One-line change to correctly dispatch to cpu function for inference #2881
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
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D48574563 |
|
This pull request was exported from Phabricator. Differential Revision: D48574563 |
TroyGarden
force-pushed
the
export-D48574563
branch
from
1b3573e to
56d5696
Compare
|
This pull request was exported from Phabricator. Differential Revision: D48574563 |
TroyGarden
added a commit
to TroyGarden/FBGEMM
that referenced
this pull request
Jul 24, 2024
…ytorch#2881) Summary: Pull Request resolved: pytorch#2881 # context * the fundamental issue is that a "dispatch_to_cpu" function is calling the "autograd" function * a side issue is the poor naming: `permute_pooled_embs_auto_grad` sounds like an "autograd version" but actually as an operator the real backend (CPU, GPU, META, AUTOGRAD, etc.) is determined by the dispatcher. * this defact exists from day-1 since this operator was developed (D31271923). * the impact is that in the inference flow where the autograd is not needed, although the dispatcher does call the CPU version of this operator, the actual function calls the autograd function, which is not desired. # how to do it properly: * For a user-faced operator, we first need to give it a good name, like `permute_multi_embedding`. * then we need to implement functions for the necessary backends, commonly 4: CPU, CUDA, META, and AUTOGRAD * it's better to name the function as {operator_name}_{backend}, and **very important to link them correctly**, here is a good example: D48720379 * in training, the dispatcher will **ALWAYS** pickup the autograd version, then [calls cuda/cpu/meta version from the autograd function](https://fburl.com/code/2d1acuop) based on device or other context. * in inference, the dispatcher will ignore the autograd, and directly calls the cpu/cuda/meta version Reviewed By: dstaay-fb, sryap Differential Revision: D48574563
TroyGarden
force-pushed
the
export-D48574563
branch
from
56d5696 to
dbd0569
Compare
…ytorch#2881) Summary: Pull Request resolved: pytorch#2881 # context * the fundamental issue is that a "dispatch_to_cpu" function is calling the "autograd" function * a side issue is the poor naming: `permute_pooled_embs_auto_grad` sounds like an "autograd version" but actually as an operator the real backend (CPU, GPU, META, AUTOGRAD, etc.) is determined by the dispatcher. * this defact exists from day-1 since this operator was developed (D31271923). * the impact is that in the inference flow where the autograd is not needed, although the dispatcher does call the CPU version of this operator, the actual function calls the autograd function, which is not desired. # how to do it properly: * For a user-faced operator, we first need to give it a good name, like `permute_multi_embedding`. * then we need to implement functions for the necessary backends, commonly 4: CPU, CUDA, META, and AUTOGRAD * it's better to name the function as {operator_name}_{backend}, and **very important to link them correctly**, here is a good example: D48720379 * in training, the dispatcher will **ALWAYS** pickup the autograd version, then [calls cuda/cpu/meta version from the autograd function](https://fburl.com/code/2d1acuop) based on device or other context. * in inference, the dispatcher will ignore the autograd, and directly calls the cpu/cuda/meta version Reviewed By: dstaay-fb, sryap Differential Revision: D48574563
|
This pull request was exported from Phabricator. Differential Revision: D48574563 |
TroyGarden
force-pushed
the
export-D48574563
branch
from
dbd0569 to
44ad142
Compare
|
This pull request has been merged in 500c5fa. |
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.
Summary: # context
Reviewed By: dstaay-fb
Differential Revision: D48574563