Integrate LLVM at 41f65666f6378bba7266be7c662c70074f04ed75 #4358
Conversation
hanhanW left a comment
hanhanW left a comment There was a problem hiding this comment.
The change looks good to me. I checked how upstream is doing and they match
| @Yu-Zhewen thanks for this change. You probably need to integrate changes in #4311 by @justin-ngo-arm |
Still carrying two reverts as #22366. New cherry-picks: - [MLIR] Revamp RegionBranchOpInterface (llvm/llvm-project@41f6566) - Corresponding torch-mlir fix (llvm/torch-mlir#4358) Also, IREE fixes on `RegionBranchPoint` and `RegionSuccessor`. ci-extra: test_torch, windows_x64_msvc Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
Thanks, @sahas3. We also need to address the deprecation from llvm/llvm-project#164649. I already have a fix, but it’s around 10k lines of changes (iree-org@553f323). Would you be okay if I revert llvm/llvm-project#164649 in this PR and then submit a separate PR for the larger fix? |
@Yu-Zhewen maybe we leave the patch here and ask the torch-mlir integrator to pick this up. We already cherry-pick the change in IREE and it works well. The last bump is 2 months ago, and I think it not IREE build gardener's responsibility to maintain torch-mlir llvm state. (cc @MaheshRavishankar ) |
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
…o enums Context: TOSA's rounding mode, nan propagation, and resize mode have been updated from string attributes to enums. This commit updates the Torch to TOSA path to align with those changes. Signed-off-by: Justin Ngo <justin.ngo@arm.com>
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
Yu-Zhewen force-pushed the users/Yu-Zhewen/integrate-llvm-41f6566 branch from 1da0692 to 9b579d8 Compare 
Carries cherry-pick iree-org/torch-mlir@42dff72 (WIP PR llvm/torch-mlir#4358) Fixes ONNX failure related to i1 tensors. --------- Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
This is a follow-up to #4358. Remove the revert by fixing all deprecated usages of `OpBuilder::create`, as requested in llvm/llvm-project@c0b42ec --------- Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>
| Hi @Yu-Zhewen, the windows CI has been failing since this commit https://github.com/llvm/torch-mlir/actions/workflows/ci_windows.yml. Can you please take a look? Thanks! |
| Zhewen and I do not work on torch-mlir project, FYI. I can try to fix it tomorrow, so take your time @Yu-Zhewen It's be good if @sjarus team can pick this up. If it becomes tricky, we can revert it and land it later. btw, is there a way to trigger windows CI without pushing it to the main repo and run the workflow manually? We may not have windows machine to reproduce the issue. |
| Trying to fix it in #4373 |
That's good to know. Thanks so much for looking into this. |
| It does not look like a trivial fix. I tried at least 10 fixes today, and all of them failed: https://github.com/llvm/torch-mlir/actions/workflows/ci_windows.yml Do we go with revert? IREE project is okay, since it already cherry-picks the patch. I'll leave this to who works on torch-mlir LLVM bump. |
| Thanks @hanhanW for looking into this. I think this might be because of the ninja issue mentioned in ninja-build/ninja#2616. I've started a CI pinning back ninja to be older than 1.13.0 in https://github.com/sahas3/torch-mlir/actions/runs/19347185649/job/55350440958. Let's see if that passes. |
5 participants
Bump LLVM to llvm/llvm-project@41f6566
Bump StableHLO to openxla/stablehlo@4c0d484
Revert llvm/llvm-project#164649, will address in a follow-up PR.
Fix RegionBranchPoint/RegionSuccessor as llvm/llvm-project#165429
Fix populateExpansionPatterns as llvm/llvm-project#151153
Integrate #4311