Adds rotary positional encoding operations #24
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
Implements two rotary embedding variants for transformer attention: - Interleaved format with real/imaginary components alternating - Contiguous format with separate real/imaginary halves Both functions support boundary checking, out-of-bounds clearing, and convert between data types for precise floating-point calculations.
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR introduces two implementations for rotary positional encoding to support transformer attention.
- Implements an interleaved variant, where real/imaginary components alternate.
- Implements a contiguous variant, with separate real/imaginary halves and additional pointer arithmetic for signal extraction.
Comments suppressed due to low confidence (2)
csrc/src/rotary.h:65
- Please replace the informal 'Idk' comment with a clear explanation detailing why the extra copy is required for convert_type to work correctly.
// Idk but I need to copy for the convert_type to work
csrc/src/rotary.h:134
- Please update the comment to explicitly clarify the necessity of copying the tensor after conversion, to improve maintainability and clarity of intent.
// Idk but I need to copy for the convert_type to work
| cute::copy(S(_, m, k), rS(_, m, k)); | ||
| if (get<1>(identity_MN(0, 0, k)) < rotary_dim) { | ||
| const bool is_left = get<1>(identity_MN(0, 0, k)) < rotary_dim / 2; | ||
| Tensor gS_other = make_tensor(S(_, m, k).data() + (is_left ? rotary_dim / 2 : -rotary_dim / 2), S(_, m, k).layout()); |
There was a problem hiding this comment.
The use of negative pointer offset here may raise safety concerns; please add a comment or safeguard to clarify that accessing memory with a negative offset is valid and correctly bounded.
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.
Implements two rotary embedding variants for transformer attention:
Both functions support boundary checking, out-of-bounds clearing, and convert between data types for precise floating-point calculations.