[FSDP][2/N] _summon_full_params -> _unshard_params
#92297
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]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/92297
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit c8420cc: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
release notes: distributed (fsdp)
This was referenced Jan 17, 2023
awgu
pushed a commit
to awgu/pytorch
that referenced
this pull request
Jan 17, 2023
ghstack-source-id: 998f0f5 Pull Request resolved: pytorch#92297
awgu
requested review from
H-Huang,
kwen2501,
mrshenli,
rohan-varma,
wanchaol and
zhaojuanmao
as code owners
**Overview**
This PR stack will add support for unsharding FSDP's sharded parameters for `fully_shard`. This PR takes the first step by doing some internal refactoring.
- The existing API for wrapper FSDP is the static method `summon_full_params()`, which calls into the helper `_summon_full_params()`.
- This PR refactors:
- `summon_full_params()` core logic to `_unshard_params()`
- `_summon_full_params()` to `_unshard_params_recurse()`, which has a `recurse: bool` argument
- Previous `_unshard_params()` to `_unshard_fsdp_state_params()`, which applies to a single FSDP state
**Details**
- This PR introduces `_get_fsdp_states_with_modules()` and `_get_root_fsdp_states_with_modules()`, which additionally return the modules along with the FSDP states. The modules are needed for handling `FlatParameter` registration.
- We may be able to remove this if we clean up the `use_orig_params=True` vs. `False` code paths because for `True`, the `FlatParameter` is not registered, meaning that it does not need to be de-registered.
- Since `fully_shard` requires `use_orig_params=True`, we may not need `_get_fsdp_states_with_modules()` and `_get_root_fsdp_root_modules()`; however, I prefer to make the separation of FSDP state and module explicit for now for clarity.
**Follow-Ups**
- `writeback=True` and `rank0_only=True` raises an error. The previous explanation was:
> is not supported, as model parameter shapes will be different across ranks, and writing to them can lead to inconsistencies across ranks when the context is exited.
I am not exactly sure what the different model parameter shapes refers to. However, I believe that we can support `writeback=True` and `rank0_only=True` by broadcasting the `FlatParameter` from rank 0 in the `finally`, writing back, and freeing. This should not increase the peak memory since rank 0 already holds the unsharded `FlatParameter` in GPU memory before writing back and nonzero ranks do not have any other unsharded `FlatParameter`s in GPU memory.
[ghstack-poisoned]
**Overview**
This PR stack will add support for unsharding FSDP's sharded parameters for `fully_shard`. This PR takes the first step by doing some internal refactoring.
- The existing API for wrapper FSDP is the static method `summon_full_params()`, which calls into the helper `_summon_full_params()`.
- This PR refactors:
- `summon_full_params()` core logic to `_unshard_params()`
- `_summon_full_params()` to `_unshard_params_recurse()`, which has a `recurse: bool` argument
- Previous `_unshard_params()` to `_unshard_fsdp_state_params()`, which applies to a single FSDP state
**Details**
- This PR introduces `_get_fsdp_states_with_modules()` and `_get_root_fsdp_states_with_modules()`, which additionally return the modules along with the FSDP states. The modules are needed for handling `FlatParameter` registration.
- We may be able to remove this if we clean up the `use_orig_params=True` vs. `False` code paths because for `True`, the `FlatParameter` is not registered, meaning that it does not need to be de-registered.
- Since `fully_shard` requires `use_orig_params=True`, we may not need `_get_fsdp_states_with_modules()` and `_get_root_fsdp_root_modules()`; however, I prefer to make the separation of FSDP state and module explicit for now for clarity.
**Follow-Ups**
- `writeback=True` and `rank0_only=True` raises an error. The previous explanation was:
> is not supported, as model parameter shapes will be different across ranks, and writing to them can lead to inconsistencies across ranks when the context is exited.
I am not exactly sure what the different model parameter shapes refers to. However, I believe that we can support `writeback=True` and `rank0_only=True` by broadcasting the `FlatParameter` from rank 0 in the `finally`, writing back, and freeing. This should not increase the peak memory since rank 0 already holds the unsharded `FlatParameter` in GPU memory before writing back and nonzero ranks do not have any other unsharded `FlatParameter`s in GPU memory.
[ghstack-poisoned]
**Overview**
This PR stack will add support for unsharding FSDP's sharded parameters for `fully_shard`. This PR takes the first step by doing some internal refactoring.
- The existing API for wrapper FSDP is the static method `summon_full_params()`, which calls into the helper `_summon_full_params()`.
- This PR refactors:
- `summon_full_params()` core logic to `_unshard_params()`
- `_summon_full_params()` to `_unshard_params_recurse()`, which has a `recurse: bool` argument
- Previous `_unshard_params()` to `_unshard_fsdp_state_params()`, which applies to a single FSDP state
**Details**
- This PR introduces `_get_fsdp_states_with_modules()` and `_get_root_fsdp_states_with_modules()`, which additionally return the modules along with the FSDP states. The modules are needed for handling `FlatParameter` registration.
- We may be able to remove this if we clean up the `use_orig_params=True` vs. `False` code paths because for `True`, the `FlatParameter` is not registered, meaning that it does not need to be de-registered.
- Since `fully_shard` requires `use_orig_params=True`, we may not need `_get_fsdp_states_with_modules()` and `_get_root_fsdp_root_modules()`; however, I prefer to make the separation of FSDP state and module explicit for now for clarity.
**Follow-Ups**
- `writeback=True` and `rank0_only=True` raises an error. The previous explanation was:
> is not supported, as model parameter shapes will be different across ranks, and writing to them can lead to inconsistencies across ranks when the context is exited.
I am not exactly sure what the different model parameter shapes refers to. However, I believe that we can support `writeback=True` and `rank0_only=True` by broadcasting the `FlatParameter` from rank 0 in the `finally`, writing back, and freeing. This should not increase the peak memory since rank 0 already holds the unsharded `FlatParameter` in GPU memory before writing back and nonzero ranks do not have any other unsharded `FlatParameter`s in GPU memory.
[ghstack-poisoned]
rohan-varma
approved these changes
Feb 1, 2023
Contributor
rohan-varma
left a comment
rohan-varma
left a comment
There was a problem hiding this comment.
Shall we add unittests for summon_full_params composable path?
| "to them can lead to inconsistencies across ranks when the " | ||
| "context is exited." | ||
| ) | ||
| # TODO: Rank 0 can broadcast the `FlatParameter` to allow all ranks to |
Contributor
There was a problem hiding this comment.
could we file an issue for this? would it work for use_orig_params=True as well?
Collaborator
Author
There was a problem hiding this comment.
I think it should work for both use_orig_params=True and False. I will file an issue.
| if recurse: | ||
| with contextlib.ExitStack() as stack: | ||
| # TODO (awgu): The traversal function does not traverse through | ||
| # incompatible composable APIs. Verify if this is the desired |
Contributor
There was a problem hiding this comment.
Could you elaborate, what's an example of this?
Collaborator
Author
There was a problem hiding this comment.
fully_shard(
Module(
replicate(
Submodule(
fully_shard(Subsubmodule),
Subsubmodule,
),
Submodule,
)
Because the traversal utils do not go through incompatible composable APIs (here, replicate), calling _unshard_params on the root Module will not unshard the parameters of the fully sharded Subsubmodule.
Collaborator
Author
Yes, this has not been added yet. (I have a local [4/N] commit that does add a frontend for that path, but I did not open a PR for it since we have not finalized what the API should look like.) I will add tests when we include that. |
**Overview**
This PR stack will add support for unsharding FSDP's sharded parameters for `fully_shard`. This PR takes the first step by doing some internal refactoring.
- The existing API for wrapper FSDP is the static method `summon_full_params()`, which calls into the helper `_summon_full_params()`.
- This PR refactors:
- `summon_full_params()` core logic to `_unshard_params()`
- `_summon_full_params()` to `_unshard_params_recurse()`, which has a `recurse: bool` argument
- Previous `_unshard_params()` to `_unshard_fsdp_state_params()`, which applies to a single FSDP state
**Details**
- This PR introduces `_get_fsdp_states_with_modules()` and `_get_root_fsdp_states_with_modules()`, which additionally return the modules along with the FSDP states. The modules are needed for handling `FlatParameter` registration.
- We may be able to remove this if we clean up the `use_orig_params=True` vs. `False` code paths because for `True`, the `FlatParameter` is not registered, meaning that it does not need to be de-registered.
- Since `fully_shard` requires `use_orig_params=True`, we may not need `_get_fsdp_states_with_modules()` and `_get_root_fsdp_root_modules()`; however, I prefer to make the separation of FSDP state and module explicit for now for clarity.
**Follow-Ups**
- `writeback=True` and `rank0_only=True` raises an error. The previous explanation was:
> is not supported, as model parameter shapes will be different across ranks, and writing to them can lead to inconsistencies across ranks when the context is exited.
I am not exactly sure what the different model parameter shapes refers to. However, I believe that we can support `writeback=True` and `rank0_only=True` by broadcasting the `FlatParameter` from rank 0 in the `finally`, writing back, and freeing. This should not increase the peak memory since rank 0 already holds the unsharded `FlatParameter` in GPU memory before writing back and nonzero ranks do not have any other unsharded `FlatParameter`s in GPU memory.
[ghstack-poisoned]
awgu
added
ciflow/trunk
ragulpr
added a commit
to ragulpr/pytorch
that referenced
this pull request
Feb 2, 2023
…n-dev-setup * origin: (898 commits) Move dynamo.optimizations.distributed to backends (pytorch#93408) Remove cuda 11.6 from nightly (pytorch#93979) Refactor dynamo register_backend/BACKENDS (pytorch#93389) Remove cuda 11.6 from CI replace with 11.7 (pytorch#93406) [Dynamo] Rename `GuardBuilder.guarded_code` -> `check_fn_manager` (pytorch#93934) Revert "Remove CUDA 11.6 from nightly builds (pytorch#93404)" Revert "[inductor] fix crash issue when input is a view tensor (pytorch#90150)" Basic Validation for FSDP `state_dict` transformations of modules with persistent buffers (pytorch#93396) Merge Inductor perf smoke test with other inductor CI tests (pytorch#93395) [inductor] Don't import torchvision (pytorch#93027) [FSDP][3/N] Refactor `summon_full_params` unit tests (pytorch#92298) [FSDP][2/N] `_summon_full_params` -> `_unshard_params` (pytorch#92297) Remove CUDA 11.6 from nightly builds (pytorch#93404) Mark buffers that reuse other buffers (pytorch#93329) Refactor to allow reuse of SchedulerNode.allocate (pytorch#93328) retire sparse_mask_helper (pytorch#91714) update fbgemm third party (pytorch#93907) [inductor] fix crash issue when input is a view tensor (pytorch#90150) [Inductor] add config for weight prepacking (pytorch#93811) Check for none for NNModuleVariable.__module__ (pytorch#93326) ...
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:
summon_full_paramsunit tests #92298 [FSDP][3/N] Refactorsummon_full_paramsunit tests_summon_full_params->_unshard_params#92297 [FSDP][2/N]_summon_full_params->_unshard_paramsOverview
This PR stack will add support for unsharding FSDP's sharded parameters for
fully_shard. This PR takes the first step by doing some internal refactoring.summon_full_params(), which calls into the helper_summon_full_params().summon_full_params()core logic to_unshard_params()_summon_full_params()to_unshard_params_recurse(), which has arecurse: boolargument_unshard_params()to_unshard_fsdp_state_params(), which applies to a single FSDP stateDetails
_get_fsdp_states_with_modules()and_get_root_fsdp_states_with_modules(), which additionally return the modules along with the FSDP states. The modules are needed for handlingFlatParameterregistration.use_orig_params=Truevs.Falsecode paths because forTrue, theFlatParameteris not registered, meaning that it does not need to be de-registered.fully_shardrequiresuse_orig_params=True, we may not need_get_fsdp_states_with_modules()and_get_root_fsdp_root_modules(); however, I prefer to make the separation of FSDP state and module explicit for now for clarity.Follow-Ups
writeback=Trueandrank0_only=Trueraises an error. The previous explanation was:I am not exactly sure what the different model parameter shapes refers to. However, I believe that we can support
writeback=Trueandrank0_only=Trueby broadcasting theFlatParameterfrom rank 0 in thefinally, writing back, and freeing. This should not increase the peak memory since rank 0 already holds the unshardedFlatParameterin GPU memory before writing back and nonzero ranks do not have any other unshardedFlatParameters in GPU memory.