[DSD] Improve the performance of distributed state_dict #125501
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/125501
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit ceb9a04 with merge base 746da87 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| @@ -132,6 +131,7 @@ class _StateDictInfo(StateDictOptions): | |||
| fsdp_modules: List[nn.Module] = field(default_factory=list) | |||
|
|
|||
|
|
|||
| @functools.lru_cache(maxsize=None) | |||
There was a problem hiding this comment.
would it be safer to set a maxsize here? otherwise this could lead to a (probably very small) memleak?
There was a problem hiding this comment.
When maxsize is None, the performance is faster as this means no LRU is used. functools.cache can achieve the same behavior but it is available only from Python 3.9. Theoretically, the get_state_dict should have the module every time so the cache of _get_fqns should stabilize after the first get_state_dict.
There was a problem hiding this comment.
I want to flag that this is not the case if get_state_dict is being used for different models in the same endpoint, and will likely be leaky.
|
@pytorchbot merge -f "The failing test is not related." |
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 |
Summary: If an object only exists on certain non-coordinator ranks, we still need to save them. Otherwise, we lose these objects. If they are duplicated, DCP will deduplicate them. Pull Request resolved: #125334 Approved by: https://github.com/wz337, https://github.com/LucasLLC ghstack dependencies: #125333, #125501
Summary: Right now DCP only flatten a mapping (e.g., dict) if that mapping has tensor objects. This behavior is odd as users may save different non-tensor objects on different ranks. Without flattening the mappings, we may lose these non-tensor objects. One use case is dataloader state_dict. We may also want to do so for a list/tuple. But this will cause extra pickles. So we don't do this for now. Pull Request resolved: #125335 Approved by: https://github.com/LucasLLC, https://github.com/wz337 ghstack dependencies: #125333, #125501, #125334
Summary: distributed_state_dict should not try to use `getattr` to get `_extra_state` as this is not well-defined. Pull Request resolved: pytorch#125336 Approved by: https://github.com/LucasLLC ghstack dependencies: pytorch#125333, pytorch#125501, pytorch#125334, pytorch#125335
* [DSD] Correctly handle _extra_state (#125336) Summary: distributed_state_dict should not try to use `getattr` to get `_extra_state` as this is not well-defined. Pull Request resolved: #125336 Approved by: https://github.com/LucasLLC ghstack dependencies: #125333, #125501, #125334, #125335 * lint * lint --------- Co-authored-by: Chien-Chin Huang <chienchin@fb.com> Co-authored-by: Andrey Talman <atalman@fb.com>
…ytorch#125337) Summary: Fixes pytorch#122792 state_dict includes only persistent buffers, while named_buffers() would include non_persistent buffers. Pull Request resolved: pytorch#125337 Approved by: https://github.com/awgu ghstack dependencies: pytorch#125333, pytorch#125501, pytorch#125334, pytorch#125335, pytorch#125336
…125337) (#127219) * [DSD] Fix to remove non_persistent buffer in distributed state dict (#125337) Summary: Fixes #122792 state_dict includes only persistent buffers, while named_buffers() would include non_persistent buffers. Pull Request resolved: #125337 Approved by: https://github.com/awgu ghstack dependencies: #125333, #125501, #125334, #125335, #125336 * lintrunner * lint --------- Co-authored-by: Chien-Chin Huang <chienchin@fb.com> Co-authored-by: Andrey Talman <atalman@fb.com>
Stack from ghstack (oldest at bottom):
Summary:
cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @awgu @penguinwu @XilunWu @wanchaol @fduwjj @wz337 @tianyu-l @wconstab @yf225 @chauhang @d4l3k @LucasLLC