Unpin Oceananigans, allow more GPUCompiler versions

[giordano]

Follow up to #799: I merged that PR quickly once I verified Reactant tests were finishing in a finite time to get CI working back, so that now we can do experiments with less pressure. The only combination I didn't try in #799 was pinning only GPUCompiler, let's see if this is sufficient to get decent CI times

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[giordano]

│ [2760] signal 8 (1): Floating point exception
│ in expression starting at /__w/Breeze.jl/Breeze.jl/test/reactant/advection_diffusion_ad.jl:116
└ Allocations: 437508571 (Pool: 437485389; Big: 23182); GC: 219

Fun 😅

In any case, JuliaGPU/GPUCompiler.jl#861 may have resolved the massive slowdown, let's see if that does the trick.

[giordano]

In any case, JuliaGPU/GPUCompiler.jl#861 may have resolved the massive slowdown, let's see if that does the trick.

Apparently it did not 😕

[giordano]

│ [2760] signal 8 (1): Floating point exception
│ in expression starting at /__w/Breeze.jl/Breeze.jl/test/reactant/advection_diffusion_ad.jl:116
└ Allocations: 437508571 (Pool: 437485389; Big: 23182); GC: 219

Fun 😅

Ok, Reactant v0.2.267 has this issue with GPUCompiler v1.22.1-3 (and with v1.22.4 it just hangs). CC @Pangoraw

[macmccallum]

I am noticing a massive slowdown in compilation today. Any chance these changes are causing this? @giordano

[giordano]

What workloads are you talking about? This is only relevant for Reactant, I don't think we're aware of other significant compilation latency regressions elsewhere. Also, this change is to avoid large compilation latency in Reactant.

[macmccallum]

I was using ExplicitTimeStepping against the baroclinic wave example. This may be some intermittent issue. I can't seem to replicate the stalling. Disregard this.