Support dim != 1 for softmax w/o using permute #845
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
facebook-github-bot
added
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D47732875 |
|
This pull request was exported from Phabricator. Differential Revision: D47732875 |
int3
added a commit
to int3/AITemplate
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: facebookincubator#845 This is a port of PyTorch's softmax implementation. Notable differences: * We use fast_exp & fast_max instead of std::max and std::exp * We don't use higher-precision types for accumulator values (doesn't look like the dim=-1 softmax code does this either) * We propagate the reduction dim size & inner size as constants We seem to be very marginally slower than PT for small batch sizes and very marginally faster for large ones. I have named this new softmax implementation "softmaxGeneral" since it is able to handle arbitrary reduction dimensions, even though we are only using it for the `dim > 1` case. Differential Revision: D47732875 fbshipit-source-id: 5118fed5cf6457bd9d27f553245c3b9985403f78
|
This pull request was exported from Phabricator. Differential Revision: D47732875 |
int3
added a commit
to int3/AITemplate
that referenced
this pull request
Jul 28, 2023
Summary: Pull Request resolved: facebookincubator#845 This is a port of PyTorch's softmax implementation. Notable differences: * We use fast_exp & fast_max instead of std::max and std::exp * We don't use higher-precision types for accumulator values (doesn't look like the dim=-1 softmax code does this either) * We propagate the reduction dim size & inner size as constants We seem to be very marginally slower than PT for small batch sizes and very marginally faster for large ones. I have named this new softmax implementation "softmaxGeneral" since it is able to handle arbitrary reduction dimensions, even though we are only using it for the `dim > 1` case. Differential Revision: D47732875 fbshipit-source-id: 1bd5a47e4293e3b800942b3d0cd0270c657d0069
|
This pull request was exported from Phabricator. Differential Revision: D47732875 |
Summary: Pull Request resolved: facebookincubator#845 This is a port of PyTorch's softmax implementation. Notable differences: * We use fast_exp & fast_max instead of std::max and std::exp * We don't use higher-precision types for accumulator values (doesn't look like the dim=-1 softmax code does this either) * We propagate the reduction dim size & inner size as constants We seem to be very marginally slower than PT for small batch sizes and very marginally faster for large ones. I have named this new softmax implementation "softmaxGeneral" since it is able to handle arbitrary reduction dimensions, even though we are only using it for the `dim > 1` case. Differential Revision: D47732875 fbshipit-source-id: 5323b5cfb0b2d3e3b983033718563b2976df5519
|
This pull request was exported from Phabricator. Differential Revision: D47732875 |
|
This pull request has been merged in 318111f. |
int3
added a commit
to int3/AITemplate
that referenced
this pull request
Aug 10, 2023
Summary: Now that facebookincubator#845 has landed, the backend supports softmax with `dim != -1` directly, and the fx converter no longer needs the workaround from facebookincubator#395. Differential Revision: D48248330 fbshipit-source-id: fad534f63b642ecbf79a90f7fae3c4cc9ad4dadf
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:
This is a port of PyTorch's softmax implementation.
Notable differences:
This is probably the reason why we are (very marginally) faster.
I have named this new softmax implementation "softmaxGeneral" since it is able to handle arbitrary reduction dimensions, even though we are only using it for the
dim > 1case.Differential Revision: D47732875