[dtensor] fix dtensor _to_copy op for mix precision #116426
Closed
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/116426
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 8579064 with merge base ca4df16 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Dec 26, 2023
github-actions
bot
added
oncall: distributed
cc mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu fduwjj wz337 tianyu-l wconstab yf225 [ghstack-poisoned]
cc mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu fduwjj wz337 tianyu-l wconstab yf225 [ghstack-poisoned]
fduwjj
approved these changes
Jan 2, 2024
bdhirsh
reviewed
Jan 2, 2024
|
|
||
| register_op_strategy( | ||
| aten._to_copy.default, schema_info=RuntimeSchemaInfo(static_kwargkey=["dtype"]) | ||
| )(default_strategy) |
There was a problem hiding this comment.
Mind if I confirm my understanding of the fix (just out of interest :) )
It looks like static_kwargkey ends up getting used in the cache-lookup for DTensor's sharding prop: https://github.com/pytorch/pytorch/blob/main/torch/distributed/_tensor/op_schema.py#L299
So was the problem that we had a graph with _to_copy() showing up twice, with different dtypes, and we ended up using the cached sharding strategy, when we should have recomputed it for a different dtype?
There was a problem hiding this comment.
Yes! this is correct, we should recomputed the dtype in this case but it was reusing the cached type
cc mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse H-Huang kwen2501 awgu penguinwu fegin XilunWu fduwjj wz337 tianyu-l wconstab yf225 [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Jan 3, 2024
Context: Existing FSDPExtension have some bug in the case when the unflatten tensor involves some compute/communications in cuda stream, the current logic of FSDPExtension unflatten tensor happens in the unshard stream, which makes runtime lost sync with the compute stream, and if there're some dependencies between the compute stream and the unflatten tensor logic, currently it would lose sync point, which could possibly lead to NaN. This PR make the FSDPExtension to record the compute stream and let DTensorExtension to directly use the compute stream for unflatten_tensor. In long term we might want to directly make the FSDP runtime logic to only make the unshard happen in unshard stream, and use unshard views to happen in the compute stream. We currently fix this in the Extension directly as this is the simplest thing to do without affecting FSDP runtime logic Pull Request resolved: #116559 Approved by: https://github.com/awgu, https://github.com/fduwjj, https://github.com/yifuwang ghstack dependencies: #116426
pytorchmergebot
pushed a commit
that referenced
this pull request
Jan 3, 2024
Disable some runtime assertion first as it does not work with torch.compile properly, I'll have a follow up fix in dynamo and reenable this check again Pull Request resolved: #116573 Approved by: https://github.com/awgu, https://github.com/XilunWu ghstack dependencies: #116426, #116559
pytorchmergebot
pushed a commit
that referenced
this pull request
Jan 3, 2024
This PR adds devices to register_backend of multithraeded pg, to avoid seeing tons of warnings. Pull Request resolved: #116678 Approved by: https://github.com/awgu, https://github.com/XilunWu ghstack dependencies: #116426, #116559, #116573
Skylion007
pushed a commit
to Skylion007/pytorch
that referenced
this pull request
Feb 12, 2024
Pull Request resolved: pytorch#116426 Approved by: https://github.com/fduwjj
This was referenced Feb 12, 2024
atalman
pushed a commit
that referenced
this pull request
Feb 14, 2024
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):
cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @rohan-varma @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @awgu @penguinwu @fegin @XilunWu @fduwjj @wz337 @tianyu-l @wconstab @yf225