perf: use persistent kernel for merging attention states #459
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
Aug 27, 2024
🤖 I have created a release *beep* *boop* --- ## [0.1.6](v0.1.5...v0.1.6) (2024-08-27) ### SM75 Support Starting from [0.1.6](v0.1.5...v0.1.6), our pre-built wheels include experimental support sm75 (Turing architecture GPUs such as Tesla T4, Quadro RTX 6000 and RTX 2080). ### API Changes #### `plan`/`run` Since [0.1.6](v0.1.5...v0.1.6) on, `begin_forward`/`forward`/`end_forward` APIs are replaced with the new `plan`/`run` API. - `forward` is renamed to `run`, which is more precise and consistent with the naming convention of cutlass's python API. - `begin_forward` is renamed to `plan`, which is consistent with the naming convention of nvmath API. - `end_forward` is deprecated and has no effect after this PR. There is some slight difference between the old `forward` and the new `run` API: - All extra arguments such as `causal` and `logits_soft_cap` will be provided in `plan` (previously `begin_forward`) API, and cached until next `plan` call, and we only need to provide query and KV-Cache tensors in `run` API. The old `begin_forward`/`forward`/`end_forward` APIs are still functional, but we will gradually deprecate them in future releases. Check [#466](#466) for more details. #### `MultiLevelCascadeAttentionWrapper` Since [0.1.6](v0.1.5...v0.1.6) on, we introduce a new `MultiLevelCascadeAttentionWrapper` API for cascade inference, which supports multi-level cascade inference where all levels' KV-Cache can be managed in a unified Paged KV-Cache. See [documentation](https://docs.flashinfer.ai/api/python/cascade.html#flashinfer.cascade.MultiLevelCascadeAttentionWrapper) and [tutorial](https://docs.flashinfer.ai/tutorials/kv_layout.html#multi-level-cascade-inference-data-layout) on API usage and layout explaination. The old `BatchDecodeWithSharedPrefixPagedKVCacheWrapper` and `BatchPrefillWithSharedPrefixPagedKVCacheWrapper` will be deprecated in future releases. ### Features * sm75 support ([#448](#448), [#449](#449)) * add `MultiLevelCascadeAttentionWrapper` API ([#462](#462)) ([1e37989](1e37989)) * add accept num, emit num metric for ChainSpeculativeSampling ([#450](#450)) ([fa38b5e](fa38b5e)) * support bmm fp8 ([#469](#469)) ([f1c0b68](f1c0b68)) ### Refactor * refactor: replace `begin_forward`/`forward`/`end_forward` with `plan`/`run` [#466](#466) ### Misc * misc: improve error handling of sampling kernels ([#456](#456)) ([0dce178](0dce178)) ### Performance Improvements * slight optimization on f16->f8 fragment layout swizzling ([#453](#453)) ([0d61871](0d61871)) * slight optimization on fragment layout swizzle ([#458](#458)) ([7c397cb](7c397cb)) * use persistent kernel for merging attention states ([#459](#459)) ([be6bf5b](be6bf5b)) ### Acknowledgement We thank [@LiuXiaoxuanPKU](https://github.com/LiuXiaoxuanPKU) on enhance of speculative sampling operator, [@merrymercy](https://github.com/merrymercy) on API change suggestion and [@zhyncs](https://github.com/zhyncs) on integrating fp8 BMM cublas implementation. --- 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>
# 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.
As observed by @MasterJH5574 , there are cases where our
VariableLengthMergeStatesKernellaunches a lot of CTAs (>=10k) while most of the CTAs only work on small number of merges, this PR fixes the issue by using a persistent kernel.There is still load imbalance issue, and I plan to resolve it inside scheduler. I'll leave it for later PRs.