update _unsafe_set_version_counter to accept lists of tensors #137921
Conversation
This file contains 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/137921
Note: Links to docs will display an error until the docs builds have been completed. ❌ 8 New Failures, 1 Cancelled Job, 10 Unrelated FailuresAs of commit ed893f9 with merge base 932ae13 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Sep 18, 2024
awgu
approved these changes
Oct 14, 2024
…ors" See the comment [here](#132014 (comment)) (cc awgu) - this PR updates `_unsafe_set_version_counter` to accept a list of tensors, for overhead-sensitive users (e.g. distributed) who need to hide VC bumps from autograd on a large list of tensors without wanting to suffer the overhead of going from python->C++ separately for every tensor in the list. I left the binding in pybind, and used a `std::vector`. if we **really** need to optimize overhead even further, we could write a manual cpython binding. I use this updated API in the next PR to fix FSDP2, so that it properly hides the VC of all `all_gather_buffer` tensors in its call to `split_with_sizes_copy.out(all_gather_buffers)`. cc XilunWu H-Huang awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
albanD
approved these changes
Oct 14, 2024
torch/autograd/grad_mode.py
Outdated
| def __init__(self, tensor: torch.Tensor) -> None: | ||
| self.tensor = tensor | ||
| self.prev_version = tensor._version | ||
| def __init__(self, tensors: Union[torch.Tensor, List[torch.Tensor]]) -> None: |
There was a problem hiding this comment.
Suggested change
| def __init__(self, tensors: Union[torch.Tensor, List[torch.Tensor]]) -> None: | |
| def __init__(self, tensors: Union[torch.Tensor, Tuple[torch.Tensor, ...]]) -> None: |
There was a problem hiding this comment.
You use both in your tests :p I would settle on tuple unless we often need the flexibility of a list.
| ) | ||
| # decrement version counters of all outputs at once | ||
| old_versions = [x._version - 1 for x in fsdp_param.all_gather_outputs] | ||
| torch._C._autograd._unsafe_set_version_counter( |
There was a problem hiding this comment.
nit: Use the preserve API here to get proper error handling?
There was a problem hiding this comment.
yeah agreed, will change (I remember @awgu saying he was worried about the overhead of the context manager, but we will now be doing a single call through the context manager, vs before we were doing N individual calls to pybind, so this should net out to being faster either way?)
There was a problem hiding this comment.
I feel like I saw another PR from someone else doing that change. @awgu am I dreaming here? :p
There was a problem hiding this comment.
…ors" See the comment [here](#132014 (comment)) (cc awgu) - this PR updates `_unsafe_set_version_counter` to accept a list of tensors, for overhead-sensitive users (e.g. distributed) who need to hide VC bumps from autograd on a large list of tensors without wanting to suffer the overhead of going from python->C++ separately for every tensor in the list. I left the binding in pybind, and used a `std::vector`. if we **really** need to optimize overhead even further, we could write a manual cpython binding. I use this updated API in the next PR to fix FSDP2, so that it properly hides the VC of all `all_gather_buffer` tensors in its call to `split_with_sizes_copy.out(all_gather_buffers)`. cc XilunWu H-Huang awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
…ors" See the comment [here](#132014 (comment)) (cc awgu) - this PR updates `_unsafe_set_version_counter` to accept a list of tensors, for overhead-sensitive users (e.g. distributed) who need to hide VC bumps from autograd on a large list of tensors without wanting to suffer the overhead of going from python->C++ separately for every tensor in the list. I left the binding in pybind, and used a `std::vector`. if we **really** need to optimize overhead even further, we could write a manual cpython binding. I use this updated API in the next PR to fix FSDP2, so that it properly hides the VC of all `all_gather_buffer` tensors in its call to `split_with_sizes_copy.out(all_gather_buffers)`. cc XilunWu H-Huang awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
…ors" See the comment [here](#132014 (comment)) (cc awgu) - this PR updates `_unsafe_set_version_counter` to accept a list of tensors, for overhead-sensitive users (e.g. distributed) who need to hide VC bumps from autograd on a large list of tensors without wanting to suffer the overhead of going from python->C++ separately for every tensor in the list. I left the binding in pybind, and used a `std::vector`. if we **really** need to optimize overhead even further, we could write a manual cpython binding. I use this updated API in the next PR to fix FSDP2, so that it properly hides the VC of all `all_gather_buffer` tensors in its call to `split_with_sizes_copy.out(all_gather_buffers)`. cc XilunWu H-Huang awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
…ors" See the comment [here](#132014 (comment)) (cc XilunWu H-Huang awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang rec) - this PR updates `_unsafe_set_version_counter` to accept a list of tensors, for overhead-sensitive users (e.g. distributed) who need to hide VC bumps from autograd on a large list of tensors without wanting to suffer the overhead of going from python->C++ separately for every tensor in the list. I left the binding in pybind, and used a `std::vector`. if we **really** need to optimize overhead even further, we could write a manual cpython binding. I use this updated API in the next PR to fix FSDP2, so that it properly hides the VC of all `all_gather_buffer` tensors in its call to `split_with_sizes_copy.out(all_gather_buffers)`. cc XilunWu H-Huang awgu kwen2501 wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
|
@pytorchbot merge |
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 |
Merge failedReason: 1 mandatory check(s) failed. The first few are: Dig deeper by viewing the failures on hud |
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.
See the comment here (cc @XilunWu @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @rec) - this PR updates
_unsafe_set_version_counterto accept a list of tensors, for overhead-sensitive users (e.g. distributed) who need to hide VC bumps from autograd on a large list of tensors without wanting to suffer the overhead of going from python->C++ separately for every tensor in the list.I left the binding in pybind, and used a
std::vector. if we really need to optimize overhead even further, we could write a manual cpython binding.I use this updated API in the next PR to fix FSDP2, so that it properly hides the VC of all
all_gather_buffertensors in its call tosplit_with_sizes_copy.out(all_gather_buffers).Stack from ghstack (oldest at bottom):
cc @XilunWu @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @rec