Bump Torch-MLIR to c26ca8b

[vivekkhandelwal1]

This commit bumps Torch-MLIR to llvm/torch-mlir@c26ca8b.

[vivekkhandelwal1]

Hi @ScottTodd, I'm not able to understand this CI failure. I don't see any particular error message apart from some dispatch regression related error.

[zjgarvey]

If I understand correctly, this is not a regression, but an improvement? Benchmark time is improved and there is one less dispatch.

[benvanik]

Do we really have an == check? Or is the printing just bad? Improvements shouldn't cause breakages and checks like this should be <=.

ERROR experimental/benchmarks/sdxl/benchmark_sdxl_rocm.py::test_sdxl_rocm_benchmark - check 1563 == 1564: SDXL punet f8 dispatch count should not regress

[zjgarvey]

Do we really have an == check? Or is the printing just bad? Improvements shouldn't cause breakages and checks like this should be <=.

Yeah. It's an equal check:

iree/experimental/benchmarks/sdxl/benchmark_sdxl_rocm.py

Lines 564 to 568 in c80fa3b

	check.equal(
	punet_int8_fp8_dispatch_count,
	goldendispatch_rocm_punet_int8_fp8,
	"SDXL punet f8 dispatch count should not regress",
	)

[benvanik]

We should fix that - a good way to make things less useful is to add noise and alias success and failure - at absolute minimum that should print a giant "CONGRATS YOU MADE THINGS BETTER! THIS IS NOT AN ERROR! WE JUST HAVE A CHANGE DETECTOR" banner :P I don't want change detectors in the repo, though. @saienduri please change the logic to be <= and only error if there's regressions.

[saienduri]

Yeah, I agree. I originally proposed to have <= checks, but was specifically requested to make that an equal check. The error message is misleading though. Sorry about that. Have a PR here #19166 :)

[zjgarvey]

Closing. See #19168