[Reland] Launch kernel on current stream & remove record_stream entirely
#150398
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
…irely Relanding #148590 due to merge conflict. This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related): 1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back. - Resolves #147729 - Resolves #146881 - Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user. 2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling. - Resolves #147168 3. Remove tensor life management when async_op=False; only use it when async_op=True. 4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](#147168 (comment)). 5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels. Joint work with @cenzhaometa who wants to remove the event sync overhead. Squashed contents: * [ptd][nccl] use current-stream as nccl-stream under async=False mode (#147820) PTD current workflow: - PTD creates its own dedicated `ncclStream` for comm operation - it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us). This diff: - async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead - async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready - pass down async from c10d down to NCCL-PG this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%** Differential Revision: D70135605 * [PGNCCL] Make avoid-record-stream default * [c10d] Add asyncOp argument to Ops * Change python side wait * Pass asyncOp at ProcessGroup level * Watchdog unstashing tensors as a safety net * Stash tensors for reduce_scatter_v and all_gather_v Pull Request approved: #149753 * [c10d] Move unstashing from watchdog to main thread Pull Request approved: #150079 * [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation Pull Request approved: #150130 [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/150398
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (2 Unrelated Failures)As of commit acf5139 with merge base 6470b37 ( UNSTABLE - The following jobs are marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
kwen2501
added a commit
that referenced
this pull request
Apr 1, 2025
…irely Relanding #148590 due to merge conflict. This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related): 1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back. - Resolves #147729 - Resolves #146881 - Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user. 2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling. - Resolves #147168 3. Remove tensor life management when async_op=False; only use it when async_op=True. 4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](#147168 (comment)). 5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels. Joint work with cenzhaometa who wants to remove the event sync overhead. Squashed contents: * [ptd][nccl] use current-stream as nccl-stream under async=False mode (#147820) PTD current workflow: - PTD creates its own dedicated `ncclStream` for comm operation - it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us). This diff: - async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead - async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready - pass down async from c10d down to NCCL-PG this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%** Differential Revision: D70135605 * [PGNCCL] Make avoid-record-stream default * [c10d] Add asyncOp argument to Ops * Change python side wait * Pass asyncOp at ProcessGroup level * Watchdog unstashing tensors as a safety net * Stash tensors for reduce_scatter_v and all_gather_v Pull Request approved: #149753 * [c10d] Move unstashing from watchdog to main thread Pull Request approved: #150079 * [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation Pull Request approved: #150130 ghstack-source-id: ce103fc Pull Request resolved: #150398
|
@kwen2501 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@pytorchmergebot merge -i |
Merge failedReason: This PR has internal changes and must be landed via Phabricator! Please try reimporting/rexporting the PR! Details for Dev Infra teamRaised by workflow job |
|
@kwen2501 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@pytorchmergebot merge -i |
Merge failedReason: This PR has internal changes and must be landed via Phabricator! Please try reimporting/rexporting the PR! Details for Dev Infra teamRaised by workflow job |
|
@pytorchmergebot merge -f "lint is green, diff will be landed with internal changes" |
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 |
Merge failedReason: This PR has internal changes and must be landed via Phabricator! Please try reimporting/rexporting the PR! Details for Dev Infra teamRaised by workflow job |
|
@pytorchmergebot merge -f "lint is green, diff will be landed with internal changes" |
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 |
|
Internal diff is D72224314 (contains internal changes) |
pytorchmergebot
pushed a commit
that referenced
this pull request
Apr 2, 2025
Update the torch-xpu-ops commit to [98c808dea6de7330c415aa777d6921944cf79887](intel/torch-xpu-ops@98c808d), include - Fixes #150001 by removing pre-CXX11 ABI logic from build script for XPU - Fixes #150430 - Fixes XCCL build issue caused by PR #150398 Pull Request resolved: #150554 Approved by: https://github.com/EikanWang, https://github.com/malfet
amathewc
pushed a commit
to amathewc/pytorch
that referenced
this pull request
Apr 17, 2025
…irely (pytorch#150398) Relanding pytorch#148590 due to merge conflict. This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related): 1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back. - Resolves pytorch#147729 - Resolves pytorch#146881 - Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user. 2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling. - Resolves pytorch#147168 3. Remove tensor life management when async_op=False; only use it when async_op=True. 4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](pytorch#147168 (comment)). 5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels. Joint work with @cenzhaometa who wants to remove the event sync overhead. Squashed contents: * [ptd][nccl] use current-stream as nccl-stream under async=False mode (pytorch#147820) PTD current workflow: - PTD creates its own dedicated `ncclStream` for comm operation - it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us). This diff: - async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead - async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready - pass down async from c10d down to NCCL-PG this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%** * [PGNCCL] Make avoid-record-stream default * [c10d] Add asyncOp argument to Ops * Change python side wait * Pass asyncOp at ProcessGroup level * Watchdog unstashing tensors as a safety net * Stash tensors for reduce_scatter_v and all_gather_v Pull Request approved: pytorch#149753 * [c10d] Move unstashing from watchdog to main thread Pull Request approved: pytorch#150079 * [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation Pull Request approved: pytorch#150130 Pull Request resolved: pytorch#150398 Approved by: https://github.com/atalman
amathewc
pushed a commit
to amathewc/pytorch
that referenced
this pull request
Apr 17, 2025
Update the torch-xpu-ops commit to [98c808dea6de7330c415aa777d6921944cf79887](intel/torch-xpu-ops@98c808d), include - Fixes pytorch#150001 by removing pre-CXX11 ABI logic from build script for XPU - Fixes pytorch#150430 - Fixes XCCL build issue caused by PR pytorch#150398 Pull Request resolved: pytorch#150554 Approved by: https://github.com/EikanWang, https://github.com/malfet
github-merge-queue bot
pushed a commit
to intel/torch-xpu-ops
that referenced
this pull request
May 16, 2025
Refer pytorch/pytorch#147820 pytorch/pytorch#150398 To launch kernels on the current stream and reduce the CPU overhead introduced by `recordStream`, an `async` option is introduced. For example, in an `allreduce` operation between two ranks: - `rank0` corresponds to `device0`, using the current device's `stream0` to create the communicator and preserving `stream0`. When `async = true`: - Both `rank0` and `rank1` perform the collective using `stream0`, which is associated with the communicator. - To prevent potential reads by `stream0` from unready tensors (e.g., from `rank1`), synchronization with the current stream is required. - After the collective completes, to prevent premature release of the input tensors, `recordStream` must be used for stream tracking, or the tensors need to be temporarily stored (e.g., in `reduce_scatter` or `all2all`). When `async = false`: - Both `rank0` and `rank1` use their respective **current streams** for collectives (i.e., `rank0` uses `stream0`, `rank1` uses `stream1`). - In this case, the collective op handles synchronization implicitly. Previously, we defaulted to `async = true`. Now, the `async` option is explicitly introduced and set to `false` by default, leveraging the current stream to avoid the overhead of stream synchronization. --------- Co-authored-by: mengfei25 <mengfei.li@Intel.com>
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):
record_streamentirely #150398Relanding #148590 due to merge conflict.
This PR has multiple changes to
ProcessGroupNCCL(which unfortunately are related):cudaStreamWaitEventin PGNCCL #146881work.wait()in distributed_c10d.py on behalf of user.record_streamand use CPU-side stashing for managing tensor lifetime against recycling.record_streamin c10d causes FSDP2 to over-allocate GPU memory #147168work.wait(), we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion here.Joint work with @cenzhaometa who wants to remove the event sync overhead.
Squashed contents:
PTD current workflow:
ncclStreamfor comm operationsuch stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us).
This diff:
this helps shave off 50% CPU overhead (70us -> 35us), which reduce total CPU/GPU from 230us to 195us by 15%
[PGNCCL] Make avoid-record-stream default
[c10d] Add asyncOp argument to Ops
Change python side wait
Pass asyncOp at ProcessGroup level
Watchdog unstashing tensors as a safety net
Stash tensors for reduce_scatter_v and all_gather_v
Pull Request approved: Stash tensors for reduce_scatter_v and all_gather_v #149753
[c10d] Move unstashing from watchdog to main thread
Pull Request approved: [c10d] Move unstashing from watchdog to main thread #150079
[PGNCCL][BE] Merge mutex into TensorShelf for encapsulation
Pull Request approved: [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation #150130
cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o
Differential Revision: D72224314