implement paged attention #367
Open
+956
−58
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
There was a problem hiding this comment.
Pull Request Overview
This PR implements a paged attention mechanism by dividing the KV cache into fixed-size pages and adds corresponding wrappers and tests.
- Introduce
paged_attentionCUDA kernel and host wrapper to manage paged KV caches - Update Python bindings in
runtime_kernel_wrapper.cuand expose new API - Add and adjust Python unit tests for paged attention and decoding with normalization
Reviewed Changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| tests/runtime_python/test_paged_attention.py | New test to compare paged_attention kernel against a Torch ref |
| tests/runtime_python/test_decoding_w_norm.py | Updated decoding test to include RMSNorm and rotary arguments |
| tests/runtime_python/runtime_kernel_wrapper.cu | Added paged_attention & updated single_batch_decoding wrappers |
| include/mirage/persistent_kernel/tasks/paged_attention.cuh | Full implementation of the paged attention device task |
| include/mirage/persistent_kernel/tasks/kernel.h | Exposed the new paged_attention task in the kernel header |
Comments suppressed due to low confidence (3)
tests/runtime_python/test_paged_attention.py:162
- [nitpick] This test only prints random ratios without any assertions. Consider adding deterministic checks (e.g.,
assert torch.allclose(mirage_output, torch_out, atol=1e-2)) to automatically validate correctness.
if torch.rand(1).item() < 0.05:
tests/runtime_python/test_decoding_w_norm.py:141
- [nitpick] This prints the output ratio but lacks assertions. Adding
assertstatements to comparemirage_outputandtorch_outputwithin tolerance will make the test fail on regressions.
print(torch_output / mirage_output)
tests/runtime_python/runtime_kernel_wrapper.cu:116
- [nitpick] The parameter name
rotary_emdappears to be a typo. Consider renaming it torotary_emborrotary_embedfor consistency and clarity.
bool rotary_emd,
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.
**Description of change
Divide the KV Cache into pages to enable more fine-grained and efficient memory management
Related Issues:
Linked Issues: