[1.8] Include oneDNN gemm fix

[leezu]

Include fix for #19185 (comment)

Thank you @pioy

Fix verified via CI run in #19185: https://jenkins.mxnet-ci.amazon-ml.com/blue/organizations/jenkins/mxnet-validation%2Fcentos-cpu/detail/PR-19185/2/pipeline/

[mxnet-bot]

Hey @leezu , Thanks for submitting the PR
All tests are already queued to run once. If tests fail, you can trigger one or more tests again with the following commands:

• To trigger all jobs: @mxnet-bot run ci [all]
• To trigger specific jobs: @mxnet-bot run ci [job1, job2]

---

CI supported jobs: [centos-cpu, windows-gpu, windows-cpu, sanity, clang, unix-cpu, website, centos-gpu, edge, miscellaneous, unix-gpu]

---

Note:
Only following 3 categories can trigger CI :PR Author, MXNet Committer, Jenkins Admin.
All CI tests must pass before the PR can be merged.

[TaoLv]

Hi @leezu, are you trying to include a unreleased version of oneDNN to MXNet v1.8? Or you just want to verify the fix? We're asking oneDNN team for a v1.6.4 patch release.

[leezu]

Could you provide more details on what fixes are missing from oneapi-src/oneDNN@5ce95ef that will go into 1.6.4?

I think it's up to @samskalicky as release manager to decide if he just includes oneapi-src/oneDNN@5ce95ef or wait for 1.6.4 (or don't include the fix at all)

[samskalicky]

Other than waiting for 1.6.4 patch release, what are our options for avoiding the issue in 1.8? Will reverting any/all of #18822 #18867 #19161 help?

I dont see any issue with setting 3rdparty/mkldnn to a unreleased commit to fix this bug either.

[samskalicky]

Do we have a test now that is passing with this fix?

[leezu]

We don't have a test in 1.x branch. One can construct a test by first calling 'tests/python/unittest/test_contrib_intgemm.py::test_contrib_intgemm_multiply[api1-16-192-4]' followed by a different test calling a hybridized Dense layer and ensuring that the output doesn't contain NaN, as per #19185 (comment)

[samskalicky]

We don't have a test in 1.x branch. One can construct a test by first calling 'tests/python/unittest/test_contrib_intgemm.py::test_contrib_intgemm_multiply[api1-16-192-4]' followed by a different test calling a hybridized Dense layer and ensuring that the output doesn't contain NaN, as per #19185 (comment)

Ok, np. did we verify the fix actually fixes the issue in 1.8.x manually?

[pioy]

Let's see if #19260 results can show that the fix really fixes the issue.

Update: It seems, that the failing tests don't exist on 1.8, so the PR is irrelevant.

[samskalicky]

Thanks @leezu do you plan to open another PR for this fix on v1.x as well?

[leezu]

@mxnet-bot run ci [unix-gpu]

[mxnet-bot]

Jenkins CI successfully triggered : [unix-gpu]

[bartekkuncer]

@leezu The number of tags has changed so for the test to pass you have to change it from 205 to 219 in MemFormat test (103 line in mkldnn_test.cc file).
If that is not a problem I will create PR for v1.x branch.
PR for v1.x branch:
#19276

[samskalicky]

@bartekkuncer @pioy looks like the CI is passing now. Is this PR ready to merge? Can you review/approve it?

[bartekkuncer]

@bartekkuncer @pioy looks like the CI is passing now. Is this PR ready to merge? Can you review/approve it?

@samskalicky Change looks good now. Ready to be merged.

[leezu]

This change is included in v1.x as part of #19276

[samskalicky]

This change is included in v1.x as part of #19276

Thanks @leezu for the tracking!