commit | 9ee061d2ec366e955f2a348225b4051a31f8f244 | [log] [tgz] |
---|---|---|
author | rohan-tan-bhowmik <46410002+rohan-tan-bhowmik@users.noreply.github.com> | Sat Sep 21 23:07:18 2024 -0700 |
committer | GitHub <noreply@github.com> | Sat Sep 21 23:07:18 2024 -0700 |
tree | 54232463ae3917ee389a23cb09b0d1b0573a093d | |
parent | 891f438c27f201ae0dde3e8408eefc6fca80bc5f [diff] |
[LinalgExt] Masked Attention Implementation (#18525) Enables float/boolean mask as parameters and created linalg generic ops to apply masking. This image (https://imgur.com/a/1MePgcy) elaborates on the main files changed and how they enable masked attention: - Blue boxes represent changed .cpp and .td files to enable/pass/decompose the mask - Yellow boxes represent the different op classes - Red boxes represent test mlir files pertaining to certain .cpp/.td implementations or ops For quick reference, AggregateOpInterfaceImpl.cpp contains the bulk of the actual mask decomposition (QK += mask) And for clarification, TileAttention.cpp only holds the convertToOnlineAttentionOp and getTileAttentionIndexingMaps functions; TilingInterfaceImpl.cpp contains the main tiling capabilities in the form of AttentionOp::getTiledImplementation and OnlineAttentionOp::getTiledImplementation. Updated version of https://github.com/iree-org/iree/pull/18461. This version was created to include scale affine map and enable fused attention (incorporated https://github.com/IanWood1/iree/tree/raikonen/sdpa_mask). - To that end, many modifications in tests are for adding the scale affine map (without much functionality change) - For tiling and decomposition tests, most functionality tests are included in "tiling.mlir" and "decompose_online_attention.mlir". On the other hand, the "tile_attention.mlir and "decompose_attention.mlir" are old paths intended to be be retired and deprecate soon. Hence, no major tests were added it there. Test directory for numerical verification: https://github.com/rohan-tan-bhowmik/iree-masked-attention-test --------- Signed-off-by: Stanley Winata <stanley.winata@amd.com> Co-authored-by: Stanley Winata <stanley.winata@amd.com> Co-authored-by: Ian Wood <ianwood2024@u.northwestern.edu>
IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels
Package | Release status |
---|---|
GitHub release (stable) | |
GitHub release (nightly) | |
Python iree-compiler | |
Python iree-runtime |
Host platform | Build status |
---|---|
Linux | |
macOS | |
Windows |
For the full list of workflows see https://iree.dev/developers/general/github-actions/.
See our website for more information.
Community meeting recordings: IREE YouTube channel
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.