AsyncCollectiveTensor: prevent wait_tensor() calls on graph inputs from getting DCEd #125677
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
…ompiled fw [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/125677
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit edba89d with merge base ab80a59 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
bdhirsh
added a commit
that referenced
this pull request
May 7, 2024
…ompiled fw ghstack-source-id: 5704a7ddf8560e00cd6bcba6ef026d7fd9ed8872 Pull Request resolved: #125677
github-actions
bot
requested review from
albanD,
antoniojkim,
ezyang,
miladm and
SherlockNoMad
…executing compiled fw" This patch is relatively low LoC but I'm not very satisfied with it (internal post coming soon) [ghstack-poisoned]
pytorch-bot
bot
added
the
oncall: distributed
bdhirsh
added a commit
that referenced
this pull request
May 7, 2024
…ompiled fw ghstack-source-id: 7e72f74afe68922c80ca817948c97aaec3fc993e Pull Request resolved: #125677
bdhirsh
changed the title
wanchaol
approved these changes
May 7, 2024
|
Very nice fix! Thank you! |
yifuwang
approved these changes
May 7, 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 |
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Outdated suggestions cannot be applied.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.
@wanchaol was seeing the loss eventually become NaN when compiling individual transformer blocks in torchtitan - with this patch I no longer see the NaN loss.
The problem is the following:
(1) It is possible to have graph inputs to a compiled region that are AsyncCollectiveTensors. In particular: when we compile individual transformer blocks in the llama model, the first layer (embedding layer) is run in eager mode, and it outputs an AsyncCollectiveTensor that is fed to the first transformer block
(2) ideally, we would like that AsyncCollectiveTensor graph input to desugar into a
wait_tensor()op that shows up at the beginning of the graph.(3) the way this is supposed to happen is: AOTAutograd traces through the torch_dispatch of AsyncCollectiveTensor, tracing out a
wait_tensor()call before dispatching to any of the other ops in the function we are tracing(4) however:
trigger_wait()was getting called in a way where we would ignore its output (and returnself.elemdirectly), which would cause thewait_tensorops to get DCE'd.Stack from ghstack (oldest at bottom):
cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @awgu @penguinwu @fegin @XilunWu @wanchaol @fduwjj @wz337 @tianyu-l @wconstab @yf225 @chauhang @d4l3k