Skip to content

Conversation

xurui1995
Copy link
Contributor

Copied from tpp-mlir project
The current MLIR repo does not offer an easy-to-use pass for decomposing linalg.softmax, and tpp-mlir provide a solution
track: #214

@ZhennanQin
Copy link
Contributor

We're waiting the upstream solution from: llvm/llvm-project#97582. @kurapov-peter would you please provide some insights?

@xurui1995
Copy link
Contributor Author

xurui1995 commented Aug 6, 2024

We're waiting the upstream solution from: llvm/llvm-project#97582. @kurapov-peter would you please provide some insights?

Oh, I did not notice this PR on llvm, @WangJialei-A found this issue when writing linalg single op validation, I think if we can get the upstream PR merged before our validation, then we can close this PR and use upstream.

@kurapov-peter
Copy link
Contributor

Unfortunately it won't land in the nearest future. The decision is to create a generic decomposition mechanism (that I'm working on now). I could put a similar implementation into GC if we need it. Do we?

@xurui1995 xurui1995 force-pushed the xurui/decompose_softmax branch from 3c72835 to c639459 Compare August 6, 2024 13:22
@xurui1995 xurui1995 linked an issue Aug 12, 2024 that may be closed by this pull request
@ZhennanQin
Copy link
Contributor

Please rebase.

@xurui1995 xurui1995 force-pushed the xurui/decompose_softmax branch from 71716fd to ed004d8 Compare August 23, 2024 07:22
@kurapov-peter kurapov-peter merged commit c1a8454 into intel:main Aug 23, 2024
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Lowering linalg softmax op failure
3 participants