Add reduce_min and reduce_max #942
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
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D49978036 |
aakhundov
added a commit
to aakhundov/AITemplate-1
that referenced
this pull request
Oct 5, 2023
Summary: TBA Differential Revision: D49978036
aakhundov
force-pushed
the
export-D49978036
branch
from
8945f2b to
ce3f076
Compare
|
This pull request was exported from Phabricator. Differential Revision: D49978036 |
aakhundov
force-pushed
the
export-D49978036
branch
from
ce3f076 to
840a9aa
Compare
aakhundov
added a commit
to aakhundov/AITemplate-1
that referenced
this pull request
Oct 6, 2023
Summary: TBA Differential Revision: D49978036
|
This pull request was exported from Phabricator. Differential Revision: D49978036 |
aakhundov
added a commit
to aakhundov/AITemplate-1
that referenced
this pull request
Oct 6, 2023
Summary: My understanding is that the max / min reduction are usually not used in the trained PT models due to their non-differentiability. Probably, that's the reason why we don't have them reflected as AIT ops. However, they may be useful in inference for a manually composed AIT model (or PT model translated by fx2ait) which was solicited in the issue [facebookincubator#941](facebookincubator#941). In this diff, we're adding the new `reduce_max` and `reduce_min` ops to AIT. The new ops are using the existing common back-end implementation used by `reduce_sum` with a small extension: the default (initial) accumulator value of 0 is not suitable for the min / max reductions, it is now made configurable in the codegen; `lowest()` and `max()` numeric limits for the ElementCompute are set by the `reduce_max` / `reduce_min` ops' codegen, respectively. fx2ait wiring of the new ops to the `torch.amax` and `torch.amin` has also been added. Differential Revision: D49978036
aakhundov
force-pushed
the
export-D49978036
branch
from
840a9aa to
793bc04
Compare
|
This pull request was exported from Phabricator. Differential Revision: D49978036 |
Summary: My understanding is that the max / min reduction are usually not used in the trained PT models due to their non-differentiability. Probably, that's the reason why we don't have them reflected as AIT ops. However, they may be useful in inference for a manually composed AIT model (or PT model translated by fx2ait) which was solicited in the issue [facebookincubator#941](facebookincubator#941). In this diff, we're adding the new `reduce_max` and `reduce_min` ops to AIT. The new ops are using the existing common back-end implementation used by `reduce_sum` with a small extension: the default (initial) accumulator value of 0 is not suitable for the min / max reductions, it is now made configurable in the codegen; `lowest()` and `max()` numeric limits for the ElementCompute are set by the `reduce_max` / `reduce_min` ops' codegen, respectively. fx2ait wiring of the new ops to the `torch.amax` and `torch.amin` has also been added. Differential Revision: D49978036
aakhundov
force-pushed
the
export-D49978036
branch
from
793bc04 to
fc62ca0
Compare
|
This pull request was exported from Phabricator. Differential Revision: D49978036 |
|
This pull request has been merged in 0a67de1. |
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.
My understanding is that the max / min reduction are usually not used in the trained PT models due to their non-differentiability. Probably, that's the reason why we don't have them reflected as AIT ops. However, they may be useful in inference for a manually composed AIT model (or PT model translated by fx2ait) which was solicited in #941.
In this diff, we're adding the new
reduce_maxandreduce_minops to AIT. The new ops are using the existing common back-end implementation used byreduce_sumwith a small extension: the default (initial) accumulator value of 0 is not suitable for the min / max reductions, it is now made configurable in the codegen;lowest()andmax()numeric limits for the ElementCompute are set by thereduce_max/reduce_minops' codegen, respectively.fx2ait wiring of the new ops to the
torch.amaxandtorch.aminhas also been added.Differential Revision: D49978036