Optimize MX4 padding to minimize need for tuning #3040
Closed
Conversation
This file contains hidden or 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: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
c8b08e2 to
3381c55
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
1 similar comment
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
force-pushed
the
export-D61816830
branch
from
3381c55 to
d1002ce
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. Differential Revision: D61816830
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
d1002ce to
1a2251d
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
1a2251d to
a3926f7
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
1 similar comment
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
a3926f7 to
acd2a7c
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
acd2a7c to
6b4d3a7
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
6b4d3a7 to
5412811
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
5412811 to
ee72991
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 26, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
dde4c7f to
bec909d
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 27, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
bec909d to
de07d2f
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
force-pushed
the
export-D61816830
branch
from
de07d2f to
8de5079
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 28, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 28, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
8de5079 to
60e8578
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 28, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
60e8578 to
a9e8892
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 28, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
a9e8892 to
cc93627
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
force-pushed
the
export-D61816830
branch
from
cc93627 to
9dfc24b
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 29, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Aug 29, 2024
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
jwfromm
force-pushed
the
export-D61816830
branch
from
9dfc24b to
d6bed7e
Compare
Summary: X-link: facebookresearch/FBGEMM#137 Pull Request resolved: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Differential Revision: D61816830
|
This pull request was exported from Phabricator. Differential Revision: D61816830 |
jwfromm
force-pushed
the
export-D61816830
branch
from
d6bed7e to
f4f7779
Compare
|
This pull request has been merged in c818b87. |
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Apr 10, 2025
Summary: Pull Request resolved: facebookresearch/FBGEMM#137 X-link: pytorch#3040 D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes. After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding. With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance. After further experimentation, we can actually remove tuning entirely and just use a reasonably large `GROUP_LOAD`. This gives good performance across all shapes and removes any chance of overhead. Empirically, `GROUP_LOAD=64` seems to be the sweet spot. Reviewed By: jianyuh Differential Revision: D61816830 fbshipit-source-id: d35460ae39d374a699f57e43b2ee5b9675f00d69
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:
D61447274 introduced a very cool way of doing 2D indexing over input tensors during MX4 quantization, however, it is fairly reliant on tuning configurations to get good performance. It turns out the use case for MX4 has highly dynamic shapes, so we spend a huge amount of time tuning those shapes.
After deep meditation I realized there's a much simpler indexing scheme we can use, which is similar to the 1D accesses we used previously but adds shifts for padding.
With this approach we should get the best of both worlds; support for padding rows not divisible by group size and minimizing tuning while maintaining good performance.
Differential Revision: D61816830