Enzyme reverse mode rules and test for index manipulations

[kshyatt]

Also consolidated the pullbacks for Enzyme and Mooncake

[codecov]

Codecov Report

❌ Patch coverage is 94.69697% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
ext/TensorKitEnzymeExt/indexmanipulations.jl	93.57%	7 Missing ⚠️

Files with missing lines	Coverage Δ
ext/TensorKitEnzymeExt/TensorKitEnzymeExt.jl	100.00% <ø> (ø)
ext/TensorKitMooncakeExt/indexmanipulations.jl	95.75% <100.00%> (-0.36%)	⬇️
src/TensorKit.jl	13.79% <ø> (ø)
src/pullbacks/indexmanipulations.jl	100.00% <100.00%> (ø)
ext/TensorKitEnzymeExt/indexmanipulations.jl	93.57% <93.57%> (ø)

... and 13 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[kshyatt]

The permute fail I really cannot reproduce locally on Ubuntu and 1.10 at all... how frustrating!

[Jutho]

General question: separating all these tests over different folders now causes a large number of CI jobs, some of which only take in the order of one minute. If I open one of them, the actual tests took 0.7 seconds, so that is one minute of overhead I guess to launch the CI job and set up the environment etc. Is that still efficient?

It seems useful to bundle CI jobs such that the actual launching overhead is perhaps in the order of 5 - 10 %, so jobs of about 10 - 20 minutes? Or am I overlooking something?

[kshyatt]

I think now that we fixed the "reinstalling and precompiling all the packages" issue by properly using the caches, these can indeed be consolidated (previously they were each taking 30mins+). I'll do that today.

[Jutho]

I think now that we fixed the "reinstalling and precompiling all the packages" issue by properly using the caches, these can indeed be consolidated (previously they were each taking 30mins+). I'll do that today.

But I guess that now the overhead is indeed "only 1 minute"; before that I assume it would have been much worse even to split up all these jobs.

[Jutho]

This looks good, but I am still surprised that we even need to define rules for flip and insert/remove unit, which are not difficult but also not entirely straightforward with the copy kwarg etc. I would think this is exactly where Mooncake and Enzyme would have made our life easier.

[kshyatt]

I think Enzyme must have made some internal fixes too. Previously they had to be split up because even after the package compilation was done, the tests themselves took 45+ mins each which exceeded the timeout.

[kshyatt]

This looks good, but I am still surprised that we even need to define rules for flip and insert/remove unit, which are not difficult but also not entirely straightforward with the copy kwarg etc. I would think this is exactly where Mooncake and Enzyme would have made our life easier.

I agree... I would like to try pruning these down after this goes in and seeing what minimal rule set we can get away with.

[kshyatt]

As discussed on Zulip, I'm going to merge this once CI passes and then take another run at getting Enzyme and Mooncake to generate good rules themselves.

[kshyatt]

AMD fail can't be related, going to merge!