[FSDP] summon offload to CPU #73904
Closed
[FSDP] summon offload to CPU #73904
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
Implement ability to offload full params to CPU in summon_full_params. Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/) [ghstack-poisoned]
rohan-varma
requested review from
H-Huang,
mingzhe09088,
mrshenli,
pritamdamania87 and
zhaojuanmao
as code owners
CI Flow Status⚛️ CI FlowRuleset - Version:
|
Contributor
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit b21c4b2 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
facebook-github-bot
added
the
oncall: distributed
rohan-varma
added a commit
that referenced
this pull request
Mar 8, 2022
Implement ability to offload full params to CPU in summon_full_params. Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/) ghstack-source-id: 150769189 Pull Request resolved: #73904
Implement ability to offload full params to CPU in summon_full_params. Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 8, 2022
Pull Request resolved: #73904 Implement ability to offload full params to CPU in summon_full_params. ghstack-source-id: 150794027 Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/)
fegin
approved these changes
Mar 8, 2022
Contributor
fegin
left a comment
fegin
left a comment
There was a problem hiding this comment.
LGTM, one comment regarding the warning of the combination of rank_0 + cpu_offload
| offload_to_cpu (bool, optional): If ``True``, full parameters are | ||
| offloaded to CPU. Note that this offloading currently only | ||
| occurs if the parameter is sharded (which is only not the case | ||
| for world_size = 1). |
Contributor
There was a problem hiding this comment.
In some rare cases, users may want to actually have multiple CPU copies, but we should add a note, warning or check to warn users if offload_to_cpu is True but rank0_only is False. At least users should be fully aware the potential CPU OOM issue that can happen.
Contributor
Author
There was a problem hiding this comment.
sounds good, added a warning
Implement ability to offload full params to CPU in summon_full_params. Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 9, 2022
Pull Request resolved: #73904 Implement ability to offload full params to CPU in summon_full_params. ghstack-source-id: 150949235 Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/)
rohan-varma
commented
Mar 9, 2022
| ) | ||
| self._update_p_data( | ||
| p, output_tensor=p._full_param_padded, # type: ignore[attr-defined] | ||
| ) |
Contributor
Author
There was a problem hiding this comment.
@zhaojuanmao just want to double check this offloading approach since it's a bit different than regular parameter offload.
Here we have full_param_padded materialized, and p.data points to it. So to offload, we:
- offload full param padded
- Update p.data to point to the offloaded full param padded.
And we restore appropriately when exiting the context.
If we directly offloaded p.data, it would still hold onto GPU memory because of full param padded.
Contributor
There was a problem hiding this comment.
yeah, this looks good to me!
zhaojuanmao
reviewed
Mar 9, 2022
Contributor
zhaojuanmao
left a comment
zhaojuanmao
left a comment
There was a problem hiding this comment.
left some minor comments, looks great!
in terms of enabling these configs in state_dict() APIs, could we make rank0_only as default setting and cpu_offloading as optional?
| ) | ||
|
|
||
| if writeback: | ||
| self._write_back_current_shard() |
Contributor
There was a problem hiding this comment.
just realized when fsdp main API cpu offloading is enabled, if writeback=True, it will make parameters point to local shard in GPUs, not local shards in CPUs. Which looks like a bug? maybe we can restrict writeback for non cpu offloading only?
Contributor
Author
There was a problem hiding this comment.
I'll look into this and file an issue.
fegin
approved these changes
Mar 10, 2022
Implement ability to offload full params to CPU in summon_full_params. Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/) [ghstack-poisoned]
rohan-varma
added a commit
that referenced
this pull request
Mar 14, 2022
Pull Request resolved: #73904 Implement ability to offload full params to CPU in summon_full_params. ghstack-source-id: 151311236 Differential Revision: [D34707801](https://our.internmc.facebook.com/intern/diff/D34707801/)
facebook-github-bot
pushed a commit
that referenced
this pull request
Mar 15, 2022
Summary: Pull Request resolved: #73904 Implement ability to offload full params to CPU in summon_full_params. ghstack-source-id: 151311236 Test Plan: CI Reviewed By: fegin Differential Revision: D34707801 fbshipit-source-id: efc9c568037ddfeb9ddaff45a1a0388c6bc85825
Contributor
|
Hey @rohan-varma. |
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 (oldest at bottom):
Implement ability to offload full params to CPU in summon_full_params.
Differential Revision: D34707801