this is to differentiate between "params" being `nn.Parameter`s and "numel" being the unit for `param.numel()`.
**Overview**
This PR introduces `ModuleWrapPolicy` as a lightweight layer over the existing `transformer_auto_wrap_policy`. The most common auto wrapping paradigm is:
```
module_classes: Set[Type[nn.Module]] = ...
auto_wrap_policy = functools.partial(
transformer_auto_wrap_policy,
transformer_layer_cls=module_classes,
)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
Now, users can instead write:
```
auto_wrap_policy = ModuleWrapPolicy(module_classes)
fsdp_model = FSDP(model, auto_wrap_policy=auto_wrap_policy, ...)
```
This hides the unused arguments expected from the callable (`recurse` and `unwrapped_params`/`nonwrapped_numel`).
`ModuleWrapPolicy` inherits from an abstract base class `FSDPPolicy` that expects a `policy` property. This decouples the construct of such `FSDPPolicy` classes and their actual `policy`, which must abide by the `_recursive_wrap` interface. Any existing auto wrap policy can be rewritten as a class that inherits from `FSDPPolicy`, so this approach is fully backward compatible from a functionality perspective.
I call this base class `FSDPPolicy` to generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructing `FlatParameter`s, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument in `fully_shard()` to simply `policy` instead of `auto_wrap_policy`.
This PR migrates usages of `transformer_auto_wrap_policy` within our unit test suite to `ModuleWrapPolicy` as much as possible.
Pull Request resolved: pytorch#88450
Approved by: https://github.com/zhaojuanmao
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack:
ModuleWrapPolicy#88453 [Dynamo][FSDP] Migrate toModuleWrapPolicyModuleWrapPolicyfor simplicity #88450 [FSDP] IntroduceModuleWrapPolicyfor simplicityBC Breaking Change
This renames
unwrapped_paramstononwrapped_numel. I prefernonwrappedoverunwrappedbecause "unwrap" suggests that some wrapping has been undone. I prefernumeloverparamsbecause that is unit of measurement; I think we should keep "params" to refer tonn.Parameters themselves.This only breaks anything that passes
unwrapped_paramsas a keyword argument, but I did not see anything that did that (except the one internal benchmark file but that does not actually depend on ourpytorchcode).In a follow-up, I want to rename
min_num_paramstomin_nonwrapped_numelinsize_based_auto_wrap_policy, which is also BC breaking. Again, this is to differentiate between "params" beingnn.Parameters and "numel" being the unit forparam.numel().Overview
This PR introduces
ModuleWrapPolicyas a lightweight layer over the existingtransformer_auto_wrap_policy. The most common auto wrapping paradigm is:Now, users can instead write:
This hides the unused arguments expected from the callable (
recurseandunwrapped_params/nonwrapped_numel).ModuleWrapPolicyinherits from an abstract base classFSDPPolicythat expects apolicyproperty. This decouples the construct of suchFSDPPolicyclasses and their actualpolicy, which must abide by the_recursive_wrapinterface. Any existing auto wrap policy can be rewritten as a class that inherits fromFSDPPolicy, so this approach is fully backward compatible from a functionality perspective.I call this base class
FSDPPolicyto generalize over the cases where we may not want to actually perform any nested wrapping. In reality, the policy is meant for constructingFlatParameters, which just happened to be induced by a nested wrapping before. Given this, I am changing the constructor argument infully_shard()to simplypolicyinstead ofauto_wrap_policy.This PR migrates usages of
transformer_auto_wrap_policywithin our unit test suite toModuleWrapPolicyas much as possible.