Stash tensors for reduce_scatter_v and all_gather_v #149753
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/149753
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit dc68c40 with merge base 8d08b49 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o [ghstack-poisoned]
#148590 removed record_stream. Since previous AVOID_RECORD flag does not cover reduce_scatter_v and all_gather_v which are in coalescing form, these two ops were missed. Causing TorchRec's Variable Length Embedding to fail. This PR adds a vector to stash tensors when coalescing is in flight. And the end of coalescing, it will hand over the tensors to Work. cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k c-p-i-o [ghstack-poisoned]
fduwjj
reviewed
Mar 21, 2025
|
|
||
| @requires_nccl() | ||
| @skip_but_pass_in_sandcastle_if(not TEST_MULTIGPU, "NCCL test requires 2+ GPUs") | ||
| def test_all_gather_v(self): |
kwen2501
added a commit
that referenced
this pull request
Mar 21, 2025
…147820) Summary: 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 lint ghstack-source-id: f391590 Pull Request resolved: #148590 Stash tensors for reduce_scatter_v and all_gather_v ghstack-source-id: f391590 Pull Request resolved: #149753
kwen2501
added a commit
that referenced
this pull request
Mar 28, 2025
…147820) Summary: 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 lint ghstack-source-id: e4b48e5 Pull Request resolved: #148590 Stash tensors for reduce_scatter_v and all_gather_v ghstack-source-id: e4b48e5 Pull Request resolved: #149753 [c10d] Move unstashing from watchdog to main thread ghstack-source-id: e4b48e5 Pull Request resolved: #150079 [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation ghstack-source-id: e4b48e5 Pull Request resolved: #150130
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-poisoned]
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
pytorchmergebot
pushed a commit
that referenced
this pull request
Apr 1, 2025
…irely (#150398) 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%** * [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 Pull Request resolved: #150398 Approved by: https://github.com/atalman
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
|
Squashed |
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 #148590#148590 removed record_stream. Since previous AVOID_RECORD flag does not cover reduce_scatter_v and all_gather_v which are in coalescing form, these two ops were missed. Causing TorchRec's Variable Length Embedding to fail.
This PR adds a vector to stash tensors when coalescing is in flight. And the end of coalescing, it will hand over the tensors to Work.
cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o