EDIT: SnoopPrecompile has been replaced by PrecompileTools . I’m pleased to announce the availability of SnoopPrecompile , a new package that can help reduce latency particularly on Julia 1.8 and higher. The concept of the package is to make precompilation easy. As described in greater detail in the d…

[image] joa-quim: #@precompile_all_calls begin #plot(rand(5,2)) #makecpt(T=(0,10)) #grdimage(rand(Float32,32,32)) #end This code looks like it needs to go in __init__. Right now it’s at the top level and is probably being executed before __init__.

That’s a perceptive comment. I hadn’t thought about putting the precompiles in the __init__ but it seems like it should work. In such cases @precompile_setup will have no impact but @precompile_all_calls should work as intended. (It’s a bit subtle because __init__ will be compiled before it runs, bu…

I’ve been messing around with this and getting some pretty awesome results :pray: However, one place it doesn’t seem to do much is for precompiling Zygote gradients. It doesn’t seem to improve TTFG (time to first gradient) by much at all. For some random thing I was just trying e.g. it goes from 13…

[image] marius311: Is this a known limitation or is there anything to do about this? (This is with 1.8 btw) This is most likely because Zygote gradients use so many closures, and closures are a different type per session so you cannot precompile them. If those closures were turned into callab…

Brought up here: Remove closures for callable types · Issue #657 · JuliaDiff/ChainRules.jl · GitHub We’ll see if we go through with it.

I’ve found a few cases where inference failed on v1.6 and I had to ntuple(i-> foo(x, i), Val{N}()) instead of ntuple(Base.Fix1(foo, x), Val{N}()) to overcome this.

Having looked into this before, most of that 11s is fixed overhead from having to precompile + codegen + run the AD transform code. Trying out SnoopPrecompile with the following MWE: using Zygote, SnoopCompileCore f(x) = (2x + 1) / 3 - 4 tinf = @snoopi_deep gradient(f, 1.) using SnoopCompile @s…

Are you sure its quite that? To pick a random rrule, it does look like its the same gensym-named type every time: $ julia -e "using ChainRules; println(typeof(ChainRules.rrule(Array, [1,2,3])[2]))" ChainRules.var"#Array_pullback#1301"{ChainRulesCore.ProjectTo{AbstractArray, NamedTuple{(:element, :a…

[image] marius311: Are you sure its quite that? To pick a random rrule, it does look like its the same gensym-named type every time: In one session, run the rule directly. In the next session, create an anonymous function first and then run the rule. #1301 means 1301 anonymous functions wer…

Is it possible to write a macro to automatically convert closures to callable structs? My confusion is that the struct definition should be evaluated in top-level scope rather than inside a parent function. I don’t know if macros can easily do that. Maybe eval can be emitted by the macro to do exact…

ANN: New package SnoopPrecompile

Community

Topic		Replies	Views
Understanding precompilation and its limitations (reducing latency) Internals & Design precompilation	13	2767	December 31, 2020
Re-using compiled code General Usage compilation , precompilation , latency	23	1407	January 30, 2023
Julia precompilation limits or are there really any? General Usage precompilation , snoopcompile , jit	20	3481	June 20, 2022
Mystery of invalidations in V>1.10. Even in precompiled code General Usage regression , method-invalidation , ttfp , ttfx	10	446	September 7, 2025
Struggling to understand how to use PrecompileTools Performance question	25	477	April 26, 2025

ANN: New package SnoopPrecompile

Related topics