Disable LTC by default until upstream revert relands #1303
Conversation
Tracked with the WIP #1292
Disable LTC in setup.py temporarily until upstream is fixed.
There was a problem hiding this comment.
I assume the issues are due to PyTorch constantly reverting and unreverting that change which changed some function signatures, right?
When @antoniojkim reverts part of his changed (or PyTorch reverts again...), I assume this should be fixed?
Is there anything we can do to improve the stability here? It seems that we kinda just have to deal with this due to usually the latest PyTorch master
|
Oh wait minor comment, perhaps we should have a comment to link back to the issue, in case people are confused about why it's disabled? This way we also know what code to "rollback" when reenabling |
|
We face a similar issue internally (even before the current revert/re-revert pattern), so we pin the PyTorch version in requirements.txt and bump it the most recent PyTorch version when we update Torch-MLIR. Is that (i.e. updating PyTorch on demand) something that we would be open to doing here as well? |
Yes, I think that moving to a model where we update PyTorch on demand would be better here. We already do that for LLVM/MLIR, and now that LTC gives us so much more exposure to PyTorch churn I think it would benefit from the same kind of approach. |
…lvm#1303) Signed-off-by: Whitney Tsang <[email protected]>
The modification to use the builder based interface to generate Krnl loop is completed (llvm#1250, llvm#1283, llvm#1285, llvm#1292, llvm#1293, llvm#1303, llvm#1308, llvm#1314, llvm#1316, llvm#1317, llvm#1326, llvm#1403), and BuildKrnlLoop is no longer needed. Signed-off-by: Whitney Tsang [email protected]
Tracked with the WIP #1292