perf: use packed bit array for attention mask #308
Merged
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
changed the title
yzh119
added a commit
that referenced
this pull request
Jun 20, 2024
🤖 I have created a release *beep* *boop* --- ## [0.1.0](v0.0.4...v0.1.0) (2024-06-20) ### Highlights * Support any GQA group size support for tensor-cores kernels. * Support any page size support for tensor-cores kernels. * Support CUDA-Graph for prefill/decode APIs. * Add an option to accelerate decode kernels with Tensor Cores. * Support custom attention mask. (https://docs.flashinfer.ai/tutorials/kv_layout.html#mask-layout-2d-ragged-tensor) * Support logits cap in Grok-1 models. * Fused GPU-sampling kernels: top-p, top-k, speculative verification. (https://docs.flashinfer.ai/api/python/sampling.html) * PyTorch wrapper of group-gemm cutlass kernels. (https://docs.flashinfer.ai/api/python/sampling.html) ### Acknowledgement We thank [@ibsidorenko](https://github.com/ibsidorenko), [@LiuXiaoxuanPKU](https://github.com/LiuXiaoxuanPKU), [@Yard1](https://github.com/Yard1) [@AgrawalAmey](https://github.com/AgrawalAmey), [@xuzhenqi](https://github.com/xuzhenqi), [@mgerstgrasser](https://github.com/mgerstgrasser), [@esmeetu](https://github.com/esmeetu), [@yz-tang](https://github.com/yz-tang), [@HSQ79815](https://github.com/HSQ79815), [@Qubitium](https://github.com/Qubitium), [@shreygupta2809](https://github.com/shreygupta2809), [@sighingnow](https://github.com/sighingnow), [@vinx13](https://github.com/vinx13), [@tqchen](https://github.com/tqchen), [@merrymercy](https://github.com/merrymercy), [@comaniac](https://github.com/comaniac) and many others for their contributions and helpful discussions for 0.0.5 release. ### Refactor * support any GQA group size for tensor-cores kernels ([#301](#301)) ([c111ca](c111ca6)) * support any page size for tensor-cores kernels ([#306](#306)) ([82fd8c](82fd8c7)) ### Features * add `use_tensor_cores` option to decode kernels to accelerate GQA ([#317](#317)) ([3b50dd5](3b50dd5)) * add group gemm operators ([#282](#282)) ([e08ba42](e08ba42)) * initial support of distributed operators ([#289](#289)) ([03553da](03553da)) * initial support of logits hook ([#298](#298)) ([ab1e2ad](ab1e2ad)) * Separate Q and KV dtypes for decode ([#286](#286)) ([5602659](5602659)) * support cuda graph for batched multi-query(prefill/append) attention ([#275](#275)) ([83ceb67](83ceb67)) * support cuda graph for batched multi-query(prefill/append) attention ([#277](#277)) ([24cc583](24cc583)) * support custom attention mask in prefill/append attention kernels ([#266](#266)) ([7304282](7304282)) * fused speculative sampilng kernels ([#259](#259)) ([cea2bb](cea2bb9)) * expose sampling APIs in pytorch ([#238](#238)) ([092902](0929023)) ### Performance Improvements * initial cuda graph support ([#256](#256)) ([7e9cc7f](7e9cc7f)) * split kv-cache for prefill/append kernels ([#310](#310)) ([f0bb0a3](f0bb0a3)) * use packed bit array for attention mask ([#308](#308)) ([3d43dc9](3d43dc9)) --- 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>
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.
float attention mask consumes too much gpu memory and makes the attention kernel slow.
In this pr we use 0/1 attention mask and uses bit-packed array (1 bit per element, 8 elements are packed together as uint8) to save gpu memory.