[Stablehlo] add refbackend #2712
Conversation
qingyunqu
force-pushed
the
add-stablehlo-refbackend
branch
from
628699e
to
4a0c621
Compare
qingyunqu
force-pushed
the
add-stablehlo-refbackend
branch
from
94a6953
to
152eb58
Compare
There was a problem hiding this comment.
I'm fine with giving this a try, but the stablehlo ref backend significantly increases the surface area of the stablehlo dep, and keeping it up to date has already been a challenge.
I think this pushes us to a policy change: if when integrating a new version of llvm, if a corresponding stablehlo version doesn't work, we will land with stablehlo disabled and fix it in post commit. If this becomes a pattern, we can factor the CI a bit differently to test these other configurations in a non gating way.
I think that removing the requirement that every commit is available in lockstep will give us all what we want without needing to engage in integrate heroics.
I agree with you. I think it's ok to disable stablehlo temporarily. |
Thanks, I forgot these config. |
|
|
||
|
|
||
| class LinalgOnTensorsStablehloBackend(StablehloBackend): | ||
| """Main entry-point for the linalg-on-tensors based TOSA backend. |
There was a problem hiding this comment.
Why is this file mentioning TOSA?
There was a problem hiding this comment.
It's a typo, I will fix it.
| if(TORCH_MLIR_ENABLE_STABLEHLO) | ||
| list(APPEND LinkedLibs StablehloPasses StablehloLinalgTransforms) | ||
| endif() |
There was a problem hiding this comment.
I'm having trouble compiling some part of this after syncing past this commit (f3e8199). My build is fine before this commit (8e2e5ee) and after this commit with -DTORCH_MLIR_ENABLE_STABLEHLO=OFF.
Error logs: https://gist.github.com/ScottTodd/332e4e76b744d0bfdd75c68cc8191970#file-torch_mlir_build-txt-L1784
System: Windows 10, MSVC (Visual Studio Build Tools 2022 Preview - amd64)
Do StableHLO and torch-mlir have Windows CI builds? I don't see issues in IREE's downstream Windows CI or locally, so either that is behind this version or something else is different about how the projects are configured.
There was a problem hiding this comment.
Torch-MLIR doesn't have Windows CI builds now. I don't have an idea to fix this compiling error.
There was a problem hiding this comment.
It seems the same error as openxla/stablehlo#1772
No description provided.