| commit | 265d41fe07f8f0523d128ad162263825b1b72e9d | [log] [tgz] |
|---|---|---|
| author | Stella Laurenzo <stellaraccident@gmail.com> | Tue Sep 05 21:17:24 2023 -0700 |
| committer | GitHub <noreply@github.com> | Tue Sep 05 21:17:24 2023 -0700 |
| tree | d3cdd7fbfe62b05243d73d3cf0b23c7ed5bbb05f | |
| parent | 13ef677e556d0a1d154e45b052fe016256057f65 [diff] |
Replace tm_tensor with torch-mlir proper (#14917) Per discussion on iree-discuss, this patch removes the original tm_tensor source snapshots and special casing from the codebase, instead pulling in the parts of torch-mlir that provide a full solution, starting from the torch dialect down. This reuses the existing `IREE_INPUT_TORCH` CMake define but routes it to a compiler plugin at `compiler/plugins/input/Torch` where all build plumbing to integrate and local pipelines exist. I have kept the original tm_tensor input type for the moment since downstreams are using that, but we will want to eventually replace that with a dedicated `torch` input type which works directly from the `torch` dialect. This patch should be NFC to all current users, however it allows programmatic and API access to the full Torch dialects and conversion pipelines, which is essential for broader scale integration. We add `third_party/torch-mlir` as a submodule at torch-mlir's current HEAD. During LLVM integrates, if there are ever API breaks that affect this, we may need to push patches to torch-mlir to address. This can be done on a torch-mlir branch as needed (or committed at head). I have been tracking changes for a couple of months and have not had a need to do this yet, which indicates that this is relatively stable. Committing with an XFAIL'd lit test, which for some reason was not running previously and now is: https://github.com/openxla/iree/issues/14916
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!
See our website for more information.
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.