| commit | 8dd1db34174c603e6a471dd7523ead112f721245 | [log] [tgz] |
|---|---|---|
| author | Ian Wood <75152913+IanWood1@users.noreply.github.com> | Wed Aug 21 09:24:13 2024 -0700 |
| committer | GitHub <noreply@github.com> | Wed Aug 21 09:24:13 2024 -0700 |
| tree | 9e7aeddfd4d3b2edfab8bfcd04e4b7d3715a0f06 | |
| parent | 6a92fb7cbf00a23796b2f1b5fc52c3382206c132 [diff] |
Bubble expand shapes through `AttentionOp`s (#18074) The idea was to incrementally land these changes (see https://github.com/iree-org/iree/pull/18030 and https://github.com/iree-org/iree/pull/18017), but it appears that they are dependent on each other to not cause regressions in sdxl fp16. I will give a summary of what the commits here do. ## 1. Make all input maps identity (originally on PR https://github.com/iree-org/iree/pull/18017) First two commits > Change ElementwiseOpInterchangePattern from making output maps identity to making the input maps identity. https://github.com/iree-org/iree/issues/18006 ## 2. Add `LinalgFusionOpInterface` to attention op (originally on PR https://github.com/iree-org/iree/pull/18030) Next two commits after merge >- Added interface to attention op so that producers/consumers can get fused into the same dispatch. >- Cleaned up interface method naming by changing `getIndexingMaps` to `getIndexingMapsArray` to better match linalg. `getIndexingMaps` also already conflicted with a method that attention already has. >- Added indexing check to attention verifier and corresponding test. ## 3. Most recent commit on this PR This was going to be the only commit of this PR and is the one that needs the most review. It creates LinalgExt::populateFoldReshapeOpsByExpansionPattern that emulates linalg::populateFoldReshapeOpsByExpansionPattern but for AttentionOp. A bit more info here https://github.com/iree-org/iree/issues/17673. --------- Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu> Signed-off-by: saienduri <saimanas.enduri@amd.com> Signed-off-by: MaheshRavishankar <mahesh.ravishankar@gmail.com> Co-authored-by: MaheshRavishankar <mahesh.ravishankar@gmail.com> Co-authored-by: saienduri <saimanas.enduri@amd.com>
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.