[FSDP2] Changed test_register_forward_method to use multiprocess test
#128100
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/128100
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 16249ba with merge base 2fc9079 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
oncall: distributed
awgu
commented
Jun 6, 2024
|
|
||
| torch.manual_seed(42) | ||
| model = Model() | ||
| for param in model.parameters(): |
weifengpy
approved these changes
Jun 6, 2024
|
@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 |
pytorchmergebot
pushed a commit
that referenced
this pull request
Jun 6, 2024
There is still ongoing discussion on how this API should work.
Current approach:
- The pre-all-gather ops run in the default stream and the all-gather is called from the default stream with `async_op=True`.
- Pros:
- The all-gather input and output tensors are allocated in the default stream, so there is no increased memory fragmentation across stream pools.
- There is no need for additional CUDA synchronization. The API is self-contained.
- Cons:
- The pre-all-gather ops (e.g. cast from fp32 -> bf16 and all-gather copy-in device copies) cannot overlap with other default stream compute. The biggest concern here is for CPU offloading, the H2D copies cannot overlap.
Alternative approach:
- Follow the default implicit prefetching approach, where the pre-all-gather ops and all-gather run in separate streams.
- Pros:
- The pre-all-gather ops can overlap with default stream compute.
- Cons:
- We require an API that should be called after the last optimizer step (namely, last op that modified sharded parameters) and before the first `unshard` call that has the all-gather streams wait for the default stream. The API is no longer self-contained and now has a complementary API.
- The all-gather input and output tensors are allocated in separate streams (not the default stream), so there can be increased memory fragmentation across pools.
Pull Request resolved: #128138
Approved by: https://github.com/wanchaol
ghstack dependencies: #128100
TharinduRusira
pushed a commit
to TharinduRusira/pytorch
that referenced
this pull request
Jun 14, 2024
…st (pytorch#128100) The test seems to be flaky due to multi-threaded process group. This PR converts the test to use normal multi-process `ProcessGroupNCCL` to fix the flakiness. This PR closes pytorch#126851. Interestingly, the original MTPG version passes for me on devgpu. Either way, the new version also passes on devgpu, so we can see in CI. Pull Request resolved: pytorch#128100 Approved by: https://github.com/weifengpy
TharinduRusira
pushed a commit
to TharinduRusira/pytorch
that referenced
this pull request
Jun 14, 2024
There is still ongoing discussion on how this API should work.
Current approach:
- The pre-all-gather ops run in the default stream and the all-gather is called from the default stream with `async_op=True`.
- Pros:
- The all-gather input and output tensors are allocated in the default stream, so there is no increased memory fragmentation across stream pools.
- There is no need for additional CUDA synchronization. The API is self-contained.
- Cons:
- The pre-all-gather ops (e.g. cast from fp32 -> bf16 and all-gather copy-in device copies) cannot overlap with other default stream compute. The biggest concern here is for CPU offloading, the H2D copies cannot overlap.
Alternative approach:
- Follow the default implicit prefetching approach, where the pre-all-gather ops and all-gather run in separate streams.
- Pros:
- The pre-all-gather ops can overlap with default stream compute.
- Cons:
- We require an API that should be called after the last optimizer step (namely, last op that modified sharded parameters) and before the first `unshard` call that has the all-gather streams wait for the default stream. The API is no longer self-contained and now has a complementary API.
- The all-gather input and output tensors are allocated in separate streams (not the default stream), so there can be increased memory fragmentation across pools.
Pull Request resolved: pytorch#128138
Approved by: https://github.com/wanchaol
ghstack dependencies: pytorch#128100
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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):
test_register_forward_methodto use multiprocess test #128100The test seems to be flaky due to multi-threaded process group. This PR converts the test to use normal multi-process
ProcessGroupNCCLto fix the flakiness.This PR closes #126851.
Interestingly, the original MTPG version passes for me on devgpu. Either way, the new version also passes on devgpu, so we can see in CI.
cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @penguinwu @fegin @XilunWu @wanchaol @fduwjj @wz337 @tianyu-l @wconstab @yf225 @chauhang @d4l3k