Prevent recursion in _eps (#207)

* Revert "Reactant: add extension to prevent stackoverflow (#206)"

This reverts commit 12b7f31.

* Change eps to not be recursive

* Update src/utils.jl

Co-authored-by: Michael Abbott <32575566+mcabbott@users.noreply.github.com>

* Update src/utils.jl

Co-authored-by: Michael Abbott <32575566+mcabbott@users.noreply.github.com>

---------

Co-authored-by: Michael Abbott <32575566+mcabbott@users.noreply.github.com>

Author: wsmoses

Project.toml

@@ -13,12 +13,10 @@ Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 [weakdeps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 EnzymeCore = "f151be2c-9106-41f4-ab19-57ee4f262869"
-Reactant = "3c362404-f566-11ee-1572-e11a4b42c853"
 
 [extensions]
 OptimisersAdaptExt = ["Adapt"]
 OptimisersEnzymeCoreExt = "EnzymeCore"
-OptimisersReactantExt = "Reactant"
 
 [compat]
 Adapt = "4"

ext/OptimisersReactantExt.jl

@@ -15,8 +15,14 @@ end
 
 ofeltype(x, y) = convert(float(eltype(x)), y)
 
-_eps(T::Type{<:AbstractFloat}, e) = T(e)
-# catch complex and integers
-_eps(T::Type{<:Number}, e) = _eps(real(float(T)), e) 
-# avoid small e being rounded to zero
+"""
+    _eps(Type{T}, val)
+
+Mostly this produces `real(T)(val)`, so that `_eps(Float32, 1e-8) === 1f-8` will
+convert the Float64 parameter epsilon to work nicely with Float32 parameter arrays.
+
+But for Float16, it imposes a minimum of `Float16(1e-7)`, unless `val==0`.
+This is basically a hack to increase the default epsilon, to help many optimisers avoid NaN.
+"""
+_eps(T::Type{<:Number}, e) = real(float(T))(e) 
 _eps(T::Type{Float16}, e) = e == 0 ? T(0) : max(T(1e-7), T(e))