-
Notifications
You must be signed in to change notification settings - Fork 2.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[CPU] Fixed accuracy of compressed matmul with scalar scale #28431
[CPU] Fixed accuracy of compressed matmul with scalar scale #28431
Conversation
@@ -74,6 +75,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_MatMulCompressedWeights_corner_cases, | |||
::testing::ValuesIn(decompression_precisions_corner_cases), | |||
::testing::Values(ov::element::undefined), | |||
::testing::ValuesIn(transpose_weights), | |||
::testing::Values(DecompressionType::full), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Maybe we can add DecompressionType::scalar
multiply instances for ARM? If ARM has the same issue as X64, the failed tests will be a good indication: we could skip them for now with the corresponding comment (e.g. with ticket number)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We are using oneDNN ref mamtul to cover compression on ARM now. So the fix doesn't affect ARM behavior.
However the idea the cover some corner cases for ARM sounds reasonable. Will follow-up in separate PR.
...ntel_cpu/tests/functional/custom/subgraph_tests/src/classes/matmul_weights_decompression.cpp
Show resolved
Hide resolved
...ns/intel_cpu/tests/functional/custom/subgraph_tests/src/x64/matmul_weights_decompression.cpp
Show resolved
Hide resolved
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
OK for GPU change
…toolkit#28431) ### Details: - This PR fixes accuracy of compressed Matmuls with scalar scale - The issue was on micro-kernel level: incorrect offset during scales load oneDNN PR: openvinotoolkit/oneDNN#270 ### Tickets: - [CVS-143420](https://jira.devtools.intel.com/browse/CVS-143420)
Details:
oneDNN PR: openvinotoolkit/oneDNN#270
Tickets: