feat: support non-contiguous (packed) input for prefill kernels #404
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
yzh119
added a commit
that referenced
this pull request
Jul 29, 2024
🤖 I have created a release *beep* *boop* --- ## [0.1.2](v0.1.1...v0.1.2) (2024-07-29) ### Bugfix * Fix the sampling kernel bug for cu118 ([#386](#386), [#387](#387)) ([0cd499](0cd4994), [dc3f18](dc3f184)) ### Features * add llama 3.1 style rope ([#401](#401)) ([4c89dec](4c89dec)) * non-inplace rope operators ([#405](#405)) ([74ffba1](74ffba1)) * sliding window attention ([#406](#406)) ([28cffd3](28cffd3)) * support non-contiguous (packed) input for prefill kernels ([#404](#404)) ([68c3719](68c3719)) ### Performance Improvements * slight optimization on merge states ([#313](#313)) ([701c813](701c813)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please). --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zihao Ye <expye@outlook.com>
yzh119
pushed a commit
that referenced
this pull request
Oct 25, 2024
## Motivation Previously, only ragged version of prefill kernel supported non-contiguous query tensor (#404). But with paged kv cache, you have to make query tensor contiguous. Libraries like vLLM or SGLang must make query tensor contiguous before calling flashinfer kernels ([vLLM call of flashinfer](https://github.com/vllm-project/vllm/blob/b7df53cd42f3eab007b4f287c151960858e949df/vllm/attention/backends/flashinfer.py#L839), [SGLang call of flashinfer](https://github.com/sgl-project/sglang/blob/87a7cfa080cec3f123618c1429b5f998bf5d99cb/python/sglang/srt/layers/attention/flashinfer_backend.py#L236)). This PR solves it, ensuring that prefill/decode kernels with paged kv cache support non-contiguous query tensor. ## Main Changes 1. Add strides of query tensor in `BatchPrefillPagedParams` and `BatchDecodeParams`. 2. Set stride parameters before calling those kernels. 3. Modify JIT compiling templates to support new kernel parameters. 4. Add some tests. The Python interfaces remain the same. Nothing changes except it accepts non-contiguous query tensors now! --------- Signed-off-by: LinHeLurking <LinHe.Lurking@gmail.com>
# for free
to join this conversation on GitHub.
Already have an account?
# 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.
This PR implements #311 , after this PR, we support packed qkv input without explictly convert make the input contiguous:
Before this PR, we need to make
q/k/vcontiguous before we launch the attention kernel, which incurs some overhead.I observe slight (<1%) performance degration after this PR for non-packed input, which is acceptable IMO.