Make @adjoint unthunk pullback inputs

[oschulz]

Required to support FluxML/Zygote.jl#966

[oschulz]

See FluxML/Zygote.jl#966 for discussion.

[oschulz]

Note: the keepthunks = false argument is intended for Zygote-internal use only.

[oschulz]

@oxinabox and @DhairyaLGandhi - if you have a moment, could you check if we can move forward with this in the current form? We'll need a release with this before FluxML/Zygote.jl#966 can go through full CI. I would suggest that this can be released as as patch release, since it's non-breaking and doesn't introduce any new public features.

[oxinabox]

I approve this.
But I won't merge til @DhairyaLGandhi says the plan in the other PR is going ahead

[oschulz]

Sure, thanks!

[oschulz]

@DhairyaLGandhi will you give us your blessing? :-)

[CarloLucibello]

bump

[oschulz]

@DhairyaLGandhi , @oxinabox , now that ChainRulesCore v1.0 is out, can we proceed with this, so that we can merge FluxML/Zygote.jl#966 ?

[DhairyaLGandhi]

I believe we have since wanted to move to not unthunk quite so aggressively. Let me revisit FluxML/Zygote.jl#966 too!

[oschulz]

Sorry to be a pest - good to merge now, @DhairyaLGandhi ?

[DhairyaLGandhi]

I think this is alright for now. I do wonder about some of the longer term effects of 966 (we should discuss those there), but seeing what implicit changes happen with generating gradients and the behaviour of rrules in general for a fairly spread out system. I have wanted to run some benchmarks with some more dP kinds of tools like Flux3D, ChemistryFeaturization and all too, which is what I've also been up to over the past couple weeks :)

[oschulz]

I do wonder about some of the longer term effects of 966 (we should discuss those there) [...] run some benchmarks

Oh sure! Can we have a release of ZygoteRules with this PR first though (it's non-breaking), so that we can run full CI on FluxML/Zygote.jl#966?

[oschulz]

Gente bump regarding merge, @DhairyaLGandhi

[oschulz]

I don't mean to nag @DhairyaLGandhi, but it's been a while. Can we move forward with this?

[DhairyaLGandhi]

I think I'm fine with merging this but be careful about 966, this has been on my radar :)

[oxinabox]

@DhairyaLGandhi I don't think @oschulz has merge rights. So if you are fine with this then you will have to merge it?

[oschulz]

I don't think @oschulz has merge rights

No, I don't, you guys will have to merge and make a release, so we can run full CI on FluxML/Zygote.jl#966:

I think I'm fine with merging this but be careful about 966, this has been on my radar

[oxinabox]

It seems I don't have merge rights either.
Which is odd, I thought I had rights on the FluxML/AutoDiff team which had Zygote, Tracker and ZygoteRules.

[CarloLucibello]

I don't have merge rights as well @DhairyaLGandhi

[DhairyaLGandhi]

Ah missed this. Thanks all

[oschulz]

Thanks @DhairyaLGandhi ! Could you release a ZygoteRules v0.2.2, so I can finish up FluxML/Zygote.jl#966 with full CI?

[oschulz]

@DhairyaLGandhi , could you tag a new release of ZygoteRules (should be v0.2.2, I guess)? It would make it much easier to proceed with FluxML/Zygote.jl#966.