perf: Dense and sparse customizable flashattention-3 template #667
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
force-pushed
the
hopper-rebased
branch
from
8021dbc to
04ee9bc
Compare
|
Finally! Congrats!!! |
|
close #521 |
|
FYI Regarding this PR, due to changes in the build method after this PR, the current nightly build is unavailable. I checked a new branch zhyncs/main, which includes the code up to this commit. It has now been updated to whl through the nightly build. |
yzh119
added a commit
that referenced
this pull request
Dec 17, 2024
This PR resolves an issue in the SWA implementation introduced in #667, where the process would hang under specific conditions.
yzh119
added a commit
that referenced
this pull request
Dec 17, 2024
🤖 I have created a release *beep* *boop* --- ## [0.2.0](v0.1.6...v0.2.0) (2024-12-17) [Release Blog](https://flashinfer.ai/2024/12/16/flashinfer-v02-release.html). ### Features * add `rotary_dim` argument to rope APIs for partial apply rope ([#599](#599)) ([eb9bc71](eb9bc71)) * add a `use_softmax` field in variant class ([#533](#533)) ([d81af97](d81af97)) * add an option `non_blocking` to plan function ([#622](#622)) ([560af6f](560af6f)) * add gemma_rmsnorm and gemma_fused_add_rmsnorm ([#477](#477)) ([1a6b17e](1a6b17e)) * add group size 3 to GQA decode dispatch ([#558](#558)) ([6227562](6227562)) * add JIT compilation support for FA3 templates ([#672](#672)) ([d4e8d79](d4e8d79)) * allow the cascade kernels to be executed using varying sequence lenghts ([#627](#627)) ([92ac440](92ac440)) * CUDAGraph compatibility of multi-level cascade inference APIs ([#586](#586)) ([2332e8a](2332e8a)) * fix the maximal grid dimension in prefill planning with CUDA graphs ([#639](#639)) ([86ca89a](86ca89a)) * improve the precision of the FusedAddRMSNormKernel function ([#587](#587)) ([c7dc921](c7dc921)) * JIT compilation ([#507](#507)) ([3613a5b](3613a5b)) * modify group-gemm stage number ([#497](#497)) ([52dab1d](52dab1d)) * non-contiguous query with paged kv cache ([#553](#553)) ([89f2c4a](89f2c4a)) * pass a dynamic token count to the cascade kernels ([#635](#635)) ([5fe9f7d](5fe9f7d)) * simplify prefill JIT compilation ([#605](#605)) ([fe4f898](fe4f898)) * specify gemm backend ([#648](#648)) ([0cc1a51](0cc1a51)) * support cached cos/sin in rope APIs ([#585](#585)) ([83e541d](83e541d)) * support huggingface transformer style rope interface ([#568](#568)) ([4f40420](4f40420)) * support sm90 cutlass group gemm ([#509](#509)) ([794bdda](794bdda)) * torch custom_op fix for rope ([#569](#569)) ([3e104bc](3e104bc)) * torch custom_op support: norm ([#552](#552)) ([f6e0010](f6e0010)) * torch.compile and custom_op support ([#554](#554)) ([9bf916f](9bf916f)) * warmup for jit kernel tests ([#629](#629)) ([8f5f349](8f5f349)) ### Bug Fixes * AOT compiler flags on non-sm90 ([#522](#522)) ([0aa4726](0aa4726)) * batch decode kernel redundant store output to gmem ([#505](#505)) ([90e42a7](90e42a7)) * compatible with torch 2.2 ([#478](#478)) ([ac41d1b](ac41d1b)) * #452 ([b53a46f](b53a46f)) * remove redundant load ([#495](#495)) ([2de16b0](2de16b0)) * update bmm fp8 test ([#487](#487)) ([45eac04](45eac04)) ### Performance Improvements * accelerate JIT compilation speed ([#618](#618)) ([eaf73fd](eaf73fd)) * Dense and sparse customizable flashattention-3 template ([#667](#667)) ([51236c9](51236c9)) * fix prefill kernel performance degradation (step 1) ([#602](#602)) ([595cf60](595cf60)) * fix the performance issue of `append_paged_kv_cache` ([#588](#588)) ([e15f7c9](e15f7c9)) * improve parallelism in RoPE with pos_ids ([#609](#609)) ([ff05155](ff05155)) * improve plan performance by using non-blocking memcpy ([#547](#547)) ([41ebe6d](41ebe6d)) * reduce the read and write of shared memory in the FusedAddRMSNormKernel ([#592](#592)) ([2043ca2](2043ca2)) * reduce total_num_tiles_q by one ([#644](#644)) ([553ace5](553ace5)) * remove unnecessary contiguous operation in block sparse attention ([#561](#561)) ([7a7ad46](7a7ad46)) * speedup jit compilation of prefill attention kernels ([#632](#632)) ([a059586](a059586)) * use cuda-core implemention for io-bound block-sparse attention ([#560](#560)) ([3fbf028](3fbf028)) --- 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.
This PR adds flashattention-3 template for improving prefill performance on hopper. Block/Vector-sparse support in FlashInfer early version are ported to FA-3 template with CustomStride abstraction in CuTE so that we can support PageAttention with any page size. The programming interface for FA3 template is exactly the same as our previous FA2 template while we add an argument
backendto allow user to select their own backend.Functionalities that are missing in current template include custom mask and we plan to support it using JIT instead of AOT.
H100 Reference performance on variable-length dense and sparse attention kernels(
causal=True) on both FA2 & 3 templates (exposed through BatchPrefillWithRaggedKVCacheWrapper and BatchPrefillWithPagedKVCacheWrapper API correspondingly), for sparse attention workload, we use PageAttention withpage_size=1:CUDA Compiler version: 12.4. FlashInfer's vector sparse (page_size=1) attention implementation can get 90% percent of the dense equivalent, reference benchmark: https://github.com/flashinfer-ai/flashinfer/blob/04ee9bceb5ab0a66c612c1abaee8fa28de2b2349/benchmarks/bench_hopper_attention .
JIT support is left to the next PR because this PR is already heavy. For fp8 support, we will incorporate SageAttention-2 algorithm for numerical stability, and it's left to v0.2.1.
Currently there is some discrepancy in attention variant interface for our FA2 and FA3 template and we will gradually fix the gap.
cc @merrymercy @zhyncs @youkaichao @WoosukKwon @jason-huang03