Add Galley option to compute()

[kylebd99]

This PR enables galley optimization through the python interface in two ways 1) as a parameter to compute 2) by setting the global scheduler with set_optimizer.

[kylebd99]

All tests pass locally if you point to the Finch main branch.

[kylebd99]

Example Script:

import finch
import time

def main():
    A = finch.lazy(finch.random((1000, 1000), density=.5))
    B = finch.lazy(finch.random((1000, 1000), density=.5))
    C = finch.lazy(finch.random((1000, 1000), density=.01))
    finch.set_optimizer("galley")
    finch.compute(A @ B @ C)
    start = time.time()
    finch.compute(A @ B @ C)
    end = time.time()
    print("Galley Exec Time: ", end-start)
    finch.clear_optimizer_cache()
    finch.set_optimizer("default")
    finch.compute(A @ B @ C)
    start = time.time()
    finch.compute(A @ B @ C)
    end = time.time()
    print("Default Exec Time: ", end-start)
    return

main()

Galley Exec Time: 0.4668142795562744
Default Exec Time: 43.4676399230957

[mtsokol]

Overall, changes LGTM!

[willow-ahrens]

This looks good, but I'd like to make the galley optimizer and the default optimizer be separate functions, so that we can pass them different options. does that make sense?

[willow-ahrens]

This should not be necessary with the latest update (Finch 1.0.0)

[kylebd99]

Do you mean the separate functions or the clearing of the cache?

I just checked and the cache is still being hit without the clearing.

[willow-ahrens]

Have you pulled the latest finch? I just want to double check. It was my intention that Finch 1.0.0 would clear the cache if you use a different tag, but not if you use the same tag. the default does not clear the cache,

[willow-ahrens]

if you want to clear it every time, you may want to give it a different tag every time

[willow-ahrens]

I am not sure if I will be able to clear the cache efficiently in future versions

[kylebd99]

Ah, I see. I need to add tag parameters to this then. That makes sense. I'll update it with that.

[mtsokol]

Let's also make this PR one that moves to 1.0.0 Finch version (but I guess this is required here anyway).

[mtsokol]

Also, in pyproject.toml you can bump finch-tensor version. We can upgrade to version = "0.1.36" or 0.2.0 if it makes more sense. WDYT?

[kylebd99]

Pending @willow-ahrens final review, I think this is good to merge!

[willow-ahrens]

a few very minor points.

[willow-ahrens]

let's put get_julia_scheduler inside the classes, so that we can pass different flags. Then we can just call opt.get_julia_scheduler() in compute.

[willow-ahrens]

I'm just commenting that the isinstance checks can be avoided here, and make it more modular to change just one scheduler or extend in the future

[willow-ahrens]

I don't know if tag=-1 is respected by Finch anymore, it will just be a tag named "-1" without a sentinel behavior.

[kylebd99]

That's fine I guess. It can just be the default tag, and programs can be cached unless the user makes a new tag.