Oct30 uplift third_party/tt-metal effort

[nsmithtt]

No description provided.

[kmabeeTT]

Plz cherry-pick a2e64a2 for some needed runtime changes to solve compile error that goes with metal PR tenstorrent/tt-metal#14394

[mtopalovicTT]

@nsmithtt
Not sure if possible to include this also?
tenstorrent/tt-metal@948fafb

[mbezuljTT]

we have a bigger problem with merges, when we solve it - these will be automated again.

[kmabeeTT]

Comment on CI fails, one error is:

2024-11-01 16:16:18,806 - DEBUG - starting loop=1/1 for binary=test_perf_slice.mlir.ttnn
                 Always | FATAL    | Physical size (32, 16) must be multiple of page size (1, 32)
2024-11-01 16:16:19,059 - ERROR - ERROR: test=test_perf_slice.mlir.ttnn experienced an error with exception=TT_FATAL @ /localdev/kmabee/mlir2/third_party/tt-metal/src/tt-metal/ttnn/cpp/ttnn/tensor/layout/tensor_layout.cpp:97: width_remainder == 0 && height_remainder == 0

Talked to Artem on Friday and he said tt-metal bug and fix is coming ...soon...

Edit: Unconfirmed if we need to fix something here, but metal PR Artem referring to is tenstorrent/tt-metal#14550

[nsmithtt]

We'll need this UMD fix: tenstorrent/tt-umd#251 (comment)

[mbezuljTT]

We'll need this UMD fix: tenstorrent/tt-umd#251 (comment)

@broskoTT we're fixing metal in mlir repo. do you know if umd uplifts regularly to metal repo or would be better for us to do it on our own?

[kmabeeTT]

Small update. Renamed PR, and cherry picked 3 metal PR's (2 from Friday, 1 from today) and the failing tests in last week's CI run we saw are gone. Rebased and force pushed this branch as we are getting close. Going to try to test on Forge-FE too.

I haven't seen need for UMD fix referenced above though, or at least I didn't do anything, I get impression it's needed only for newer tt-metal not the Oct30 base we are still targeting with this uplift. tt-mlir is currently using Oct25 tt-metal.