[Unity][BYOC] Add fused patterns for stacked attention #14608
Merged
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
In some models, the input Q, K and V for attention ops are from a stacked tensor initially, and then they are splitted and reshaped to call attention op, like stacked_qkv -> split -> reshape -> attention. Actually, we could to skip the split and reshape ops, by manipulating the layout parameters in codegen. This PR adds the such fused patterns for stacked attention in BYOC. So that we are able to codegen directly from stacked_qkv.
Collaborator
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.
Generated by tvm-bot |
vinx13
approved these changes
Apr 12, 2023
vinx13
reviewed
Apr 12, 2023
| *make_attention_pattern(with_bias=True), | ||
| ), | ||
| ( | ||
| "cutlass.stacked_attention", |
Member
There was a problem hiding this comment.
does the order of patterns here matter? If we have a subgraph containing both reshape and attention, will cutlass.attention that matches only a single attention operation be selected first?
Contributor
Author
There was a problem hiding this comment.
the order matters here. I tried to change the order with stacked attention first, however, the original attention matches first.
tqchen
pushed a commit
that referenced
this pull request
Apr 13, 2023
* [Unity][BYOC] Add fused patterns for stacked attention In some models, the input Q, K and V for attention ops are from a stacked tensor initially, and then they are splitted and reshaped to call attention op, like stacked_qkv -> split -> reshape -> attention. Actually, we could to skip the split and reshape ops, by manipulating the layout parameters in codegen. This PR adds the such fused patterns for stacked attention in BYOC. So that we are able to codegen directly from stacked_qkv. * fix lint * fix lint
cyx-6
added a commit
to cyx-6/tvm
that referenced
this pull request
Apr 17, 2023
This PR expands the support for fused stacked attention patterns strating with `strided_slice`. Initially, we only support fused stacked attention pattern starting with `split` in apache#14608. But with the help of apache#14583, we may have similar patterns starting with `strided_slice` as well.
masahi
pushed a commit
that referenced
this pull request
Apr 18, 2023
* [Unity][BYOC] Fuse attention pattern with `strided_slice` This PR expands the support for fused stacked attention patterns strating with `strided_slice`. Initially, we only support fused stacked attention pattern starting with `split` in #14608. But with the help of #14583, we may have similar patterns starting with `strided_slice` as well. * remove useless code
cyx-6
added a commit
to cyx-6/tvm
that referenced
this pull request
Apr 18, 2023
This PR is a follow up for apache#14608 and apache#14649. In this PR, we add the checks for the fused stacked attention patterns. So we only enable the fusion of `stacked_qkv` with `ndim=3` and the `split/strided_slice axis=2`.
cyx-6
added a commit
that referenced
this pull request
Apr 19, 2023
* [Unity][BYOC] Add check for stacked attention patterns This PR is a follow up for #14608 and #14649. In this PR, we add the checks for the fused stacked attention patterns. So we only enable the fusion of `stacked_qkv` with `ndim=3` and the `split/strided_slice axis=2`. * check the order of strided_slice
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
In some models, the input Q, K and V for attention ops are from a stacked tensor initially, and then they are splitted and reshaped to call attention op, like
stacked_qkv -> split -> reshape -> attention.
Actually, we could to skip the split and reshape ops, by manipulating the layout parameters in codegen.
This PR adds the such fused patterns for stacked attention in BYOC. So that we are able to codegen directly from stacked_qkv.